20954 THE WORLD BANK ECONOMIC REVIEW Volume 14 Nlav 20()0 Numnbcr 2 msv, 02,58-67'7() i O 9 4. 3 - I,8 - .... 1 .z ' `.... " - " " .-.r.> 'i, V .. _ 1 . -, .. -r ,* .1 1. i r , -.F . 11 .-l - , - .. I -. I . I . THE WORLD BANK ECONOMIC REVIEW ED)I OR Francois Bourguignon CONStULTING EDITOR Ilyse Zable FDITORIAL BOARD Kaushik Basu, Cornell University and University of Delhi Stijn Claessens Carmen Reinhart, University of Marvland David Dollar Mark R. Rosenzweig, University of Pennsylvania Gregory K. Ingram L. Alan Winters, University of Sussex Martin Ravallion The World Bank Economic Review is a professional journal for the dissemination of World Bank- sponsored research that informs policy analyses and choices. It is directed to an international readership among economists and social scientists in government, business, and international agencies, as well as in universities and development research institutions. The Review emphasizes policy relevance and opera- tional aspects of economics, rather than primarily theoretical and methodological issues. It is intended for readers familiar with economic theory and analysis but nor necessarilv proficient in advanced mathemati- cal or econometric techniques. Articles will illustrate howy professional research can shed light on policv choices. Inconsistency with Bank policy will not be grounds for rejection of an article. Articles will be drawn primarilv from work conducted by World Bank staff and consultants. Before being accepted for publication by the Editorial Board, all articles are reviewed by two referees who are not members of the Bank's staff and one World Bank staff member; articles must also be recommended by at least one external member of the Editorial Board. The Review may on occasion publish articles on specified topics by non-Bank contributors. Any reader interested in preparing such an article is invited to submit a proposal of not more than two pages in length to the Editor. The views and interpretations expressed in this journal are those of the authors and do not necessarily represent the views and policies of the World Bank or of its Executive Directors or the countries they represent. The World Bank does not guarantee the accuracy of the data included in this publication and accepts no responsibility whatsoever for any consequences of their use. When maps are used, the bound- aries, denominations, and other informationi do not imply on the part of the World Bank Group any judgment on the legal status of any territory or the endorsement or acceptance of such boundaries. Comments or brief notes responding to Review articles are welcome and will be considered for publica- tion to the extent that space permits. Please direct all editorial correspondence to the Editor, The World Bank Economic Review, The World Bank, Washington, D.C. 20433, U.S.A. The World Bank Economic Review is published three times a year (January, May, and September) by the World Bank. Single copies may be purchased at $12.95. Subscription rates are as follows: Individuals Institutions 1-year subscription US$30 US$50 2-year subscription US$55 US$95 3-year subscription US$70 US$130 Orders should be sent to: World Bank Publications, Box 7247-7956, Philadelphia, PA 19170-7956, U.S.A. Subscriptions are available without charge to readers with mailing addresses in developing countries and in socialist economies in transition. Written request is required every two years to renew such subscriptions. 0 2000 The International Bank for Reconstruction and Development / THF WORLD BANK 1818 H Street, N.W., Washington, D.C. 20433, U.S.A. All rights reserved Manufactured in the United States of America ISBN 0-8213-4607-5; ISSN 0258-6770 Material in this journal is copyrighted. Requests to reproduce portions of it should be sent to the Office of the Publisher at the address in the copyright notice above. The World Bank encourages dissemination of its work and will normally give permission promptly and, when the reproduction is for noncommercial purposes, without asking a fee. Permission to make photocopies is granted through the Copyright Clearanice Center, Suite 910, 222 Rosewood Drive, Danvers, MA 01923, U.S.A. This journal is indexed regularly in Current Contents/Social & Behavioral Sciences, Index to Interna- tional Statistics, Journal of Economic Literature, Public Affairs Information Service, and Social Sciences Citation IndexO. It is available in microform through Universitv Microfilms, Inc., 300 North Zeeb Road, Ann Arbor, MI 48106, U.S.A. THE WORLD BANK ECONOMIC REVIEW Volume 14 May 2000 Number 2 Understanding Patterns of Economic Growth: Searching for Hills among Plateaus, Mountains, and Plains 221 Lant Pritchett Macroeconomic Fluctuations in Developing Countries: Some Stylized Facts 251 Pierre-Richard Agenor, C. John McDermott, and Eswar S. Prasad Monitoring Banking Sector Fragility: A Multivariate Logit Approach 287 Asli Demirguic-Kunt and Enrica Detragiache Trade Reform Dynamics and Technical Efficiency: The Peruvian Experience 309 Ila M. Semenick Alam and Andrew R. Morrison Monitoring Targeting Performance When Decentralized Allocations to the Poor Are Unobserved 331 Martin Ravallion Child Labor, Child Schooling, and Their Interaction with Adult Labor: Empirical Evidence for Peru and Pakistan 347 Ranjan Ray A NEW DEVELOPMENT DATABASE A Cross-Country Database for Sector Investment and Capital 371 Donald F. Larson, Rita Butzer, Yair Mundlak, and Al Crego THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2: 221-50 Understanding Patterns of Economic Growth: Searching for Hills among Plateaus, Mountains, and Plains Lant Pritchett The historical path of gross domestic product (GDP) per capita in the United States is, except for the interlude of the Great Depression, well characterized by reasonably stable exponential trend growth with modest cyclical deviations: graphically, it is a modestly sloping, slightly bumpy hill. However, almost nothing that is true of U.S. GDP per capita (or that of other countries of the Organisation for Economic Co-operation and Devel- opment) is true of the growth experience of developing countries. A single time trend does not adequately characterize the evolution of GDP per capita in most developing countries. Instability in growth rates over time for a single country is great, relative to both the average level of growth and the variance across countries. These shifts in growth rates lead to distinct patterns. While some countries have steady growth (hills and steep hills), others have rapid growth followed by stagnation (plateaus), rapid growth fol- lowed by decline (mountains) or even catastrophic falls (cliffs), continuous stagnation (plains), or steady decline (valleys). Volatility, however defined, is also much greater in developing than in industrial countries. These stylized facts about the instability and volatility of growth rates in developing countries imply that the exploding econometric growth literature that makes use of the panel nature of data is unlikely to be informa- tive. In contrast, research into what initiates (or halts) episodes of growth has high potential. The aspect of economic growth that makes it "hard to think about anything else" (Lucas 1988) is the implication for human well-being of large and persis- tent differences in growth rates. The power of compound interest, over long pe- riods, turns even small differences in growth into huge shifts in living standards and sustained large differences into seismic shifts. From 1870 to 1980 the United States grew 1.84 percent a year, Great Britain grew 1.24 percent, and Japan grew 2.64 percent (Maddison 1995). The cumulative effect of this 0.6 percentage point lag in British growth relative to U.S. growth resulted in Great Britain's decline from reigning as the world's economic superpower to having to play catch up. The cumulative effect of Japan's 0.8 percentage point edge over the United States Lant Pritchett is with the East Asia and the Pacific Country Units at the World Bank. His e-mail address is lpritchett@worldbank.org. The author would like to thank William Easterly, Ross Levine, Deon Filmer, Dani Rodrik, and the participants at seminars at the World Bank and the Massachusetts Institute of Technology's Urban Studies Group for suffering through a much rougher version of this article and making useful comments. He also thanks Dennis Tao for timely assistance. 2000 The International Bank for Reconstruction and Development / THE WORLD BANK 221 222 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 was Japan's transformation from economic backwater to superpower. The huge growth spurts of some East Asian countries, sustained only over the past few decades, have changed the global economic map. However, because of this fixa- tion on long-run (even, possibly, steady-state) differences in growth, recent theo- retical and empirical growth research has underestimated the importance of in- stability and volatility in growth rates, especially in developing countries. Which aspects of countries' growth are growth theories trying to explain? Does explaining Brazil's growth mean explaining its 4.2 percent growth from 1965 to 1980 or explaining its stagnation from 1980 to 1992?1 Or does theory ignore this break and explain the 1960-92 average of 3.14 percent? Between 1960 and 1980 C6te d'Ivoire grew at 3.1 percent, something of an African growth miracle, while between 1980 and 1992 its gross domestic product (GDP) per capita fell 4.1 percent a year, a growth disaster. Ignoring this break, average growth was 0.22 percent. Nearby Senegal stagnated throughout the same period, with stable growth of 0.18 percent. In what relevant sense are these two growth expe- riences the same? This article has linked halves. The first half provides a set of descriptive statis- tics characterizing the evolution of GDP per capita for a broad cross section of countries, emphasizing the instability in growth rates and the volatility of output. The second half discusses the implications of these facts for recent econometric research. The use of high-frequency panel data, particularly with fixed effects, to investigate long-run growth correlates is almost certainly pointless. Instead, the nature of growth instability suggests future research into the determinants of shifts in growth rates focused on episodes of growth or policy changes. I. DATA AND METHODS The output variable I use throughout is the chain-linked index of real GDP per capita measured in 1985 purchasing power parity dollars (P$) from the Penn World Tables Mark 5.6.2 I use the data beginning in 1960 for the 111 countries with at least 25 years of data. Since the final year of data varies from 1985 to 1992, I refer to it as the "most recent" year. I calculate statistics describing three aspects of growth for each country: average levels, instability, and volatility. The procedures and statistics reported in each category are described in table 1. I separate countries as developing or industrial (table 2). I define industrial countries primarily by membership in the Organisation for Economic Co-opera- tion and Development (OECD), before any recent expansion, and developing coun- 1. Unless otherwise noted, all growth rates are gross domestic product (GDP) per capita per year. 2. It is most likely that none of the results about growth and its characteristics would differ much if I had used the World Bank's national accounts data on real per capita GDP in constant local currency prices. The Penn World Tables Mark 5.6 provide information about the level of per capita GDP in comparable terms, but since for nearly all developing countries there are few benchmark points, most of the time-series content of the Penn World Tables data actually comes from the World Bank data. Pritchett 223 Table 1. Description of the Calculated Statistics on Growth Rates Statistic Reported Basic statistics on output level and average growth Ordinary least squares growth rate: the Growth 1960-most recent year, 1960-73, estimated coefficient b from a trend line 1973-82, 1982-most recent year regression, y, = a + bt + et Initial income GDP per capita for the first year, 1960 Final income GDP per capita for the final year, generally 1992 Average annual growth (YTIY0 ) (1/T) (average of annual growth rates) Ratio of final income to maximum (minimum) YT/max (Yf) and YT/min (Y,) income Statistics on instability in growth rates Growth differences based on the best single Year of breakpoint (t*) breakpoint in trend: if y, = a,I, (t < t*) + Growth before the break (gb) b1 t*l(t < t*) + a11*I(t > t*)+ b1l t*I(t > t*) + t Growth after the break (g ) where t(.) is an indicator function and t* is Difference in growth rates (g6 - g) chosen to minimize the sum of squared errors over all t, such that t* - to 6 and T - t 2 6) Explanatory power of a single trend: y = a + bt + et R2 of the trend regression Statistics on volatility in output Variability of deviations from a single trend: Standard deviation of et et = y, - a' - b'y, is the deviation from a single estimated trend In first difference: In(y, - y X) Coefficient of variation Standard deviation Mean In second differences: ln(y, - y,2) Median of the absolute value Forecast errors: fe, (10,3) = y, - y,', the actual Absolute value of the mean less the predicted value three years ago. The Maximum of the absolute value prediction is y,' = y,-3 + b'3, where b' is estimated on data from the 10 years prior to the forecast date (t- 3, t- 3 - 10) Note: Y is GDP per capita, y is InY, and T is total number of years in the panel. tries as the rest.3 This definition does not correspond to a ranking by initial in- come (in 1960 Republica Bolivariana de Venezuela had higher GDP per capita than France, Iraq than Japan, Mexico than Greece), but I believe the OECD classi- fication better captures the nature of industrial countries than does a classifica- tion based on GDP per capita.4 Using this definition affects the results, as one of the notable features of the data is the very strong performance over this period by the members of the OECD that were poorer intially: Greece, Ireland, Italy, Japan and Portugal. 3. There are three exceptions. I include two Mediterranean islands, Malta and Cyprus, in the industrial category, even though they are not part of the OECD, and I exclude Turkey, even though it is. 4. This definition also differs from the World Bank's "high-income" category by consistently excluding oil producers (such as Kuwait and Saudi Arabia) and by not adding new entrants as they pass an income threshold (such as Singapore and Hong Kong). Table 2. Summary Statistics on Basic Growth Rates Least Growth rates by period GDP per capita (1995 squares (percent) purchasing power parity dollars) growth 1960- 1973- 1982- Finall Final! Statistic (percent) 73 82 recent Initial Final maximum minimum Developing countries Mean 1.64 2.68 1.74 0.10 1,385 2,639 0.82 2.04 Median 1.51 2.72 1.99 -0.13 1,103 1,869 0.88 1.61 4, Standard deviation 1.98 2.20 3.22 2.94 1,089 2,696 0.18 1.37 Industrial countries Mean 2.90 4.26 2.05 2.47 5,430 12,665 0.98 2.69 Median 2.86 3.97 1.79 2.10 5,553 13,118 1.00 2.42 Standard deviation 1.05 1.57 1.51 1.14 2,368 3,062 0.04 0.98 Source: Author's calculations based on Penn World Tables Mark 5.6. See table 1 for a description of procedures. Pritchett 225 II. RESULTS Growth rates were substantially higher in the industrial countries than in the developing countries (table 2). The median growth rate in the industrial coun- tries was 2.86 percent, almost twice the rate of the developing countries (1.51 percent). As many other authors have emphasized (Quah 1996), incomes be- tween industrial and developing countries diverge absolutely; the correlation be- tween initial income and growth rates is positive, 0.22, and the ratio of median incomes increased from 5:1 to more than 7:1 (P$13,118 compared with P$1,869). The period since the 1980s has been very bad for most developing countries (with the important exception of the world's two largest countries, China and India). The gap in growth rates between industrial and developing countries grew substantially in 1982-92.5 Median growth rates for the industrial and develop- ing groups were 4.0 and 2.7 percent in the 1960s. Then in the period after the oil shock (1973-82), growth rates were slightly lower for industrial countries, 1.8 percent compared with 2.0 percent for developing countries. But since 1982 me- dian growth rates have been 2.1 percent for industrial countries and negative (-0.1 percent) for developing countries. The variance in growth rates across countries is also much larger among de- veloping countries. The standard deviation of growth rates is around 1 percent for industrial countries and nearly twice as large, around 2 percent, for develop- ing countries. Figure 1 shows the scatter plot of initial income against subsequent growth rates. Whereas the positive correlation between growth and initial in- come is barely visible, the much larger variance in growth rates among countries that began the period below P$3,000 is striking.6 The wider range of growth experience among developing countries is also seen in comparing the extremes. The industrial countries' growth rates fall into a narrow range as the fifth fastest, Greece, grew at 3.6 percent, while the fifth slowest, Australia, grew at 2.0 per- cent: a difference of only 1.6 percentage points. In contrast, the fifth fastest de- veloping country, Botswana, grew at 6.0 percent, while the fifth slowest, Soma- lia, shrank at 1.4 percent: a difference of more than 7 percentage points (table 3).7 Growth differentials of this magnitude produce rapid shifts in relative in- comes: the Republic of Korea has gone from having less per capita income than Angola to having 10 times more in just 30 years. 5. This diversion has continued since 1992. The growth of population-weighted average gross national product per capita for the decade 1985-95 was -1.4 percent for low-income countries (excluding China and India), -1.3 percent for lower-middle-income countries, 0.2 percent for upper-middle-income countries, and 1.9 percent for high-income countries (World Bank 1997: table 1). 6. The slow and decelerating growth among many of the poorest countries (particularly in Africa), combined with the continued higher-than-world-average growth rates and absolute convergence in levels among the poorer but still well-off European countries, contributes to an emerging "twin peaks" in the distribution of world income (Quah 1996). 7. Of course, the absolute magnitude of the growth differential between the fifth fastest and fifth slowest developing country is also larger because there are more developing than industrial countries (differences in these order statistics tend to grow with sample size). But this does not explain all of the gap. 226 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Figure 1. Growth Rates and Initial Per Capita Income 8 l l l I KOR (5GP 6 TWN HKG iDNT MYS PN 4 LS"O H$CpPU X CHN ~~ PCB RB ISL NOR CPV E-gCOG BRA A ALCA R E(7f SAU BE NL.FRA DE C LJX 2 N L C~S ffiTN PH I ~~~~~~~TTO C~HE C ElZAF CHL ROC NZL_ a SLy ~~~~~~~~ARC X C R ~NG PE R RN ; 0 M8 N VEN -2 TCD -4 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 GDP per capita (in $10,000) Note: See appendix for country names. Source: Author's calculations. Growth Instability Although average growth is of interest, the evolution of most countries' GDP per capita is not well captured by a single trend growth rate. Rather, countries show large shifts in growth rates, often identifiable in episodes (table 4).8 The first noticeable aspect of these shifts is the enormous deceleration of growth. On average, a country's growth has decelerated 2 percentage points. For the indus- trial countries this deceleration is largely the result of two phenomena: the global deceleration following the oil shocks and the deceleration of the European coun- 8. Graphs of each country's growth, including growth statistics, are available from the author. Ben- David and Papell (1997) introduce the analysis of growth shifts with country-specific breaks. The only difference between my approach and theirs is that Ben-David and Papell report summary statistics of changes in growth only for the changes that are statistically significant. This confounds two issues: the magnitude of the shift in growth and the power of the test for the shift. The issue of statistical power is especially problematic given the differing volatilities of the series. We are less able to detect a shift in a more volatile growth series. Given two shifts in growth rates of equal magnitude, but in countries with different underlying volatilities, one shift might be statistically significant and the other not. I take the view that growth rates are simply a convenient summary statistic of the GDP per capita time series. Just as one does not report country growth rates only for those countries in which the rates are statistically different from zero, so I report the before-and-after growth rates as a way of summarizing the GDP per capita time series. Whether those shifts are statistically significant is a different question and should not be the basis for sample selection. Pritcbett 227 Table 3. Five Highest and Lowest Growth Rates of GDP Per Capita Since 1960 (percent) Developing countries Industrial countries Rank Country Growth rate Country Growth rate Five highest growth rates Singapore 6.95 Malta 6.03 Korea, Rep. of 6.85 Japan 4.63 Taiwan (China) 6.29 Cyprus 4.29 Hong Kong 6.15 Portugal 4.10 Botswana 6.03 Greece 3.61 Five lowest growth rates Somalia -1.36 Australia 1.99 Angola -1.97 Sweden 1.88 Madagascar -2.12 United States 1.81 Mozambique -2.25 Switzerland 1.49 Chad -2.75 New Zealand 1.19 Source: Author's calculations based on Penn World Tables Mark 5.6. tries from their rapid post-World War II catch-up of the 1950s and 1960s. For the developing countries deceleration arises from a larger variety of events, dis- cussed below. Second, differences in growth rates within a country over time are large. Among the developing countries the absolute value of the shift in growth rates averages 3.4 percentage points, which is much larger than either the cross-sectional vari- ance of 2.0 percent or the median growth rate of 1.5 percent. Growth rates in 55 of the 111 countries either decelerated or accelerated more than 3 percentage points within the period. Figure 2 shows these shifts with a 2 percentage point band around the 45 degree line (along which growth was equal in both periods), identifying countries whose growth decelerated (those located above the band) or accelerated (those located below the band) by more than 2 percentage points. Third, the evolution of GDP per capita in the developing countries is not well characterized by a single exponential trend. The R2 of fitting a single time trend Table 4. Statistics on Instability of Growth Rates Summary from "best break analysis Percentage Growth before Growth after R2 Statistic Year point shift break (percent) break (percent) of trend Developing countries Mean 1977 -2.58 2.62 0.05 0.58 Median 1978 -2.21 2.86 -0.04 0.67 Standard deviation 4 3.53 2.23 2.99 0.32 Industrial countries Mean 1975 -1.91 4.07 2.17 0.94 Median 1974 -1.93 3.83 1.84 0.95 Standard deviation 4 1.46 1.53 1.04 0.03 Source: Author's calculations based on Penn World Tables Mark 5.6. See table 1 for a description of procedures. 228 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Figure 2. Growtb Rates Bejbre and After Break Point 9 l 8 -PN Above 5. bott 7 - (5;AJ NAA GRC HKG THKG KOR ;6 -YUG EPR ~ LY ^1 5 Deceleration (> 2) JOR RNSYR D -PfRO N~~ IC Z M/ v V2 N /2 3 percent, Japan, Malta (Rep. of), g > 3 percent) Malaysia, .) Singapore, Taiwan (China), Thailand Hills 27 Australia, Austria, China, Myanmar, Bangladesh, Israel, Tunisia, Barbados, Tanzania (gh > 1.5 percent, Belgium, Canada, Philippines Pakistan Turkey Colombia, Costa ga > 1.5 percent) Denmark, Finland, Rica, Mexico France, Germany, Greece, Italy, Portugal, Spain, Switzerland, United States Plateaus 16 Iceland, Morocco Brazil, Ethiopia, The (gb > 1.5 percent, Netherlands, New Dominican Gambia, Guinea- O < g < 1.5 percent) Zealand, Sweden Republic, Bissau, Kenya, El Salvador, Lesotho, Malawi, Guatemala Swaziland Mountains 33 United Kingdom Namibia, Papua Algeria, Egypt, Argentina, Cameroon, Congo, (gb >1.5 percent, New Guinea Iran, Iraq, Bolivia, Ecuador, C6te d'Ivoire, g, < 0 percent) Jordan, Saudi Guyana, Gabon, Liberia, Arabia, Syrian Honduras, Mozambique, Arab Republic Jamaica, Niger, Nigeria, Nicaragua, Sierra Leone, South Panama, Africa, Togo, Paraguay, Peru, Zaire, Zambia Suriname, Trinidad and Tobago Plains 17 Nepal Haiti, Republica Angola, Burundi, (gb < 1.5 percent, Bolivariana de Benin, Central ga < 1.5 percent) Venezuela African Republic, Guinea, Burkina w Faso, Madagascar, Mali, Mauritania, Rwanda, Senegal, Somalia, Uganda, Zimbabwe Accelerators 7 Indonesia India, Sri Lanka Chile, Uruguay Ghana, Mauritius (gh < 1.5 percent, g& >1.5 percent) Number of 111 23 12 5 11 24 36 countries Note: gb (g,) is growth measured before (after) the structural break, using procedures described in table 1. Source: Author's calculations. 232 THE WORLD BANK ECONOMIC REVIEW, VOL 14, NO. 2 Figure 4. GDP Per Capita, by Pattern of Growth, 1960-92 Figure 4a. Steep hills (In) GDP per person Thailand (In) GDP per person Korea 19(X) 192 4 1968 1972 1976 1986 1984 1998 1992 196 1964 1968 1972 1976 1981) 194 1988 1992 Figure 4b. Hills (In) GDP per person Ulnited States (In) GDP per person Pakistan 7.9 7. 7 636 1960 1964 1868 1972 1976 1980 1984 1988 1992 1960 196s 1968 1972 1971,6 1981) 1984 1988 1992 Figure 4c. Plateas In) GDP per person Brazil (In) GDP per person Guatemala 75 73 1960 1964 1968 1972 1976 1981) 1984 1988 1992 1960 1964 1968 1972 1976 1989 1984 1988 1992 Nigeria, Saudi Arabia), a number of commodity exporters that experienced positive commodity price shocks followed by negative shocks ( Cote d' Ivoire, Guyana, Jamaica, Zambia), and Latin American countries affected by the debt crisis (Argentina, Bolivia, Paraguay). The mountains include some countries with cliffs, very sharp drops, usually resulting from war or civil unrest (Liberia, Mozambique, Nicaragua). Because of a sharp break in their growth, the mountain countries show a lov trend R2 (for example, Cote d'Ivoire, 0.013; Argentina, 0.204; Nicaragua, 0.190). Pritchett 233 Figure 4. (continuzed) Figure 4d. Mountains (In) GDP per person C6te d'Ivoire (In) GDP per person Guyana 6.S~~~~~~~~~~~~~~~~~~~~ 68 7.1~~~~~~~~~~~~~~~~~~~~. 1900 1964 1968 1972 1976 19890 1984 1988 1992 196) 1964 1968 1972 1976 1980 1981 1988 1992 Figure 4e. Plains (In) GDP per person Senegal (In) GDP per person Madagascar ~~~~~~~~~~~~~~~~~~~~~7 '1 7 . - 1 6,8 6.8 625 1960 1964 1968 1972 1976 1980 1984 1988 1992 1960 1964 1968 1972 1976 1980 1984 1988 1992 Figure 4f. Accelerators (In) GDP per person Indonesia (In) GDP per person India 7 8 8.0 6.9 i7 ov' '' .'I6 6.0 6.8 1960 196i 1968 1972 1976 1980 1984 1988 1992 1960 1964 1968 1972 1976 1980 1984 1988 1992 Note Dotted line is the trend. Souerce Author's calculations. * Plains. These 17 countries had growth rates less than 1.5 percent both before and after their structural break (figure 4e). Nearly all of these countries (14 of the 17) are in Sub-Saharan Africa. Senegal is a classic plain, with continuous stagnation and a fairly steady growth rate around zero (0.18 percent). Hence GDP per capita is reasonably characterized by a single trend, but nevertheless has a low R2 (0.213). Included among the plains countries are those with consistently negative growth rates, such as Mozambique, which could be characterized as valleys. 234 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 * Accelerators or "Denver."'1 These seven countries did not have growth rates above 1.5 percent before their structural break, but did afterward (figure 4f). This class includes a number of clear successes, like Indonesia after 1966 and Mauritius after 1970, as well as less clear-cut successes, like India. For some countries, such as Ghana, the acceleration was from low or negative rates to respectable, but unimpressive, rates. This classification scheme captures some interesting stylized facts about differ- ences in growth across regions." The OECD countries are nearly all hills or steep hills (18 of 23), and even the five exceptions are borderline. Nearly all of the plains are countries in Sub-Saharan Africa (14 of 17), but not all Sub-Saharan African countries are plains (only 14 of 36); a nearly equal number of countries are mountains (13 of 36). This even division contrasts with two other regions with slow overall growth: Latin America, with very few plains (2 of 24) but many mountains (12 of 24), and the Middle East and North Africa, where 7 of 11 countries are mountains. One implication of large changes in growth rates is that there is relatively little correlation across periods. Since the study by Easterly and others (1993) is de- voted entirely to this point, I will not belabor it. Although casual discussions of high-performing and low-performing countries make it seem as if relatively time- persistent characteristics account for the bulk of the variation in growth across countries, the cross-national (rank) correlation of countries' growth before and after their structural break is only 0.24. Growth Volatility If a time series can be well represented by a single stable growth rate, then measuring the volatility around that trend is relatively straightforward.12 How- ever, since nearly every developing country exhibits a large shift in trend over time, simple measures based on the residuals from a single trend do not give a good indication of the pure volatility of output. For instance, if one country has high volatility around a stable trend while another has very stable output in each of two subperiods, but around two different trends, the two countries would appear to have similar volatilities. This is the reason for using a variety of mea- sures, including measures that allow for a shifting or rolling trend. The correla- 10. The only geographic metaphor I thought of was Denver, where the Great Plains meet the Rocky Mountains. 11. The classification also throws up a few anomalies, which reveal some limitations of the method. For instance, China is a consistent growth performer (a hill) because the data only allow one break, smoothing over the disasters of the Great Leap Forward. Similarly, the data break Tanzania's growth at 1980 and give two reasonably high-growth subperiods, smoothing over the disastrous years from 1978 to 1984. Great Britain is a mountain because the data break their otherwise very smooth series at the peak of a business cycle in 1987. 12. Although I use the simple trend and deviations throughout, I suspect that I would find similar results about differences in volatility and shifts in the drift parameter among countries if I were to treat the series as difference-stationary. Pritchett 235 tions among the measures are high, but not high enough to suggest that there are not real differences as to the aspects of volatility the different statistics are capturing. However measured, volatility is much higher in the developing countries. The median standard deviation of the deviation from trend is twice as high in the developing countries as in the industrial countries, 0.10 compared with 0.05 (table 6). The median forecast error is also nearly twice as large in the developing countries (0.095 compared with 0.054), and the typical maximum forecast error is also twice as large (0.28 compared with 0.14). The coefficient of variation of the (natural) log of first differences of GDP per capita is four times as high in the median developing country as in the median industrial country (4.3 compared with 1.04). Figure 5 shows the scatter plot of one mea- sure of volatility (the standard deviation of the deviations from a single trend) against initial GDP per capita. III. So, You THINK YOU WANT To RUN A GROWTH REGRESSION? What are the implications of the instability and volatility of per capita GDP for empirical research into the determinants of economic growth? Recently, this research has expanded to regressions using higher-frequency data and country- specific growth effects, motivated by two arguments. First, there is a naive no- tion that one should use all the data so as not to throw away information. Second, the correlation between the included growth correlates and unobserved country-specific growth effects could generate misleading results. However, the commonly proposed cure of higher-frequency data and country-specific effects could easily worsen the disease. Given the instability and volatility of output, moving to shorter and shorter time periods and eliminating long-period vari- ance are likely to entangle dynamics, specification, endogeneity, and statistical power, which will ultimately confuse, not clarify, issues of growth, especially in developing countries. Preliminaries A theory of economic growth relates the level of income at each point in time (and hence its growth rate) to another set of variables. Three dimensions of growth are traditionally distinguished based on the notion of the equilibrium level of income as a function of the underlying (X) variables, denoted y, *(X), and the actual level of output, y. * The steady state refers to the growth rate of the steady-state level of output, Y*(-)- * Transitional dynamics refer to movements of output as it adapts to changes in the steady-state level or steady-state growth rate of y*(.). * The business cycle refers to the dynamics in actual output, y, without shifts in either the level or growth rate of y*(.). Table 6. Summary Statistics on the Volatility of Output Standard Median absolute deviation of First differences value of second deviation from Coefficient Standard differences Forecast error (3,10) Indicator single trend of variation deviation Mean (*100) Mean Maximum Developing countries Mean 0.112 10.597 0.065 0.018 0.7 0.114 0.314 Median 0.100 4.314 0.059 0.014 0.5 0.095 0.283 Standard deviation 0.052 19.449 0.027 0.015 0.7 0.053 0.166 Industrial countries Mean 0.063 1.140 0.031 0.029 0.4 0.056 0.153 Median 0.054 1.040 0.027 0.028 0.3 0.054 0.135 Standard deviation 0.026 0.473 0.013 0.011 0.4 0.022 0.086 Source: Author's calculations, based on Penn World Tables Mark 5.6. See table I for a description of procedures. Pritchett 237 Figure 5. Volatility and Per Capita Income 0.24 _0 NAM C 0.20 GUY C N E NIC X UGA SWSYR C 0.16 - IS AGSOZ SUR GNB YG .5 ~~G 8# J OR MU S .> TGO DM E C LW 1 M FAN Fq o 0.12 - JRY4 SGP BWM QJ[ PN 5 . q9N D F~C~ ESpIs R -~~~ 4, ~~CY ~IF4IISR .SK BRB ~ 0.08 _ SBB - S REVS URY, ~~~~~~ISL P ~~~~~~FIN FRkLD SUN -UUL CAN LUX RL TA AT ~ CN u a ~~~~4~NCOL HKG R U NOR CHE - 0.04 5IIT SE-;N R CBH L SA 0.00I l ll l ll ll l l l 0.0 0.1 0.2 0.3 0.4 0.5 o.6 0.7 0.8 0.9 1.0 Initial GDP per capita (in $10,000) Note: See appendix for country names. Source: Author's calculations. Growth, g,(N), at time t over horizon N, has a steady-state, a transitional, and a cyclical component: (1) g,(N) =Y Yt- =(Yt Yt-n) + (y[- Y[ ) + (yt- Yt-n) Obviously, the fraction of the variation of growth (over time and across coun- tries) due to each of these three components varies with the horizon N. The second element necessary to organize a discussion of the implication for growth research is to classify the myriad possible growth correlates according to three features: time-series persistence, exogeneity, and model rationale. Persis- tence ranges from country-specific, time-invariant variables, such as latitude (Hall and Jones 1999) and access to the sea (Gallup and Sachs 1999), to quantities that evolve very slowly, such as population size and human capital stock (Barro 1997) and trust (Knack and Keefer 1997), and to highly volatile series, such as black mar- ket premia, capital inflows, and terms of trade. Whereas stable, high-persistence variables are also usually exogenous, with little or no feedback from growth, the volatile variables can be either exogenous to a country's growth (terms of trade shocks) or highly endogenous (foreign investment). Potential growth correlates can be further classified according to their model rationale, the postulated causal chain from the variable to growth. That is, every model tells a story in which one thing is affected by another, which then affects a third, which in turn affects output. Hence the model rationale includes (too often 238 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 only implicitly) information relevant to the extent of exogeneity and the compo- nent of growth affected (steady-state, transitional, cyclical) and the expected time scale of the impact. This information can range from exogenous variables acting over decades to simultaneously determined variables affecting output over sev- eral months. Some types of model rationales include structural, shock, produc- tion function, policy, intermediate-outcome, and institutional (table 7). Four Problems with Growth Regressions in Higher Frequencies With the classification of the three components of growth and the classifica- tion of potential growth correlates by their persistence, exogeneity, and model rationale, the elements are in place to argue that the use of shorter panels is as likely to hurt as to help studies of long-run growth. There are four problems with using higher-frequency data, particularly with techniques that remove country- specific effects: lower power, greater measurement error, endogeneity, and dy- namic misspecification (which itself comes in three flavors). LOWER POWER. The lack of identification of country-specific, time-invariant variables using fixed effects in panel data is merely the limiting case of the decline Table 7. Classification of Variables Included in Growth Regressions Persistence Endogeneity Stable Medium Volatile Low Structural: geographic Shocks: terms of (land-locked, distance trade, spillovers from from the equator), financial crises, climatic (rainfall), weather resource endowment (minerals) Institutional: for example, ethnic diversity, political system, language, colonial experience, type of legal system Medium Policy: quantities over which some individual or entity has more or less direct control (such as tariff rates) High Intermediate-outcome: Intermediate- for example, trade outcome: for example, ratio, inflation, budget foreign direct deficit, financial depth investment, export growth, budget deficit, black market premium Source: Author. Pritchett 239 in statistical power as the "between" country variance in time-persistent right- side variables is swept out by the fixed effects. Table 8 compares the fraction of total variation in five-year panel data that is due to variation over time within countries as a measure of a variable's persistence. The fraction of variance in per capita GDP growth that is within country is 0.73, reflecting instability of growth as well as volatilitv of output. In contrast, for many growth correlates the within- country variance is very low, only 0.22 for investment rates, 0.07 for level of education, and 0.02 for population size. Other growth determinants also have strong persistence: Isham, Kaufman, and Pritchett (1997) show high persistence for measures of democracy and civil liberties, while Deininger and Squire (1996) report that in their panel data on inequality only 10 percent of the variation in inequality is within country. The difference in persistence between growth and growth correlates implies that fixed effects will sweep out much of the variation in right-side variables, while increasing the proportion of variance due to the volatile components of growth. We can estimate how much lower the t-statistics would be from fixed effects relative to time-averaged cross sections using a Monte Carlo evaluation of a simple bivariate regression with 100 countries and six periods (table 8). For a variable with the same persistence as level of education, the use of five-year- horizon fixed-effects estimates would cut t-statistics to a third of their level in the cross section. This implies that in the cross section the t-statistic would have to be 6 or higher to avoid being made statistically insignificant simply by the lower power of fixed effects. MEASUREMENT ERROR. A well-known problem of panel econometrics (pointed out early, and subsequently often, by Grilliches and others) holds that if the cross- Table 8. Differences in Persistence between Economic Growth and Typical Explanatory Variables in Growth Regressions and the Implications for Statistical Power Ratio of t-statistics from fixed elfects versus cross- section regressions for a Ratio of right-side variable with within- the ratio of within-country Number Number country to total variance indicated of of variance to (based on Monte Carlo Variable countries observations total variance simulation) Growth of GDP per capita 126 756 0.73 Population growth 126 756 0.31 0.62 Investment rates 126 756 0.22 0.54 Level of education 84 504 0.07 0.32 In of population site 126 756 0.02 0.18 Note: The predicted ratio comes from a Monte Carlo simulation of data for 100 countries over six periods with growth and the right-side variables having different time-series persistence. By varying the degree of persistence of right-side variables while holding the persistence of growth at 0.75, a tight predicted relationship with t-statistics is produced. 240 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 sectional variance is large relative to the time-series variance and if the measure- ment-error variance has a large time series relative to the cross-sectional vari- ance, then the use of fixed effects exacerbates the degree of measurement error and hence increases the attenuation bias. This effect can be enormous, especially with high-persistence variables, in which the variance of the measurement error is constant for each repeated measurement. The use of fixed effects with vari- ables that have persistence characteristics similar to those in table 8 could easily increase attenuation bias due to measurement error by a factor of 10, making even large effects disappear. ENDOGENEITY BIAS. Short panels can exacerbate the endogeneity problem, as seen in a simple hypothetical model. Suppose there are long-run growth, g., and business cycle effects in the determination of current output for each country i at time t: (2) g' = i. In addition, suppose that there is a causal relationship from the country-specific average of x to long-run growth (gj): ) Yt = Yt -1 (1 += g ) (1 + gc ) Finally, suppose that although the average level of x is the result of a policy choice, the cyclical component of x is a result of the business cycle component of growth: (4) (x - xi) = 691 . In this simple setup it is easy to see that if one is trying to identify the long-run impact on growth of policy changes in x, j, then moving from the time-averaged cross section to fixed effects on panels will be a disaster. Instead of identifying , the fixed-effects estimator identifies the impact of growth on x, 6, which repre- sents an entirely different phenomena. This problem has wide implications be- cause many growth correlates (especially intermediate-outcome variables) are endogenous. DYNAMIC MISSPECIFICATION. The fourth problem is dynamic misspecification created by arbitrarily changing the time span over which the regression is esti- mated. Dynamic misspecification raises two distinct problems. The first is the dynamic misspecification of the time scale over which the growth correlates have an effect. Arbitrarily parsing time series into shorter periods imposes the assump- tion that the dynamics are invariant across growth correlates.13 Although includ- ing lagged income levels creates some ad hoc adjustment dynamics, this still as- sumes that the speed of adjustment is equal across the right-side variables. But, in fact, although some growth effects are contemporaneous, especially macroeco- 13. Arbitrary is not too strong a word: I have seen published growth regressions at 1-, 3-, 4-, 5-, 7-, and 10-year horizons, justified only on the grounds that data were available at those frequencies or the researcher wanted to divide the whole period into equal chunks. Pritcbett 241 nomic and cyclical factors, others could take several years, such as transitional dynamics due to changes in investment incentives, and still others could take decades, such as the impact of changes that affect the rate of technical progress. Some right-side variables may have output or growth effects at all horizons- cyclical, transitional, and steady-state-and there is no reason to believe that these effects are of similar magnitude, nor have the same sign, because some policy choices may lead to temporary booms but ultimately to busts. The second problem is that emphasizing the higher-frequency components of the right-side variables assumes that transitory and permanent changes in a growth correlate have the same impact on growth rates at every frequency. This is false for most variables in dynamic-optimization macroeconomic models in which re- sponses depend on expectations. The output response to changes in a growth correlate (tariffs, taxes, terms of trade, investment incentives) can be of a com- pletely different order of magnitude, or even sign, depending on whether the change is perceived as permanent or temporary. For any given growth correlate, x, there are six underlying sets of output- response coefficients: the steady-state (*), transitional (T), and cyclical (C) dy- namics in output in response to either a permanent (P) or temporary (R) change in the growth correlate. These are sets of coefficients, as each has its own growth dynamics with possible lags from zero (contemporaneous) to k with the resulting k long-run impact L = pm,n n =O k k (5) Yt = IPt-n XtL_n t-n t-n n=O n=O k k I = XPt[fX n +Pt-n Xt-n n=O n=0 k k YtC= Y, P +tC Xt-nxt-n n=0 n=0 A typical growth regression specification is: (6) gt(n) = E)(n)f (x,, XI-J + ?Yt-n where f (.) is a function of the N annual observations on x (usually a simple average, beginning-of-period values, Xt - n, or end-of-period values, xt) when in levels or usually just xt, X,x when in differences. The resulting coefficient, e(n), is a complex weighted average of the underlying Ps. Since the weights vary with the variance components of y* yT, and yC and with the permanent and tempo- rary components of x reflected in the data, e(n) depends strongly on the chosen horizon, N. Since the Ps need not even have the same sign,6(1), 6(5), and 0(30) 242 THE WORLD BANK ECONO-MIC REVIEW, VOL. 14, NO. 2 are not estimates of the same parameter using data at different horizons; rather they are estimates of different underlying combinations of parameters. The same is true of HFE (n) and e9OLS (n). Taking out the fixed effect enhances the relative signal of output that is cyclical and the relative signal of the right-side growth correlates that is temporary (versus permanent). By using the within- country variation to identify coefficients, this approach may completely miss, or change the sign of, important long-run impacts on growth. Implications for Reading the Growth Literature These four problems make it impossible to assert that the higher-frequency re- gressions have done a better job of estimating the structural relationship between growth and a candidate growth determinant. For any growth correlate the empiri- cal findings using time-averaged cross sections and those using higher-frequency data with country-specific effects can differ in four possible ways. The correlation coefficient of x could not be robust, could fall in magnitude, could rise in magni- tude or change sign, or could be unstable over time. None of these possibilities is of any particular interest in helping us understand long-term growth. NOT ROBUST. A common finding is that many estimates are not robust to the inclusion of country effects, in the mechanistic sense that variables that are statis- tically significant in the cross section are not statistically significant in panels with country effects. For instance, although cross-sectional studies typically find that, conditional on initial income, the level of education (or the enrollment rate) is a significant determinant of subsequent growth (Mankiw, Romer, and Weil 1992 and Barro and Sala-i-Martin 1995), panel regressions typically find that the level of education is insignificant (Islam 1995).14 In many cases the confi- dence interval of the fixed-effects estimate is large enough to include zero (and hence not be statistically significant), even though the point estimate is larger than the cross-section estimate. In this case all that is learned is that the new estimate is less precise. A failure to reject is completely uninformative unless accompanied by a serious analysis of statistical power (Andrews 1989). LOWER IN MAGNITUDE. A second possibility is that in moving to estimates with country effects, the magnitude of the point estimate falls. But given the substan- tial measurement error in most of the growth correlates and the well-known exacerbation of attenuation bias by transformations that reduce signals more than noise, a smaller coefficient is completely uninteresting-without a serious remedy for the measurement error bias. A smaller and statistically insignificant growth correlation from panel estimation is twice as uninteresting. HIGHER IN MAGNITUDE (OR CHANGE IN SIGN). If robustness and falling magni- tude were the only two problems, the situation would not be completely hope- 14. The growth-human capital regressions that mix economic growth with education levels or enrollment rates have additional, extremely serious, empirical and theoretical problems (Pritchett 1996b). Pritchett 243 less, as the direction of the changes from these two problems (larger standard errors and attenuation to zero) are well known. The results might still be interest- ing if the estimated magnitude of the partial correlation rose or if the sign of the partial correlation changed. However, once the problems of endogeneity and dynamic misspecification enter, anything can happen. The interpretation of the differences between the cross-sectional and fixed-effects estimates depends en- tirely on the underlying theories of the cyclical and adjustment dynamics of out- put-which typically are not developed within growth theories. An example illustrating these two problems is a regression relating growth and budget deficits. If a country pursues a countercyclical fiscal policy, then peri- ods of low cyclical growth would correspond to periods of high deficit. Using short-period data (and, given the large magnitude of output volatility relative to trend, "short" could be quite long), estimates could easily show that budget defi- cits have a large negative effect on growth when, in fact, the causality is exactly the reverse. An even more likely possibility is that moving to fixed effects in- creases the omitted variable bias from dynamic structural misspecification. There are many kinds of temporary shocks-war, political disruption, terms-of-trade movements, adverse weather-that affect output in a variety of ways. These shocks are more highly correlated with the time-series dimension of fiscal deficits than with cross-national, long-period averages. Time averaging reduces the correla- tion between growth and the unobserved (or not included) shock variables and the correlation between growth and fiscal deficits-hence reducing the omitted- variable bias. However, moving to higher frequencies and removing country ef- fects worsen these sources of bias (Easterly and others 1993). Changes in magnitude and sign can be the result of arbitrarily changing the time scale of regressions without considering the dynamics implicit in any given model rationale. For instance, several recent studies find that the negative corre- lation between inequality and economic growth in cross sections is not robust, and even reverses sign, when using five-year panels (Forbes 1997). But certainly the models that propose a model rationale running from inequality to median- voter preference to politically determined tax rates to investment to growth (Persson and Tabellini 1994) or a rationale running from inequality to political instability to investment to growth (Alesina and Perotti 1993) are not meant to be tested using a contemporaneous relationship between short-run deviations of growth from its long-run average and short-run deviations of inequality from its long-run average, while sweeping out permanent cross-national differences. Moreover, one can easily imagine models in which a short-run increase in growth causes a short-run increase in inequality, whereas a long-run increase in inequal- ity causes a decrease in output or growth."5 A finding that the short-run impact is 15. Julst to prove that it is easy to imagine such a model, here is one (of many possibilities). A positive shock increases the returns to human capital, given the role of human capital in adapting to disquilibrium. Thus there is a temporary increase in growth and a temporary increase in inequality. At the same time persistent inequalities in income distribution lead to unequal educational opportunities across talent levels, thus lowering the long-run quality of human capital. 244 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 different than the long-run impact does not contradict or refute the robustness of the long-run growth correlation. An important point that seems to have been overlooked is that a Hausman- Wu type specification test is an omnibus specification test. Many studies frame the test as one of the bias due to correlation with omitted country-specific vari- ables. Hence they interpret a rejection as evidence that ordinary least squares is an inconsistent estimator of the parameter of interest. This reasoning is incorrect. Any misspecification that prevents eFE and oOLS (n) from converging to the same parameter vector 00 can cause a rejection, including the endogeneity, mis- specification, and measurement-error problems described above. A rejection of the null-Ho: OOLS (30) = EOLS (5), either informally or through a vigorous specification test-does not lead to the conclusion that ordinary least squares is an inferior estimate of the long-run impact of permanent changes in 'LR. The rejection that may arise may be due not to the bias in E3OLS (30) as an estimate of 3LRP or to the problem of correlation with unobserved and omitted country-specific effects, but rather to dynamic misspecification or the exacerbation of endogeneity in FE (5) as an estimate of 1LR' With intermediate-outcome variables (such as the black market premium, in- flation, or the ratio of trade to GDP or investment), which are themselves time- varying, the combined problems of endogeneity and dynamic misspecification may be overwhelming. Intermediate-outcome variables are not, strictly speaking, "policy" variables, since they are determined both by policy actions and by re- sponses to those policies. For these cases, although simple time averaging might not be the optimal filter, it may be better than no filter at all, as the additional signal in higher-frequency data may not be a signal relevant to identifying the long-run impact, P'LR- UNSTABLE. The final empirical problem often found with the use of higher frequencies is that the regression parameters are not stable over time (with or without fixed effects). In regressions run on 10- or 5-year periods the statistical significance and even the signs of coefficients are not the same as those in regres- sions run on time-averaged data. Kelley and Schmidt (1994) regress GDP per capita growth on population growth decade by decade and find a mildly negative coef- ficient in the 1960s, a mildly positive coefficient in the 1970s, and a larger and negative coefficient in the 1980s. Similarly, Knack and Keefer's (1997) study on social characteristics and growth shows substantial parameter instability-the social variables have different signs in the regressions on growth after 1980 than in regressions on growth prior to 1980, even though one would suspect that the underlying social characteristics actually changed very slowly (the study has data for only one year). Similarly, one of the best-researched findings in the growth literature is the connection among financial depth, characteristics of the banking sector, and growth (Levine 1997 and Levine and Zervos 1998). However, even this relationship shifts in decade-by-decade data. Vavamkidis (1997) looks at the Pritchett 245 relationship between growth and trade over the long run, using data going back into the 1900s, and finds that the partial correlation of the two shifts over time. Well, so what? So absolutely nothing. Parameter instability is not particu- larly informative because it is hard to reconcile with any of the coefficients that identify a "true" invariant structural parameter that links a variable with out- put (or growth) at all frequencies.16 But, then, how does one interpret these parameter shifts? Does the underlying structural relationship produce the reduced-form shifting? Does this reflect shifts in temporary growth dynamics or in the steady-state components of growth? Do these shifts reveal true strate- gic opportunities: it was good to be open to trade in the 1960s, but not in the 1930s? It was good to be financially deep in the 1980s, but not in the 1970s? To make assertions about time-varying relationships between determinants and growth requires a growth theory that specifies not only what these relation- ships are but also how they shift over time. Without such a theory, it is impos- sible to say which time-varying parameter is relevant for the current or future periods. For instance, suppose one really believed that the association between population growth and per capita GDP growth changed from the 1960s to the 1970s to the 1980s. What could be inferred from the past about the expected relationship in the 1990s? Without a (verified) explanation of the causes of the parameter shifts, the answer is: nothing. The Way Forward Many agree that growth regressions as a tool for investigating long-run growth have passed the point of diminishing returns (see my requiem for growth regres- sions, Pritchett 1997). There is a group that sees a way forward in combining higher-frequency data with new econometric techniques for dynamic panels. These techniques are aimed primarily at solving another, more narrow econometric problem that arises in panels, that of the bias in dynamic regressions with a lagged endogenous variable. Although it is possible that these techniques might address some of the problems raised here (such as measurement error), so far most of the work has depended on more or less arbitrary restrictions on the time- series properties of the data for identification. Opinions still differ on whether using dynamic generalized methods of moments estimators identified from as- sumptions on dynamics is a fruitful approach for understanding countries' eco- nomic growth and its determinants. I have become more, rather than less, doubt- ful by the results produced. But since it is hard to think about anything other than long-term growth, let me end on a positive note and offer some suggestions on ways to research the determinants of growth that might even contribute to policy. First, analysis of the episodes of growth acceleration or the onset of growth deceleration has prom- 16. Parameter stability is also an omnibus specification test so that a failure to reject is some evidence of a correct specification, but a rejection of parameter stability could arise from any number of specification problems. 246 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 ise. 7 For each of the plateau, mountain, and cliff countries there is usually an easily identifiable turning point after which growth is much slower. Moreover, for many of the steep hill and accelerator countries there is an identifiable takeoff date after which growth is much more rapid. One research strategy that seems promising is to examine the economic, political, institutional, and policy condi- tions that accompany these break points: why did growth suddenly change sharply at a certain time? Research also could identify which political conditions made the adoption of such reforms possible or made the adoption of policies that would avoid a decline in growth unlikely (Rodrik 1996). Studies of rapid growth in Korea in the early 1960s (Haggard, Byung-Kook, and Moon 1990), Mauritius in 1970 (Romer 1993), Indonesia after 1966, China since 1978, or Chile since 1988 are extremely useful in establishing the condi- tions that initiate episodes of growth. Similarly, we can also ask why some coun- tries hit periods of slow or negative growth and were unable to reverse the de- cline. This question is particularly interesting in a comparative context in which countries experienced similar shocks, but their growth responses differed widely: Korea's response to the debt shock of the early 1980s compared with Brazil's, Chile's response to the long-run collapse in the price of copper compared with Zambia's, Nigeria's response to the decline in oil prices in the 1980s compared with Indonesia's, or, more recently, Indonesia's response to the Asian crisis com- pared with Thailand's. The second type of useful study is an analysis of discrete episodes in the evolu- tion of potential growth determinants. For instance, Bruno and Easterly (1998) show that it is impossible to estimate the impact of inflation on growth from either long-period averages or panels. But when one analyzes episodes of infla- tion, one finds clear and robust, if surprising, results. Similarly, Krueger's (1978) study of discrete episodes of changes in exchange rate regimes and import liber- alization in 10 countries, although "ancient," still has, in my opinion, more in- sights on the impact of policies on growth and ultimate persuasive power for policymaking than the hundreds of growth regressions with openness as a right- side variable.18 If variable x is a powerful growth determinant, then large changes in x should be followed by large shifts in output and growth. Hence beginning from known episodes of policy change is a promising avenue. The third approach that has great potential is cross-sectional analysis of changes in growth rates over time. This approach has great potential because it is almost unexplored. In fact, when I wrote the first draft of this article, I could claim that 17. A rich economic literature in a number of fields relies on episodic analysis, based on both statistical tests and case studies. The studies on the effects of devaluation by Kamin (1988) and Edwards (1989) rely on identification of discrete devaluation episodes; studies of the impact of debt crises, banking crises, and currency crises similarly rely on episodic analysis. 18. Pritchett (1996a) explains that the lack of internal coherence of the various indicators of openness makes the usual interpretation of the results dubious in any case. Pritchett 247 there were no studies looking systematically at the determinants of changes in growth rates. However, Rodrik (2000) investigates the change in growth be- tween the early and later periods of the data (using at one stage the data from this work). He shows that growth deceleration is determined by a combination of shocks to the economy and countries' ability to adjust to those shocks, which in turn is determined by social and political factors. Rodrik's article shows the prom- ise of taking instability seriously and examining shifts in growth rates as well as levels. IV. CONCLUSION The exception should prove the rule, not be mistaken for it. Although stable growth rates are the exception outside of the OECD, those exceptions-the East Asian countries that sustained rapid growth rates for three decades (at least until 1997)-have captured the imagination. Both casual talk and an academic litera- ture on growth have focused on why some GDP per capita hills are steeper that others. However, the rule of growth in developing countries is that anything can happen and often does. The instability of growth rates makes talk of the growth rate almost meaningless. Moreover, the enormous volatility of growth around its trend (however defined) means that even over periods as long as a decade, growth can be dominated by shocks and recovery. This implies that the arbitrary parsing of the entire time series of output into different lengths is unlikely to lead to significant, policy-relevant insights into growth. Although we have learned some things from examining growth corre- lates with multivariate regressions of various types, there is little more to be learned by moving to panels. This approach leads to low power, greater measurement error bias, confusion about causality and endogenity, and dynamic misspecification of many stripes, all of which cloud the interpretation of regressions using higher frequencies. One can certainly question the usefulness of a technique that might cause the estimated partial correlation to rise, fall, shift, or lose statistical signifi- cance, when any of these would have no impact on inferences about the impor- tant question of interest: the impact of permanent policy shifts on long-run out- put or growth. A more promising approach to understanding the determinants of growth in developing countries, particularly in a way that is relevant to policy, is more care- ful research into three questions: * What are the conditions that initiate an acceleration of growth or the conditions that set off sustained decline? * What happens to growth when policies-trade, macroeconomic, invest- ment-or politics change dramatically in episodes of reform? * Why have some countries absorbed and overcome shocks with little impact on growth, while others have been completely overwhelmed? 248 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Appendix 1. Country Acronyms and Names Code Country name Code Country Name Code Country Name AGO Angola GNB Guinea-Bissau NIC Nicaragua ARG Argentina GRC Greece NLD Netherlands AUS Australia GTM Guatemala NPL Nepal AUT Austria GUY Guyana NZL New Zealand BDI Burundi HKG Hong Kong PAK Pakistan BEL Belgium HND Honduras PAN Panama BEN Benin HTI Haiti PER Peru BGD Bangladesh HVO Burkina Faso PHL Philippines BOL Bolivia IDN Indonesia PNG Papua New Guinea BRA Brazil IND India PRT Portugal BRB Barbados IRL Ireland PRY Paraguay BWA Botswana IRN Iran, Islamic Rep. of RWA Rwanda CAF Central African Republic IRQ Iraq SAU Saudi Arabia CAN Canada ISL Iceland SEN Senegal CHE Switzerland ISR Israel SGP Singapore CHL Chile ITA Italy SLE Sierra Leone CHN China JAM Jamaica SLV El Salvador CIV C6te d'lvoire JOR Jordan SOM Somalia CMR Cameroon JPN Japan SUR Suriname COG Congo KEN Kenya SWE Sweden COL Colombia KOR Korea, Rep. of SWZ Swaziland CRI Costa Rica LIB Liberia SYR Syrian Arab Republic CYP Cyprus LKA Sri Lanka TGO Togo DEU Germany LSO Lesotho THA Thailand DNK Denmark MAR Morocco TTO Trinidad and Tobago DOM Dominican Republic MDG Madagascar TUN Tunisia DZA Algeria MEX Mexico TUR Turkey ECIJ Ecuador MLI Mali TWN Taiwan (China) EGY Egypt MLT Malta TZA Tanzania ESP Spain MMR Myanmar UGA Uganda ETH Ethiopia MOZ Mozambique URY Uruguay FIN Finland MRT Mauritania USA United States FRA France MUS Mauritius VEN Venezuela GAB Gabon MWI Malawi ZAF South Africa GBR United Kingdom MYS Malaysia ZAR Zaire GHA Ghana NAM Namibia ZMB Zambia GIN Guinea NER Niger ZWE Zimbabwe GMB Gambia, The NGA Nigeria Pritchett 249 REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Alesina, Alberto, and Roberto Perotti. 1993. "Income Distribution, Political Instability, and Investment." NBER Working Paper Series 4486. National Bureau of Economic Research, Cambridge, Mass. Processed. Andrews, Donald W. K. 1989. "Power in Econometric Applications." Econometrica 57(5):1059-90. Barro, Robert J. 1997. Determinants of Economic Growth. Cambridge, Mass.: MIT Press. Barro, Robert J., and Xavier Sala-i-Martin. 1995, Economic Growth. New York: McGraw Hill. Ben-David, Dan, and David Papell. 1997. "Slowdowns and Meltdowns: Post-war Growth Evidence from 74 Countries." NBER Working Paper Series 6266. National Bureau of Economic Research, Cambridge, Mass. Processed. Bruno, Michael, and William Easterly. 1998. "Inflation Crises and Long-run Growth." Journal of Monetary Economics 41(1):2-26. Deininger, Klaus, and Lyn Squire. 1996. "A New Data Set Measuring Income Inequal- ity." The World Bank Economic Review 10(September):565-91. Easterly, William, Michael Kremer, Lant Pritchett, and Lawrence H. Summers. 1993. "Good Policy or Good Luck: Country Growth Performance and Temporary Shocks." Journal of Monetary Economics 32(December):459-83. Edwards, Sebastian. 1989. Real Exchange Rates, Devaluation, and Adjustment: Exchange Rate Policy in Developing Countries. Cambridge, Mass.: MIT Press. Forbes, Kristin. 1997. "A Reassessment of the Relationship between Inequality and Growth." Department of Economics, Massachusetts Institute of Technology, Cam- bridge, Mass. Processed. Gallup, John Luke, and Jeffrey Sachs. 1999. "Geography and Economic Growth." In Boris Pleskovic and Joseph E. Stiglitz, eds., Annual World Bank Conference on Devel- opment Economics 1998. Washington, D.C.: World Bank. Haggard, Stephan, Kim Byung-Kook, and Chung-in Moon. 1990. "The Transition to Export-Led Growth in South Korea, 1954-66." Policy Research Working Paper 546. Development Research Group. World Bank, Washington, D.C. Processed. Hall, Robert E., and Charles Jones. 1999. "Why Do Some Countries Produce So Much More Output per Worker Than Others?" Quarterly Journal of Economics 14(1):83- 116. Isham, Jonathan, Daniel Kaufman, and Lant Pritchett. 1997. "Civil Liberties, Democ- racy, and the Performance of Government Projects." The World Bank Economic Re- view 11(May):219-42. Islam, Nazrul. 1995. "Growth Empirics: A Panel Data Approach." Quarterly Journal of Economics 110(November):1127-70. Kamin, Steven. 1988. "Devaluation, External Balance, and Macroeconomic Performance: A Look at the Numbers." International Finance Section, Princeton University, Princeton, N.J. Processed. Kelley, Allen C., and Robert Schmidt. 1994. Population and Income Change: Recent Evidence. World Bank Discussion Paper 249. Washington, D.C.: World Bank. 250 THE WORLD BANK ECONOMIC REVIEW. VOL. 14, NO. 2 Knack, Stephen, and Philip Keefer. 1997. "Does Social Capital Have an Economic Pay- off? A Cross-country Investigation." Quarterly Journal of Economics 112(Novem- ber):1251-88. Krueger, Anne 0. 1978. Liberalization Attempts and Consequences. Cambridge, Mass.: Ballinger for the National Bureau of Economic Research. Levine, Ross. 1997. "Financial Development and Economic Growth: Views and Agenda." Journal of Economic Literature 35(June):688-726. Levine, Ross, and Sarah Zervos. 1998. "Stock Markets, Banks, and Economic Growth." American Economic Review 88(June):537-58. Lucas, Robert E., Jr. 1988. "On the Mechanics of Economic Development: W. A. Mack- intosh Lecture 1985." Journal of Monetary Economics 22(1):3-42. Maddison, Angus. 1995. Monitoring the World Economy, 1820-1992. Paris: Develop- ment Centre of the Organisation for Economic Co-operation and Development. Mankiw, Gregory N., David Romer, and David Weil. 1992. "A Contribution to the Empirics of Economic Growth." Quarterly Journal of Economics 107(May):407-38. Persson, Torsten, and Guido Tabellini. 1994. "Is Inequality Harmful for Growth? " Ameri- can Economic Review 84(June):600-21. Pritchett, Lant. 1996a. "Mind Your P's and Q's: The Cost of Public Investment Is Not the Value of Public Capital." Policy Research Working Paper 1660. Development Re- search Group, World Bank, Washington, D.C. Processed. . 1996b. "Where Has All the Education Gone?" Policy Research Working Paper 1581. Development Research Group, World Bank, Washington, D.C. Processed. .1997. "Our Dearly Departed: A Requiem for Growth Regressions." Paper pre- sented at the 1997 American Economic Association meetings. January. Quah, Danny T. 1996. "Twin Peaks: Growth and Convergence in Models of Distribution Dynamics." Economic journal 106(July):1045-55. Rodrik, Dani. 1996. "Understanding Economic Policy Reform." Journal of Economic Literature 34(March):9-41. . 2000. "Where Did All the Growth Go?: External Shocks, Social Conflict, and Growth Collapses." Forthcoming in the Journal of Economic Growth. Romer, Paul. 1993. "Two Strategies for Economic Development: Using Ideas and Pro- ducing Ideas." In Lawrence Summers and Shekhar Shah, eds., Proceedings of the World Bank Annual Bank Conference on Development Economics 1992. Washington, D.C.: World Bank. Sala-i-Martin, Xavier. 1997. "I Just Ran Two Million Regressions." American Economic Review, Papers and Proceedings 87(2):178-83. Vavamkidis, Athanasios. 1997. "International Integration and Economic Growth." Un- published Ph.D. dissertation. Department of Economics, Harvard University, Cam- bridge, Mass. World Bank. 1997. World Development Report 1997. New York: Oxford University Press. THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2: 251-85 Macroeconomic Fluctuations in Developing Countries: Some Stylized Facts Pierre-Richard Agenor, C. John McDermott, and Eswar S. Prasad This article documents the main stylized features of macroeconomic fluctuations for 12 developing countries. It presents cross-correlations between domestic industrial output and a large group of macroeconomic variables, including fiscal variables, wages, infla- tion, money, credit, trade, and exchange rates. Also analyzed are the effects of economic conditions in industrial countries on output fluctuations in the sample developing coun- tries. The results point to many similarities between macroeconomic fluctuations in de- veloping and industrial countries (procyclical real wages, countercyclical variation in government expenditures) and some important differences (countercyclical variation in the velocity of monetary aggregates). Their robustness is examined using different detrending procedures. Understanding and distinguishing among the factors that affect the short- and long-run behavior of macroeconomic time series have been among the main ar- eas of recent research in quantitative macroeconomic analysis. Using a variety of econometric techniques, a substantial body of literature has documented a wide range of empirical regularities in macroeconomic fluctuations and business cycles across countries. These stylized facts have often been used as an empirical basis for formulating theoretical models of the business cycle and as a way to discrimi- nate among alternative classes of models. Most of the new research in this area has focused on industrial countries, paying less attention to developing countries.' At least two factors may account for this. First, limitations on the quality and frequency of data may be constrain- ing factors. For instance, quarterly data on national accounts are available for only a handful of developing countries, and even where they are available, they are considered to be of significantly lower quality than annual estimates. Second, developing countries tend to be prone to sudden crises and marked gyrations in 1. For an overview of the literature on industrial countries, see, for example, Backus and Kehoe (1992), Fiorito and Kollintzas (1994), and van Els (1995). Pierre-Richard Agenor is lead economist and director of the Macroeconomic and Financial Management Program at the World Bank, C. John McDermott is an advisor at the Reserve Bank of New Zealand, and Eswar Prasad is a senior economist at the International Monetary Fund. Their e-mail addresses are pagenor@worldbank.org, mcdermottj@rbnz.govt.nz, and eprasad@imf.org. The authors would like to thank Nadeem Haque, Alexander Hoffmaister, Philip Lane, James Nason, and Julio Santaella for helpful discussions and comments; Marianne Baxter for the computer code for the band-pass filter; Brooks Calvo for excellent research assistance; and three anonymous referees for their comments. 2000 The International Bank for Reconstruction and Development / THE WORLD BANK 251 252 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 macroeconomic variables, often making it difficult to discern any type of cycle or economic regularity. At the same time documenting the stylized facts on macroeconomic fluctua- tions in developing countries could be useful for a number of reasons. Such an exercise could be valuable for analyzing whether similar empirical regularities are observed across countries with different income levels. Differences in the types of reduced-form relationships observed in industrial countries could provide an empirical basis for constructing analytical models of short-run fluctuations that incorporate features particularly important to developing countries. In addition, as argued, for instance, by Ag6nor and Montiel (1996), these findings may have important policy implications. They may, for example, be crucial for designing stabilization and adjustment programs. A burgeoning literature has begun to document these stylized facts for devel- oping countries. Some of the studies focus on specific stylized facts and construct theoretical models that can replicate those facts. Mendoza (1995), for instance, documents a strong positive correlation between terms-of-trade and output fluc- tuations in developing countries. Other studies in this genre include Kouparitsas (1997) and Kose and Riezman (1998), although these articles focus on one or two specific sets of bivariate correlations. Another set of articles documents a broader set of cross-correlations, but typically only for one country. Further, most use only one detrending procedure. Representative papers include Kydland and Zarazaga's (1997) work on Argentina and Rodriguez-Mata's (1997) analy- sis of fluctuations in Costa Rica. This article builds on the existing literature by systematically documenting a wide range of regularities in macroeconomic fluc- tuations for a large group of developing countries. We chose the countries in our sample on the basis of several considerations. The first was the desire to select a group of countries for which we could as- semble data of reasonable quality, thereby addressing the criticism that such ex- ercises have limited validity because of data inaccuracies. The second consider- ation was the need to include different geographic areas and a wide range of macroeconomic experiences, at the same time selecting countries that did not suffer substantial economic turmoil (in the form of, say, sustained episodes of hyperinflation) over the relevant sample period. With this criterion we avoid crisis-prone countries and the difficulties associated with data interpretation in such cases. Moreover, by looking for a consistent set of relationships among macroeconomic variables in a relatively large group of countries that have had diverse experiences with structural change, we provide a set of stylized macro- economic facts that are unlikely to reflect country-specific episodes. Our study of business cycle regularities is based on quarterly data for a group of 12 middle-income countries: Chile, Colombia, India, the Republic of Korea, Malaysia, Mexico, Morocco, Nigeria, the Philippines, Tunisia, Turkey, and Uru- guay. On the one hand, the decision to use quarterly, rather than annual, data imposes an additional constraint on the size of our sample, because relatively few developing countries produce quarterly output indicators. On the other hand, Age'nor, McDermott, and Prasad 253 quarterly data provide us with sufficiently long time series for reliable statistical inference.2 The data cover a wide range of macroeconomic variables and include indus- trial output, prices, wages, monetary aggregates, domestic private sector credit, fiscal variables, exchange rates, and trade variables. (See the appendix for a de- scription of the data and sources.) Thus we are able to examine macroeconomic fluctuations in various dimensions, in contrast to earlier studies. In addition, we examine the relationship between economic fluctuations in these countries and two key indicators that proxy for economic activity in industrial countries-an index of industrial-country output and a measure of the world real interest rate. Two methodological aspects of this article are worth highlighting at the out- set. First, in line with the recent literature on business cycles for industrial coun- tries, many of the results discussed in the article are based on unconditional correlations between different variables. We naturally recognize that such cor- relations do not imply causal relationships and, in some cases, attempt to comple- ment our correlation results by examining bivariate exogeneity tests. We also recognize that reduced-form relationships between certain variables depend crucially on the sources of macroeconomic shocks. Nevertheless, our results are useful in that they indicate the types of shocks that could be important for different countries and set the stage for more formal structural models of busi- ness cycle fluctuations. Second, many of the macroeconomic series used in this article have distinct trends over time and, hence, need to be rendered stationary prior to empirical analysis. Empirical results could, of course, be sensitive to the choice of econo- metric procedure used to remove long-term trends from the data and derive cycli- cal components. This article makes an additional methodological contribution by examining the sensitivity of correlations and other stylized facts to the detrending procedure used. We use two detrending techniques: a modified ver- sion of the Hodrick-Prescott (1997; HP) filter developed by McDermott (1997) and the band-pass (BP) filter proposed by Baxter and King (1995).3 Thus this article's main contribution is to document a comprehensive set of stylized facts that are comparable across countries and to examine their sensitiv- ity to different detrending techniques. Consistent with the work of other authors (such as Mendoza 1995), we find that output volatility is greater in developing countries than in industrial countries, terms-of-trade and output fluctuations are strongly positively correlated, and there is no consistent relationship between the 2. There are two additional considerations in choosing quarterly rather than annual data. First, some of the series we use have been readily available (and comparable across countries) for only a limited time. For instance, our data on effective exchange rates have been published by the International Monetary Fund only since 1978. Second, establishing large enough samples on an annual basis would imply going back to the early 1960s. It is likely that the quality of the data, where available, was substantially lower in those earlier years. 3. In Agenor, McDermott, and Prasad (1998) we provide robustness checks for our results using two other detrending techniques-first differences and a nonparametric technique. 254 TIIE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 trade balance and fluctuations in domestic output. As in Kouparitsas (1996), we find some evidence that output fluctuations in developing countries are posi- tively correlated with business cycles in industrial countries and negatively corre- lated with real interest rates in industrial countries. We also find evidence of procyclical variation in monetary aggregates and real wages and countercyclical variation in government expenditures. The remainder of the article is organized as follows. Section I briefly describes the detrending procedures used. Section II describes a number of economic fea- tures of the countries included in the data set and presents summary statistics for the behavior of output. Section III provides a more rigorous characterization of macroeconomic fluctuations in these countries and contrasts the results with avail- able stylized facts of business cycles in industrial and developing countries. Sec- tion IV summarizes the main results of the article. Section V offers some final remarks and suggestions for further empirical and theoretical analysis. I. UNIVARIATE DETRENDING TECHNIQUES As indicated earlier, the objective of our article is to examine economic fluc- tuations at business cycle frequencies rather than to study longer-term growth.4 To do so, it is necessary to decompose all of our macroeconomic series into nonstationary (trend) and stationary (cyclical) components, because certain em- pirical characterizations of the data, including cross-correlations, are valid only if the data are stationary. For a given series, in finite samples, stationary components obtained using different filters can often display very different time-series properties (see Canova 1998). In this article we take an agnostic approach and report results obtained using the two filters mentioned above. The variant of the HP filter we use here chooses the smoothing parameter optimally for each series rather than imposing the same exogenous smoothing parameter for all series (see McDermott 1997).5 II. KEY CHARACTERISTICS OF SAMPLE COUNTRIES In this section we describe a number of important economic features of the developing countries in our sample that are relevant for our analysis. In addition, we present summary statistics for output and inflation and provide a preliminary characterization of business cycle fluctuations in our group of countries. We also compare the properties of business cycles in these countries with those observed in industrial countries. The sample period for most of the data series used in this study runs from the first quarter of 1978 to the fourth quarter of 1995. The data sources are described in detail in the appendix. 4. The real business cycle literature makes no clear distinction between trend and cycles since both short- and long-term fluctuations are regarded as being driven by the same stochastic process. 5. A detailed discussion of the detrending techniques and the algorithms for these filters, along with a discussion of their properties, can be found in Ag6nor, McDermott, and Prasad (1988). Agenor, McDermott, and Prasad 255 Most of the countries in our sample could be reasonably characterized as middle- income countries. Although India and Nigeria have relatively low per capita in- comes, we include them in the sample because they are among the largest market economies in Asia and Africa (figure la). The urbanization rate and the propor- tion of agricultural output as a share of gross domestic product (GDP) indicate that agriculture is an important, but not dominant, sector in most of the sample (figures lb and lc). Because we were unable to obtain reliable quarterly GDP data for all of the countries in our sample, we use indexes of industrial output to construct mea- sures of the aggregate business cycle. The manufacturing sector accounts for a significant fraction of total GDP (figure ld). Except for Nigeria, this share is more than 15 percent for all countries in our sample, compared with an average share of 25 to 30 percent for most industrial countries. In addition, because output in the industrial sector roughly corresponds to output in the traded goods sector (excluding primary commodities) and is most closely related to what are tra- ditionally thought of as business cycle shocks, either exogenous or policy- determined, we argue that this variable is a reasonable proxy for measuring the aggregate cycle.6 For all countries except Nigeria, export growth is an important contributor to overall GDP growth (figure lh). Standard measures of openness to international trade-as indicated by the average openness ratio (the ratio of the sum of im- ports and exports to GDP)-illustrate the importance of foreign trade transac- tions in our sample (figure 1i). Hence an important part of our analysis focuses on the relationship between the domestic business cycle and the prices and quan- tities related to international trade. An important consideration in choosing our sample was to exclude countries that had suffered sustained episodes of hyperinflation during the period under study. Although some of the countries in the sample (such as Mexico, Turkey, and Uruguay) had high levels of inflation over the past two decades, none suf- fered sustained episodes of hyperinflation (figure lk). This is also apparent from the average annual rates of consumer price inflation and the volatility of infla- tion, as measured by the standard deviation of annual inflation rates (last two columns of table 1). A key issue concerning business cycle fluctuations in developing countries is whether aggregate fluctuations are characterized by basic time-series properties, such as volatility and persistence, that are similar to those observed in industrial countries. A simple way of approaching this issue is by examining summary sta- tistics for the stationary components of industrial output (table 1). The first two columns of table 1 report means and standard deviations of output growth rates as well as standard deviations of the cyclical components of output derived using 6. In general, the use of GDP data for measuring business cycle activity in a developing country can be problematic. Agriculture, which still accounts for a large share of aggregate output in many developing countries (including several in our sample) is influenced more by weather conditions than by cyclical factors. Poor measurement of services and informal sector activities may also impart significant biases. 256 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Figure 1. Economic Indicators in Selected Developing Countries (data are for 1993, unless otherwise indicated) Figure lI. Per capita income (U.S. dollar.s) Figure lb. Urbanization rate (percent) Chile _ Chile Cololbia Colombia India India Korea, Rep. of Korea, Rep. of Malaysia Malaysia Mexico Mexico Morocco Morocco Nigeria Nigeria Philippines Philippines Tainisia Tunisia T-irkey Turkey Uruguay Urtiguay __ 0 2,0(00 4,000 6.000 800 ( 20 40 60 80 100 Figurtw Ic. Agriculture/GDP (percent) Figure ld. Manufacturing/GDP (percent) Chile Chile Colombia Colombia India India Korea, Rep. of Korea, Rep. of Malaysia Malaysia Mexico _Mexico Morocco Morocco Nigeria, Nigeria ___ Philippines Philippines Tunisia Tunisia Turkey Turkey t.lrlguay_ Uruguiay _. 0 10 20 30 40 0 S 10 15 20 25 30 35 Figure le. Total government expenditures/GEP (percent) Figure If TFotal gosvennanent revenue/GDP (percent) Chile Chile Colombia Colombia India India Korea, Rep. of Korea, Rep. of Malaysia Malaysia Mtexico Mexico Morocco Morocco Nigeria Nigeria Philippines Philippines Ttinisia Tunisia Turkey Turkey Uruguay tFlrugiay 0 10 20 30 40 0 5 10 15 20 25 30 35 Agenor, McDermott, and Prasad 257 Figure 1. (continued) Figure Ig. Import growth (percent)' Figure Ih, Export growth (ptrcentr` Chile hChile _ Colombia - Colrnmbia - India - India - Korea, Rep. of - Korea, Rep, of - Malaysia Nlaysia Mexico _Mexico - Morocco _ Morocco _ Nigeria Nigera - Philippines - Philippines - Tunisia Tutnisia - Turkey Turkey _ Uruguay L,rLguay -15 -10 -; 0 5 10 15 5 0 5 10 15 Figure Ij. Exteral debt service/exports of goods Figure Ii. Openness (percent) and services (percent) Chile Chile Colombia Colombia India India Korea, Rep. of Korea, Rep. of Malaysia Malaysia Mexico Mexico Morocco Morocco Nigeria Nigeria Philippines Philippines Tunisia Tnisia Turkey Turkey Uruguay Uruguay 0 50 C100 150 0 1(1 20 30 40 50 60 70 Figure lk. Consumer price inflation (percenrn Figure 11. Broad money growth (percent) Chile Chile Colombia Colombia India India Korea, Rep of Korea, Rep. of Malaysia Malaysia Mexico Mexico Morocco Morocco Nigeria Nigeria Philippines Philippines Tunisia Turnsia Turkey Turkey Uruguay . e._._._._._._._._._. Uruguay 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 a. Average annual growth mute, 1980-93. b Average annual ratio of swun of exports and impris to tGDP, in percent, 1')8G93. Soure:- Intemabonal Monetarv Find and World Bank. 258 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 1. Summary Statistics for Industrial Output and Inflation Output Annual inflation Country and Mean Standard Autocorrelations Mean Standard filter (percent) deviation Lag 1 Lag 2 Lag 3 Lag 4 (percent) deviation Chile Growth 3.81 7.03 0.75 0.53 0.25 -0.07 19.69 8.69 HP 4.53 0.68 0.51 0.27 0.00 BP 1.45 0.56 0.54 0.42 0.25 Colombia Growth 2.57 4.61 0.70 0.52 0.39 0.15 24.19 4.02 HP 2.33 0.51 0.27 0.17 0.02 BP 1.40 0.63 0.65 0.59 0.49 India Growth 6.02 4.31 0.67 0.54 0.27 0.0S 9.34 2.78 HP 2.45 0.48 0.35 0.10 0.02 BP 1.13 0.24 0.49 0.28 0.27 Korea, Rep. of Growth 9.22 6.06 0.75 0.49 0.22 -0.11 8.23 7.34 HP 3.47 0.71 0.44 0.20 -0.14 BP 1.48 0.67 0.57 0.61 0.37 Malaysia Growth 9.22 6.79 0.71 0.29 -0.04 -0.29 3.79 2.55 HP 4.06 0.69 0.30 0.07 -0.16 BP 1.41 0.46 0.23 0.41 0.15 Mexico Growth 2.39 6.21 0.79 0.51 0.24 0.00 48.55 40.36 HP 3.31 0.72 0.40 0.14 -0.13 BP 1.42 0.76 0.64 0.53 0.30 Morocco Growth 2.57 4.44 0.12 0.27 0.06 -0.32 7.03 3.38 HP 2.77 0.06 0.25 0.08 -0.18 BP 1.14 0.01 0.43 0.33 0.13 Nigeria Growth 3.05 12.34 0.62 0.33 0.17 0.00 29.54 23.56 HP 6.69 0.45 0.09 -0.06 -0.12 BP 3.29 0.50 0.43 0.44 0.40 Philippines Growth 13.85 11.69 0.63 0.37 0.03 -0.29 13.69 11.48 HP 7.45 0.63 0.42 0.10 -0.15 BP 2.62 0.18 0.41 0.19 0.04 Tunisia Growth 2.34 4.79 0.77 0.57 0.30 0.13 7.50 2.35 HP 2.72 0.63 0.42 0.13 0.06 BP 1.25 0.61 0.70 0.44 0.46 Turkey Growth 6.19 6.14 0.48 0.27 0.11 -0.23 61.78 25.45 HP 3.67 0.38 0.14 0.06 -0.12 BP 1.42 -0.08 0.20 0.20 0.07 Uruguay Growth -0.94 8.55 0.72 0.55 0.34 0.04 62.04 23.93 HP 4.94 0.63 0.50 0.27 -0.01 BP 2.37 0.62 0.75 0.63 0.53 Note: Growth refers to the four-quarter differences of the log levels of relevant variables (as in text). HP and BP refer to the stationary components of output derived using the modified Hodrick-Prescott and band-pass filters, respectively. Source: Authors' calculations based on IMF data. Agenor, McDermott, and Prasad 259 the HP and BP filters.7 Growth rates are measured here as four-quarter differences of the log levels of the relevant variables. Mean annual growth rates of industrial output over the past two decades var- ied substantially across the countries in our sample, ranging from almost 14 per- cent for the Philippines to about 2.5 percent for Colombia, Mexico, Morocco, and Tunisia. Uruguay, in fact, recorded a negative mean growth rate over this period. The volatility of growth rates also varies markedly across countries. On average, volatility in our sample is much higher than the level typically observed in industrial countries. A similar picture emerges from the standard deviations of the filtered cyclical components of industrial output.8 Because the filters used here tend to eliminate more of the low-frequency variation than, say, a first-difference filter, these stan- dard deviations are generally lower. However, the ordering of countries by their cyclical volatility is similar, and their volatility is generally higher than that ob- served for industrial countries. The volatility of the cyclical components obtained using the BP filter is generally much lower than that using the HP filter; the BP filter eliminates some of the high-frequency variation in the data, whereas the HP filter eliminates only low-frequency variation. To examine the persistence of business cycle fluctuations, we also measure the first four autocorrelations of the filtered series (table 1). The autocorrelations are generally strongly positive, indicating considerable persistence in the cyclical com- ponents. These results suggest that it is appropriate to view the developing coun- tries in our sample as having short-term fluctuations that could reasonably be characterized as business cycles. III. MAIN FEATURES OF MACROECONOMIC FLUCTUATIONS We measure the degree of comovement of a series Yt with industrial output x: by the magnitude of the correlation coefficient A(j), j E (O, +1, +2, . . .). These correlations are between the stationary components of Yt and x,, with compo- nents in both derived using the same filter. In the discussion that follows, we consider the series y, to be procyclical, acyclical, or countercyclical if the contem- poraneous correlation coefficient A(0) is positive, zero, or negative, respectively. In addition, we deem the series yt to be strongly contemporaneously correlated if 0.26 S 1A(0)1 < 1, weakly contemporaneously correlated if 0.13 1A(0)1 < 0.26, and contemporaneously uncorrelated with the cycle if 0 1A(0)1 < 0.13.9 7. These filters, by construction, deliver stationary components that have zero means. The output series as well as all the other time series used in this article were deseasonalized using the X-11 procedure. 8. We can interpret these standard deviations as quarterly percentage standard deviations. For purposes of comparison, the standard deviation of HT-filtered postwar quarterly industrial production for the United States is about 2 percent. 9. The approximate standard error of these correlation coefficients, computed under the null hypothesis that the true correlation coefficient is zero and given the average number of observations per country in our sample, is about 0.13. 260 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 The cross-correlation coefficients A(j), j E {O, 1, 2, ...} indicate the phase- shift of yt relative to the cycle in industrial output. We say that Yt leads the cycle by j periods if IA(j)I is a maximum for a positive j, is synchronous with the cycle if IA(j)l is a maximum for j = 0, and lags the cycle if IA(j)l is a maximum for a negative j. To conserve space, we report only contemporaneous correlations and correlations at the fourth and eighth lags and leads. Results for a larger set of lags and leads can be found in Agenor, McDermott, and Prasad (1998). Correlations with Industrial-Country Business Cycles Here we examine the relationship between fluctuations in domestic industrial output in our sample countries and variables that represent economic activity in the main industrial countries-a relationship that could be particularly impor- tant for developing countries that have substantial trade links with industrial countries.10 As discussed earlier, the magnitude of the links between macroeco- nomic fluctuations in industrial and developing countries and the channels through which shocks propagate between these two sets of countries are of considerable interest from a number of different perspectives. The contemporaneous correlations are positive for a majority of the sample countries, indicating that business cycle fluctuations in developing countries tend to be correlated with business cycle fluctuations in industrial countries (table 2). For many of the countries that have positive contemporaneous correlations, the correlations generally peak at or near a zero lag, suggesting that output fluctua- tions in industrial economies are transmitted fairly quickly." These results are generally robust across filters, barring a couple of excep- tions. For instance, in the case of Mexico the BP filter yields a strong negative contemporaneous correlation, whereas the HP filter yields a positive correlation. The correlations at the four-quarter lag are, however, all strongly positive, indi- cating that industrial-country output has a lagged effect on Mexican output. The contemporaneous correlations are close to zero for Morocco and Nigeria and marginally negative for Turkey. For these countries there is some evidence that industrial-country output has a positive effect on domestic industrial output with a lag of about four to eight quarters. Business cycle conditions in industrial economies also could influence fluctua- tions in developing economies through the world real interest rate. The world real interest rate is likely to have an important effect on economic activity in the developing world, not only because it affects domestic interest rates, but also because it reflects credit conditions in international capital markets. These capi- tal markets may be especially important for developing countries (even those in the middle-income range) that do not have well-developed domestic capital mar- kets. To examine this issue, we measure correlations of industrial output in our 10. The industrial-country variables used in this section are described in the appendix. 11. Business cycles in the industrial economies are, of course, not perfectly synchronized. But Lumsdaine and Prasad (1997), among others, argue that there is a substantial common component in business cycle fluctuations across the main industrial economies. Age'nor, McDermott, and Prasad 261 Table 2. Cross Correlations between Domestic Output and Industrial-Country Output Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP -0.04 0.09 0.52 -0.11 -0.48 BP 0.24 0.38 0.32 -0.03 -0.28 Colombia HP -0.44 -0.07 0.43 0.21 -0.12 BP -0.37 0.02 0.49 0.55 0.35 India HP 0.13 0.15 0.24 -0.20 -0.11 BP 0.32 0.56 0.46 0.07 -0.15 Korea, Rep. of HP 0.00 -0.49 0.36 0.22 0.22 BP 0.28 0.07 0.36 0.66 0.49 Malaysia HP -0.49 0.14 0.59 0.08 -0.40 BP -0.29 0.30 0.57 0.21 -0.35 Mexico HP 0.08 0.38 0.19 -0.62 -0.37 BP 0.36 0.35 -0.29 -0.83 -0.70 Morocco HP 0.10 -0.07 -0.06 0.02 -0.03 BP 0.20 0.06 -0.05 -0.06 -0.17 Nigeria HP 0.22 0.26 0.03 -0.22 -0.05 BP 0.59 0.31 -0.15 -0.41 -0.38 Philippines HP -0.63 -0.05 0.53 0.10 -0.48 BP -0.32 -0.11 0.38 0.26 -0.36 Tunisia HP -0.43 -0.24 0.45 0.13 -0.36 BP 0.80 -0.48 0.04 0.07 -0.27 Turkey HP 0.23 0.04 -0.14 -0.02 0.12 BP 0.24 -0.32 -0.36 -0.16 0.43 Uruguay HP 0.33 0.08 0.21 -0.30 -0.29 BP 0.66 0.49 0.18 0.01 0.17 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescort and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of industrial-country output, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. sample countries with a weighted index of real interest rates in the major indus- trial countries (table 3). For most of the countries in our sample the contemporaneous correlations between HP-filtered output and the world real interest rate are positive. This could reflect the facts that the real interest rate in industrial economies tends to be procyclical and that changes in industrial-country output, through trade links, 262 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 3. Cross Correlations between Domestic Output and the World Real Interest Rate Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP 0.06 -0.43 0.17 0.10 0.06 BP -0.18 -0.42 -0.25 0.06 0.16 Colombia HP -0.50 -0.33 0.22 0.32 0.33 BP -0.55 -0.13 0.24 0.58 0.39 India HP 0.00 0.14 0.29 -0.28 -0.28 BP -0.08 0.14 0.09 -0.36 -0.25 Korea, Rep. of HP 0.12 -0.28 0.34 -0.08 0.08 BP -0.21 -0.08 0.05 0.16 0.06 Malaysia HP -0.03 -0.18 0.18 -0.09 -0.04 BP 0.56 -0.02 -0.12 -0.24 -0.29 Mexico HP -0.32 0.00 0.22 -0.03 -0.14 BP 0.11 0.11 0.09 -0.02 -0.06 Morocco HP 0.24 0.18 -0.16 0.00 -0.04 BP 0.32 0.23 -0.06 -0.25 -0.22 Nigeria HP -0.23 0.32 -0.01 -0.05 0.22 BP -0.04 0.01 0.08 0.05 -0.17 Philippines HP -0.15 -0.08 0.26 0.06 -0.34 BP -0.04 0.07 0.21 0.17 -0.41 Tunisia HP 0.26 -0.25 0.04 0.07 -0.14 BP 0.27 0.13 0.10 -0.01 -0.32 Turkey HP 0.13 0.02 -0.22 -0.05 0.22 BP 0.05 -0.25 -0.46 -0.04 0.44 Uruguay HP -0.10 -0.32 0.19 0.13 -0.02 BP -0.20 -0.35 -0.19 0.04 0.22 Note: The world real interest rate is proxied by a weighted index of real interest rates in the major industrial countries. HP and BP refer to the stationary components derived using the modifed Hodrick- Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of the world real interest rate, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. have positive spillover effects on output in these middle-income countries.12 Morocco and Turkey are the only sample countries for which this correlation is negative using either filter. For a few countries the lagged correlations are nega- tive. Mexico is an interesting case: the contemporaneous correlation is positive, 12. The correlation between the cyclical components of the output and real interest rate indexes for industrial countries is strongly positive for 1975-95, irrespective of the derrending procedure used. Agenor, McDermott, and Prasad 263 but most of the correlations at short leads and lags are close to zero, indicating that the effects of changes in the world interest rate are transmitted quite rapidly to Mexican industrial output. This is not surprising given Mexico's physical prox- imity to and close trade links with the United States, which is the dominant in- dustrial economy and therefore has a high weight in the composite index of industrial-country output and our proxy for the world real interest rate. Overall, these results suggest that the level of economic activity in industrial countries has a significant, positive relationship with industrial output in the middle-income countries in our sample.13 Because real interest rates are procyclical in industrial countries, their relationship to industrial output in developing coun- tries may be muted by the opposite, indirect effect of aggregate economic activity in industrial countries. Further research is needed to separate out the quantita- tive importance of these different influences on business cycle propagation. An important issue in this context (which we return to later) is the inability to mea- sure interest rates that individual countries face on world capital markets with- out adequate data on country-specific premiums. Cyclical Behavior of Public Sector Variables The relationship between fluctuations in aggregate output and the compo- nents of aggregate demand has been well documented for industrial countries. Unfortunately, we were unable to obtain consistent and sufficiently long series of quarterly data on consumption and investment for all countries in our sample. We were, however, able to obtain data on the public sector, although only for a limited set of countries. Examining the relationship between aggregate economic activity and public sector expenditures and revenue has analytical value from the perspective of business cycle modeling and is important from a policy perspec- tive, including in the design of macroeconomic stabilization programs. There is a robust negative relationship between government expenditures and the domestic business cycle in all four countries for which we have data-Chile, Korea, Mexico, and the Philippines (top panel of table 4). Thus there is fairly clear evidence of a countercyclical role for government expenditures. These re- sults contrast with those obtained for industrial countries. Fiorito and Kollintzas (1994), for instance, find no clear pattern. The negative contemporaneous corre- lation between government consumption expenditures and industrial output is consistent with the prediction of a variety of models, such as the class of intertemporal optimizing models with imperfect capital mobility and flexible prices (see Agenor 1997). In these models an increase in public spending leads to a net increase in domestic absorption (if the degree of intertemporal substitution in consumption is not too large), a real exchange rate appreciation, and a fall in output of tradable goods on impact. Government revenues are significantly countercyclical in Korea, the Philip- pines, and Uruguay (second panel of table 4).14 This negative correlation may 13. This finding is consistent with the results of Kouparitsas (1996). 14. Rodriguez-Mata (1997) establishes a countercyclical pattern of government revenue for Costa Rica. 264 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 4. Cross Correlations between Domestic Output and Government Expenditures, Government Revenue, and the Fiscal Impulse Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Between domestic output and government expenditures Chile HP -0.13 -0.04 -0.16 0.30 0.00 BP -0.27 -0.01 -0.13 0.43 -0.34 Korea, Rep. of HP -0.03 -0.17 -0.39 0.13 0.32 BP -0.14 -0.33 -0.46 -0.02 0.39 Mexico HP -0.10 -0.11 -0.21 0.22 0.36 BP -0.11 -0.35 -0.10 0.21 0.27 Philippines HP 0.59 0.22 -0.72 0.00 0.57 BP 0.69 0.10 -0.54 -0.06 0.25 Between domestic output and government revenue Colombia HP -0.03 0.05 -0.17 0.03 0.23 BP 0.15 0.20 -0.01 -0.03 0.14 Korea, Rep. of HP -0.06 -0.17 -0.28 0.14 0.30 BP -0.18 -0.20 -0.31 0.08 0.41 Mexico HP 0.07 0.21 -0.08 -0.03 0.17 BP -0.11 0.15 0.13 0.15 0.41 Philippines HP 0.49 0.32 -0.69 -0.14 0.45 BP 0.35 0.22 -0.57 -0.22 0.24 Uruguay HP -0.13 -0.09 -0.26 0.28 0.22 BP -0.27 -0.26 -0.13 0.15 0.12 Between domestic output and the fiscal impulse' Korea, Rep. of HP -0.03 -0.09 -0.23 0.00 0.11 BP -0.02 -0.23 -0.27 -0.08 0.15 Mexico HP -0.13 -0.25 -0.16 0.28 0.24 BP -0.07 -0.40 -0.14 0.17 0.10 Philippines HP 0.32 -0.10 -0.30 0.16 0.30 BP 0.46 -0.16 -0.17 0.15 0.10 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of government expenditures, government revenue, or the fiscal impulse, with both variables detrended using the same filter. The data series and sources are described in the appendix. a. The fiscal impulse is defined as the ratio of government expenditures to government revenue. Source: Authors' calculations based on IMF data. Agenor, McDermott, and Prasad 265 result from the negative effects of increases in tax revenues (possibly induced by increases in effective tax rates) on disposable income and aggregate demand.'" In Mexico the relationship appears to be acyclical, although this result is sensitive to the choice of filter. To examine the net effect of government revenue and expenditures on the domestic business cycle, we construct a measure of the fiscal impulse-the ratio of government expenditures to government revenue-for the three countries for which both revenue and expenditure series were available. The fiscal impulse is negatively correlated with the business cycle, either contem- poraneously or at short lags, in Korea, Mexico, and the Philippines (third panel of table 4). Thus the fiscal impulse is countercyclical and plays a role in short-run macroeconomic stabilization. To summarize, the correlations examined in this subsection suggest that the government balance does play a significant role in dampening domestic fluctua- tions in Korea, Mexico, and the Philippines. However, the countercyclical be- havior of government revenue in some countries highlights the need to reexamine revenue sources to ensure that they do not exacerbate domestic fluctuations. An alternative possibility is that tightening government finances could raise future output growth by, for instance, crowding in private investment or by signaling the future stability of domestic macroeconomic policy, thereby stimulating for- eign investment. Based on the negative lagged correlations, there is some evi- dence of this effect in Korea. Correlations with Wages and Prices Establishing stylized facts about the cyclical behavior of wages and prices has important implications for discriminating among different classes of models (based on their predictions concerning such behavior). For instance, Keynesian models imply that real wages are countercyclical, whereas equilibrium models of the business cycle imply that real wages are procyclical (Abraham and Haltiwanger 1995). Similarly, the implications of the cyclical behavior of prices, inflation, and (as discussed later) various monetary aggregates for discriminating among differ- ent classes of business cycle models have been the subject of considerable debate in the business cycle literature recently (Chadha and Prasad 1994). We begin by examining correlations between average nominal wages in the industrial sector and industrial output. Consistent time-series data on wages were available for only 5 of the 12 countries in our sample. The cyclical behavior of nominal wages varies markedly across the five countries (first panel of table 5). In Chile nominal wages appear to be procyclical, whereas there is some evidence that nominal wages are countercyclical in Korea, Colombia, and Mexico. In interpreting these results, it is useful to look at the cyclical behavior of real wages, often the relevant wage variable for analyzing business cycles. We con- struct real wages by deflating nominal wages by the consumer price index (cPI). 15. A highly positive short-run correlation between current income and expenditures in developing countries has been well documented and has been attributed to the existence of liquidity constraints and finite horizons. See Agenor and Montiel (1996: ch. 3). 266 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 5. Cross Correlations between Domestic Output and Nominal Wages and between Domestic Output and Real Wages Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Between domestic output and nominal wages Chile HP -0.13 -0.08 0.08 0.03 0.06 BP -0.54 -0.29 0.38 0.53 0.08 Colombia HP -0.05 0.07 -0.13 0.41 0.15 BP -0.03 -0.35 -0.45 -0.47 0.41 Korea, Rep. of HP -0.05 -0.28 -0.08 0.06 0.29 BP -0.58 -0.56 -0.43 -0.12 0.33 Mexico HP 0.16 -0.07 -0.15 -0.17 0.1 BP 0.73 0.20 -0.30 -0.29 -0.18 Turkey HP 0.02 0.32 0.08 -0.20 -0.22 Br 0.47 0.50 0.05 -0.53 -0.36 Between domestic output and real wages Chile HP 0.06 -0.27 0.31 -0.01 0.04 BP -0.15 -0.04 0.15 0.06 0.00 Colombia HP -0.24 0.02 0.68 -0.44 -0.07 BP 0.27 0.09 0.27 -0.43 -0.46 Korea, Rep. of HP -0.33 -0.21 0.38 0.21 -0.11 BP -0.24 0.01 0.32 0.34 0.34 Mexico HP -0.40 -0.26 0.64 0.15 -0.24 BP 0.15 0.28 0.47 0.14 -0.48 Turkey HP 0.07 0.19 0.43 -0.21 0.64 BP 0.38 0.61 0.15 -0.63 -0.64 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of nominal wages or real wages, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. Alternative theories offer different predictions of wage behavior. For instance, traditional Keynesian models of the business cycle posit short-run movement along a stable labor demand schedule and, therefore, predict that real wages are countercyclical. However, real business cycle models, as well as new Keynesian macroeconomic models with imperfect competition and countercyclical mark- ups, predict procyclical wages.16 Finally, efficiency wage models predict no tight contemporaneous relationship between output (employment) and real wages. 16. See Rotemberg and Woodford (1991) for a discussion of macroeconomic models with imperfect competition and countercyclical markups. Agenor, McDermott, and Prasad 267 More generally, as Abraham and Haltiwanger (1995:1230) note, different types of shocks can have very different implications for the cyclicality of the real wage. Technology shocks tend to produce procyclical movements of the real wage, whereas nominal shocks (such as money supply shocks) generate countercyclical movements. The correlations between industrial output and real wages are striking (lower panel of table 5). In all five countries for which data are available, and with both filters, we find strong evidence of procyclical real wage variation. This result is consistent with the implications of real business cycle models that ascribe a domi- nant role to technology shocks that shift the labor demand schedule in the short run. It is also in line with the evidence on real wage rate variation in the United States (see Kydland and Prescott 1994).'7 Next we turn to the correlations between prices and output. A substantial literature documents the countercyclical behavior of prices in industrial econo- mies (see, for instance, Backus and Kehoe 1992, Fiorito and Kollintzas 1994, Kydland and Prescott 1994, and Cooley and Ohanian 1991). Many of these studies argue that the countercyclical behavior of price levels provides support for supply-driven models of the business cycle, including real business cycle mod- els that depict technology shocks as predominant in driving business cycle fluc- tuations. However, Chadha and Prasad (1994) argue that the correlation be- tween inflation and cyclical output is the appropriate correlation for discriminating between demand- and supply-driven models of the business cycle.8 They show that inflation in the G-7 countries has in fact been procyclical during the postwar period. We therefore examine the cyclical behavior of both the price level and the inflation rate. The contemporaneous correlations between industrial output and the aggre- gate consumer price index are generally negative for Colombia, India, Korea, Malaysia, Morocco, Nigeria, and Turkey, indicating countercyclical variation of the price level (table 6). For a few countries, including Chile and Uruguay, how- ever, the correlations are significantly positive. Thus, unlike industrial countries, our sample countries do not show a consistent negative relationship between the stationary components of output and price levels. The contemporaneous correlations between the level of inflation and the cy- clical component of output indicate that there is little strong evidence of procyclical inflation for most countries in our sample, although the lagged correlations are positive for Chile and Uruguay (table 7)19 The correlations at the leads do not 17. Our analysis of real wage cyclicality only considers the consumption wage in the manufacturing sector, not the producer wage. The two measures could display very different behavior over time, as Abraham and Haltiwanger (1995) illustrate for the United States. 18. Also see Judd and Trehan (1995). 19. Although we exclude from the sample countries with sustained hyperinflationary episodes, unit root tests for inflation indicate that, for about half of the countries in the sample, we cannot reject the null hypothesis of nonstationarity. Hence, for all countries we detrend inflation using the same filters that we use for output. We also examine the correlations using the raw series for inflation and filtered output. For the countries with significant correlations (reported in table 7), the choice of filtered or unfiltered inflation does not matter. 268 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 6. Cross Correlations between Domestic Output and the Price Level (Consumer Price Index) Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP -0.41 0.27 0.42 0.03 -0.14 BP -0.03 0.28 0.51 0.15 -0.23 Colombia HP 0.10 -0.06 -0.43 0.14 -0.05 BP -0.37 -0.60 -0.67 -0.46 0.11 India HP -0.27 -0.21 -0.06 0.03 0.21 BP -0.11 -0.56 -0.47 -0.12 0.09 Korea, Rep. of HP 0.00 0.02 -0.26 -0.22 0.27 BP -0,37 -0.45 -0.58 -0.53 -0.18 Malaysia HP 0.13 -0.17 -0.05 -0.15 -0.09 BP 0.10 -0.04 -0.19 -0.34 -0.37 Mexico HP 0.29 0.17 0.46 -0.26 0.23 BP 0.72 0.08 -0.55 -0.50 -0.18 Morocco HP -0.06 -0.05 -0.28 0.01 0.22 BP -0.38 -0.39 -0.39 -0.10 0.21 Nigeria HP 0.01 0.14 -0.23 -0.05 0.16 BP -0.01 -0.14 -0.21 -0.08 0.10 Philippines HP -0.43 -0.62 0.44 0.38 -0.22 BP -0.36 -0.56 0.00 0.47 0.50 Tunisia HP 0.19 0.28 -0.15 0.03 0.03 BP 0.10 0.50 0.37 0.29 -0.08 Turkey HP 0.26 0.24 -0.31 -0.15 0.13 BP 0.55 0.15 -0.47 -0.43 0.17 Uruguay HP -0.18 0.47 0.40 -0.27 -0.48 BP 0.22 0.44 0.40 0.15 0.02 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of the price level, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMr data. provide a clear indication of a positive relationship between output and lagged inflation, as Phillips curve-type models, for instance, would predict. Indeed, for some countries, such as Mexico and Turkey, we find negative correlations be- tween inflation and the cyclical component of output, indicating countercyclical variations in inflation. Agenor, McDermott, and Prasad 269 Table 7. Cross Correlations between Domestic Output and Inflation Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP -0.08 0.49 0.09 -0.33 -0.14 BP 0.32 0.39 0.16 -0.20 -0.18 Colombia HP -0.02 -0.17 -0.23 0.48 -0.03 BP -0.50 -0.15 0.03 0.30 0.66 India HP -0.40 -0.02 0.10 0.14 0.12 BP -0.26 -0.42 0.04 0.36 0.34 Korea, Rep. of HP --0.16 -0.01 -0.19 0.05 0.40 BP 0.19 0.17 0.27 0.43 0.56 Malaysia HP -0.25 -0.23 0.10 -0.05 0.06 BP 0.14 -0.02 -0.17 -0.22 0.08 Mexico HP 0.23 -0.09 -0.48 0.11 0.39 BP 0.02 -0.52 -0.52 0.08 0.38 Morocco HP 0.07 -0.02 -0.13 0.23 0.21 BP -0.32 -0.02 0.00 0.31 0.35 Nigeria HP -0.02 0.08 -0.27 0.16 0.14 BP 0.00 -0.12 -0.08 0.10 0.04 Philippines HP -0.64 -0.14 0.76 0.04 -0.40 BP -0.49 -0.22 0.52 0.42 -0.05 Tunisia HP 0.23 0.12 -0.25 0.03 0.00 BP 0.58 0.42 -0.05 -0.26 -0.48 Turkey HP 0.25 0.00 -0.39 0.15 0.24 BP 0.23 -0.30 -0.46 -0.04 0.49 Uruguay HP 0.18 0.49 -0.06 -0.56 -0.28 BP 0.44 0.36 0.00 -0.44 -0.08 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of inflation, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. For countries like Mexico and Turkey the procyclical behavior of real wages and the countercyclical behavior of both the price level and the inflation rate suggest that supply shocks may have been a key determinant of domestic macro- economic fluctuations over the past two decades. It is worth emphasizing that, for our sample of middle-income countries, the term "supply shocks" could have 270 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 a different connotation than it has for large industrial economies. In particular, these developing countries could be subject to large terms-of-trade shocks rather than prototypical productivity shocks, although terms-of-trade shocks could, in principle, have both supply-side and demand-side effects. Money and Credit In recent years it has become increasingly evident that equilibrium business cycle models often need to incorporate monetary variables to capture important business cycle phenomena. The relationship between monetary variables and the business cycle has, therefore, become a topic of increasing interest (see, for in- stance, Kydland and Prescott 1994). This relationship is particularly relevant to middle-income countries, where the monetary mechanism could play an impor- tant stabilizing role. A large literature has evolved around the question of whether monetary vari- ables influence output in industrial countries or, in more loaded terminology, whether money causes output. From a different perspective, King and Plosser (1984) argue that the positive correlation between money and the business cycle largely reflects the endogenous response of inside money to exogenous shocks that drive business cycle fluctuations rather than a causal relationship from money to output. Given this debate and uncertainties regarding the definition of money that theoretical models use, we examine money-output correlations using several definitions of monetary aggregates. We estimate correlations between industrial production and an index of broad money (M2). Broad money roughly corresponds to the definition for industrial economies. Although in some cases the sign (and statistical significance) of the correlations is affected by the detrending procedure, the contemporaneous corre- lations are broadly positive for a majority of the sample countries, including Chile, Colombia, India, Morocco, the Philippines, Tunisia, and Uruguay (table 8). Among the remaining countries, the contemporaneous correlations are often close to zero, although for Korea, Malaysia, and Mexico, there is some evidence of countercyclical variation in broad money. Of the countries that show positive correlations between money and output, the pattern of lead-lag correlations and, in particular, the lag at which the peak positive correlation occurs could be interpreted as an indication of the speed with which innovations in monetary variables are transmitted to real activity. For these countries the peak positive correlations generally occur at very short lags, suggesting that the transmission of monetary shocks to real activity is fairly rapid. Of course, as noted earlier, this could simply reflect the endogenous response of money to output fluctuations that are driven by nonmonetary shocks. Indeed, when we run bivariate Granger-causality tests on these two variables, we find little evidence that money Granger-causes output, even in those countries where the correlations between the two variables are strongly positive. The patterns of correlations are similar when we use two alternative monetary aggregates-reserve (or base) money and narrow money (currency in circulation Agenor, McDermott, and Prasad 271 Table 8. Cross Correlations between Output and Broad Money (M2) Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP -0.51 -0.20 0.37 0.32 0.08 BP -0.22 -0.25 0.23 0.50 0.24 Colombia HP -0.23 -0.44 0.18 0.24 0.21 BP -0.36 -0.24 0.06 0.54 0.45 India HP 0.15 0.14 0.07 -0.41 -0.01 BP 0.13 0.25 0.35 -0.24 -0.10 Korea, Rep. of HP -0.27 0.17 0.03 -0.15 -0.11 BP -0.65 -0.39 -0.27 -0.34 -0.20 Malaysia HP -0.10 -0.04 -0.26 0.05 0.20 BP -0.05 -0.29 -0.32 -0.14 -0.04 Mexico HP 0.04 0.08 -0.09 -0.25 0.28 BP 0.32 -0.03 -0.24 0.00 0.15 Morocco HP -0.29 -0.07 0.24 -0.07 0.21 BP -0.37 0.06 0.21 0.17 0.53 Nigeria HP -0.05 -0.18 -0.14 0.22 0.00 BP -0.53 -0.38 -0.09 0.07 0.36 Philippines HP -0.45 -0.21 0.48 0.25 -0.22 BP -0.56 -0.34 0.34 0.61 0.21 Tunisia HP -0.15 -0.08 0.19 0.34 -0.16 BP -0.09 0.20 0.42 0.26 -0.45 Turkey HP -0.04 -0.07 0.17 -0.39 0.00 BP 0.01 -0.03 -0.18 -0.24 0.18 Uruguay HP 0.03 -0.02 -0.04 0.04 0.04 BP -0.09 0.26 0.45 0.37 0.25 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of broad money, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. plus sight deposits in the banking system).20 The main features of the results derived from using broad money are preserved when using the other monetary aggregates. The contemporaneous correlations are positive for about half of the countries in the sample, generally statistically insignificant for many of the oth- ers, and, in the case of Nigeria, clearly negative. Overall, therefore, we find lim- 20. These results are available on request. 272 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 ited evidence in our sample of the type of procyclical behavior of monetary ag- gregates that has been documented for many industrial countries (see, for in- stance, Backus and Kehoe 1992). More important, we are unable to detect evi- dence of Granger causality from money to output. These results may indicate the need to develop a different analytical framework for studying the relationship between monetary policy and macroeconomic fluctuations in developing coun- tries. We discuss this issue below. We also examine the cyclical behavior of measures of velocity corresponding to the alternative definitions of monetary aggregates discussed above. Again, to conserve space, we present only the results for the measure of velocity based on broad money.21 These correlations are striking (table 9). For 11 of the 12 coun- tries in our sample (Mexico being the exception) and with both filters, the con- temporaneous correlations between the velocity of broad money and industrial output are strongly negative. From a quantity theory perspective, of course, the countercyclical behavior of velocity is be expected, given the procyclical behav- ior of broad money and countercyclical variation in the aggregate price level in a majority of the sample countries. This result stands in sharp contrast to the weakly procyclical behavior of velocity in the G-7 countries, as documented by Fiorito and Kollintzas (1994). Finally, we consider another monetary variable that could have a significant influence on economic activity-domestic private sector credit. This variable is especially relevant for middle-income countries, where equity markets tend to be weakly capitalized relative to industrial-country markets and private sector credit typically plays an important role in determining investment and the financing of working capital needs-and thus overall economic activity, especially in the in- dustrial sector.22 Note that changes in credit could partly reflect the derived de- mand for credit, which in turn could be affected by exogenous shocks that influ- ence the level of industrial activity. Nevertheless, even in these circumstances changes in the availability of credit could dampen the effects of these shocks on industrial output. Thus the pattern of these correlations is still of considerable analytical value. A number of countries, including Colombia, India, Mexico, and Turkey, show a positive contemporaneous association between domestic credit and industrial output (table 10). Chile and Uruguay, in contrast, show a negative correlation. In the countries where the association is positive, the correlations peak at or close to a zero lag, indicating that the availability of domestic credit affects activity in the industrial sector fairly rapidly. However, this could simply reflect cyclical fluc- 21. Measures of velocity corresponding to rhe reserve-money and narrow-money aggregates yield velocity-output correlations that are broadly similar to the results discussed in this paragraph. These results are available on request. 22. Although the role of private sector credit in many developing countries is well documented, few studies have quantitatively assessed the relative importance of money and credit in the transmission of monetary policy. We intend to pursue this issue in future work. Agenor, McDermott, and Prasad 273 Table 9. Cross Correlations between Domestic Output and M2 Velocity Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP 0.26 -0.26 -0.76 0.34 0.56 BP -0.26 -0.62 -0.70 -0.08 0.38 Colombia HP -0.04 -0.42 -0.27 0.16 0.39 BP -0.06 -0.34 -0.38 0.28 0.35 India HP 0.48 0.11 -0.73 -0.22 0.21 BP 0.25 0.13 -0.68 -0.34 0.01 Korea, Rep. of HP 0.13 0.25 -0.69 0.13 0.07 BP -0.08 -0.29 -0.65 -0.24 0.03 Malaysia HP 0.09 0.13 -0.82 0.19 0.30 BP 0.18 -0.32 -0.80 -0.09 0.33 Mexico HP -0.14 0.01 0.05 -0.02 0.28 BP -0.52 -0.21 0.17 0.42 0.32 Morocco HP -0.13 0.22 -0.48 0.22 0.18 BP -0.02 0.21 -0.44 0.14 0.43 Nigeria HP 0.15 -0.10 -0.48 0.21 -0.02 BP -0.23 -0.38 -0.53 -0.19 0.02 Philippines HP 0.43 0.43 -0.71 0.07 0.46 BP 0.19 0.16 -0.60 0.18 0.31 Tunisia HP 0.21 -0.18 -0.62 0.20 0.19 BP 0.04 -0.38 -0.61 -0.26 -0.25 Turkey HP -0.11 -0.17 -0.10 -0.12 0.05 BP -0.09 -0.16 -0.43 0.05 0.33 Uruguay HP 0.14 -0.24 -0.59 0.39 0.47 BP -0.37 -0.50 -0.51 0.04 0.51 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of M2 velocity, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. tuations in the demand for private sector credit, where the demand for loans is determined primarily by other factors. To test this hypothesis, we run bivariate Granger-causality tests between the stationary components of private sector credit and industrial output. For some of the countries with positive correlations between these two variables, we do find that private sector credit has predictive power for industrial output in the Granger- 274 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 10. Cross Correlations between Domestic Output and Private Sector Credit Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Cbile HP 0.05 -0.43 -0.26 0.32 0.18 BP -0.54 -0.66 -0.50 -0.35 -0.32 Colombia HP -0.45 -0.31 0.28 0.37 0.18 BP -0.62 -0.11 0.36 0.64 0.34 India HP 0.17 0.29 0.21 -0.17 -0.19 BP 0.01 0.55 0.47 -0.07 -0.27 Korea, Rep. of HP -0.24 -0.01 0.12 0.12 -0.19 BP 0.02 0.03 0.12 0.04 0.19 Malaysia HP 0.07 0.08 -0.04 0.00 0.08 BP -0.46 -0.32 0.13 0.41 0.10 Mexico HP -0.39 -0.15 0.65 0.09 -0.31 BP -0.45 0.38 0.84 0.48 -0.11 Morocco HP -0.34 -0.09 0.09 0.02 0.31 BP -0.38 -0.22 0.04 0.27 0.54 Nigeria HP -0.13 -0.07 0.14 0.06 -0.01 BP -0.17 -0.05 0.10 0.02 -0.29 Philippines HP 0.05 0.56 0.00 -0.20 -0.07 BP -0.35 0.32 0.55 0.12 -0.38 Tunisia HP 0.14 -0.03 -0.10 0.25 0.18 BP -0.10 0.23 0.53 0.69 0.29 Turkey HP -0.32 -0.25 0.44 0.05 -0.28 BP -0.52 0.01 0.52 0.37 -0.37 Uruguay HP 0.08 -0.07 -0.27 0.18 0.39 BP -0.14 -0.38 -0.34 -0.04 0.18 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of private sector credit, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. causal sense. However, for some of these countries there is also evidence of re- verse causation from output to credit. Thus we do not find robust evidence of a unidirectional causal relationship from credit to economic activity. Nevertheless, the strong positive association between private sector credit and the domestic business cycle in some of the sample countries has important implications for the Agenor, McDermott, and Prasad 275 design of stabilization programs. Ignoring this link may exacerbate the output cost of a restrictive monetary policy aimed at lowering inflation. Foreign Trade and the Business Cycle In this subsection we explore the relationship between domestic business cycle fluctuations and fluctuations in price and quantity variables that are relevant to international trade. In particular, we examine correlations of output fluctuations with fluctuations in merchandise trade and measures of both nominal and real effective exchange rates. An adequate measure of foreign trade transactions is the trade balance, con- structed as the difference between real exports and real imports and divided by real GDP in order to control for scale effects. In the absence of reliable data on price deflators for exports and imports, many authors use the ratio of the sum of nominal exports and imports to output. Unfortunately, we are even more con- strained because we have only real industrial output data for most of the coun- tries in our sample. Hence we use the ratio of exports to imports at current prices as a rough measure of the trade balance. Because changes in the terms of trade could be large and important for these countries, we return to that issue later. For Chile, Mexico, Turkey, and Uruguay, the contemporaneous correlations between our proxy for trade balance movements and industrial output are nega- tive irrespective of the filter used (table 11). This pattern is similar to that found for industrial countries, as reported by several authors-see Fiorito and Kollintzas (1994), Prasad and Gable (1998), and the references therein. However, for cer- tain countries-including Morocco and Nigeria-the contemporaneous correla- tions are strongly positive. This result may reflect a strong link between changes in industrial output and exports of manufactures, or it may reflect the fact that merchandise imports are not highly sensitive to fluctuations in domestic demand. In addition, where we do find significant correlations between the trade ratio and domestic output, these correlations peak at (or near) lag zero. We interpret this as indicative of the close relationship between international trade and industrial output in these middle-income economies, with industrial output being a good proxy for output in the traded goods sector (other than primary commodities). We were able to obtain unit values of imports and exports and to construct a quarterly index of the terms of trade for only three of the countries in our sample- Colombia, Korea, and Mexico. These three countries show a strong positive cor- relation between the cyclical components of industrial production and the terms- of-trade index (table 12). For Colombia and Korea the BP-filtered data yield the strongest correlations. This suggests that the positive relationship between out- put and the terms of trade might be obscured when using the HP filter because of the large amount of high-frequency variation in the terms-of-trade data. Because these middle-income countries are unlikely to affect the world price of any industrial commodity, the positive correlations could be seen as consistent with demand shifts that lead to simultaneous increases in world prices and in the export demand for the industrial sector output of these countries. For the three 276 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 11. Cross Correlations between Domestic Output and the Trade Balance Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP 0.23 0.54 -0.54 -0.49 0.28 BP 0.46 0.32 -0.48 -0.48 0.18 Colombia HP 0.30 0.05 -0.20 -0.16 0.25 BP 0.23 0.02 -0.08 -0.15 0.05 India HP 0.10 -0.05 -0.10 0.08 -0.03 BP 0.18 0.19 -0.10 0.10 0.24 Korea, Rep. of HP 0.00 0.40 0.04 -0.09 -0.15 BP 0.13 0.43 0.39 -0.03 -0.34 Malaysia HP -0.15 0.17 0.16 -0.02 -0.19 BP -0.15 0.04 0.15 0.10 0.16 Mexico HP 0.42 -0.09 -0.71 0.08 0.53 BP 0.24 -0.49 -0.60 0.17 0.75 Morocco HP 0.04 -0.28 0.31 -0.16 -0.30 BP 0.15 -0.12 0.23 -0.14 -0.22 Nigeria HP 0.00 -0.02 0.46 -0.19 -0.32 BP -0.03 0.20 0.20 -0.47 -0.18 Philippines HP -0.34 -0.27 0.24 0.05 0.12 BP -0.23 -0.35 -0.09 0.02 0.62 Tunisia HP 0.07 -0.06 0.00 0.04 0.04 BP 0.08 -0.03 -0.08 0.10 0.04 Turkey HP -0.06 0.09 -0.49 0.28 0.22 BP -0.33 -0.18 -0.18 0.37 0.52 Uruguay HP -0.12 -0.15 -0.30 -0.11 0.46 BP -0.25 -0.35 -0.36 -0.11 0.28 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported arc between the contemporaneous values of domestic output and the jth lag or lead of the trade balance, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. countries for which we have terms-of-trade data, the strong positive correlations between lagged values of the terms-of-trade index and contemporaneous output provide further support for this interpretation. Overall, our results are consistent with those of Rodrfguez-Mata (1997) for Costa Rica, Kose and Riezman (1998) for Sub-Saharan Africa, and Mendoza (1995), who suggest that almost half of the output fluctuations in developing countries can be explained by terms-of- Agenor, McDermott, and Prasad 277 Table 12. Cross Correlations between Domestic Output and Terms of Trade Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Colombia HP 0.18 0.15 0.10 0.15 0.08 BP 0.57 0.42 0.34 0.06 -0.29 Korea, Rep. of HP 0.08 0.08 0.41 -0.17 -0.14 BP 0.00 0.38 0.62 0.50 0.26 Mexico HP -0.28 0.13 0.46. -0.30 -0.24 BP -0.10 0.49 0.46 0.14 0.32 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of terms of trade, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. trade disturbances. Our results are also consistent with those of Deaton and Miller (1995), who find evidence that export price shocks have had substantial contem- poraneous effects on output in Sub-Saharan Africa. Cyclical Behavior of Exchange Rates The interpretation of the unconditional correlations between industrial out- put and measures of nominal and real effective exchange rates is complicated by the fact that their short-run relationship depends crucially on the sources of mac- roeconomic fluctuations.23 Nonetheless, it is useful to look at unconditional cor- relations for two reasons. First, the sign and magnitude of these correlations could indicate the types of shocks that have dominated fluctuations over a period of time. Second, these correlations could help in interpreting the correlations between output and trade variables. In India, Morocco, Tunisia, and Turkey, there is some evidence of a positive relationship between nominal effective exchange rates and industrial output, while the correlations are generally negative for Chile, Nigeria, and Uruguay (table 13). The correlations between output and real effective exchange rates show a similar pattern, but with a few important differences (table 14). The contempo- raneous correlations for Mexico and Uruguay are positive, while the correlations for Morocco and the Philippines are close to zero. However, many of the con- temporaneous correlations are not significantly different from zero. The absence 23. There are, of course, important differences in the exchange rate arrangements of the sample countries. However, since we use trade-weighted measures of both real and nominal effective exchange rates, the fact that certain bilateral exchange rates could be fixed does not, in principle, affect the interpretation of our results. We also define the effective exchange rates such that an increase in the exchange rate implies an appreciation of the currency (in real or nominal terms, as the case may be). Thus a positive correlation indicates that the exchange rate tends to appreciate when the cyclical component of output rises. 278 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 13. Cross Correlations between Domestic Output and the Nominal Effective Exchange Rate Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP 0.13 -0.57 -0.05 0.32 -0.11 BP -0.04 -0.40 -0.36 -0.13 0.00 Colombia HP -0.29 -0.21 0.19 0.13 -0.23 BP -0.14 -0.40 -0.40 -0.27 -0.08 India HP 0.03 0.45 0.25 -0.19 -0.32 BP -0.26 0.21 0.18 -0.37 -0.48 Korea, Rep. of HP 0.11 -0.43 -0.02 0.45 0.06 BP -0.39 -0.55 -0.09 0.45 0.56 Malaysia HP -0.03 0.17 0.10 0.12 -0.51 BP -0.25 0.14 0.19 -0.06 -0.34 Mexico HP 0.02 0.69 -0.05 -0.41 -0.10 BP -0.05 0.52 0.38 0.10 0.36 Morocco HP 0.04 0.06 0.17 0.04 -0.12 BP -0.05 0.18 0.33 0.29 -0.02 Nigeria HP 0.06 -0.13 -0.04 -0.13 -0.14 BP -0.41 -0.47 -0.39 -0.25 0.10 Philippines HP 0.48 -0.01 -0.38 0.26 0.02 BP -0.06 0.29 0.25 -0.02 -0.46 Tunisia HP -0.32 -0.31 0.33 0.49 0.05 BP -0.21 0.14 0.67 0.79 0.31 Turkey HP 0.02 -0.26 0.36 -0.09 -0.22 BP 0.28 0.34 0.09 -0.15 -0.32 Uruguay HP 0.22 -0.22 -0.27 0.03 0.16 BP -0.18 -0.41 -0.57 -0.31 0.24 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of the nominal effective exchange rate, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on tF data. of a systematic relationship between real exchange rates and the business cycle is consistent with the notion that this relationship is affected by an amalgam of supply, real demand, and nominal shocks, each of which could affect this corre- lation in different ways. One interesting aspect of these results is that, for many countries, the correla- tions are quite similar using either nominal or real measures of effective exchange Agenor, McDermott, and Prasad 279 Table 14. Cross Correlations between Domestic Output and the Real Effective Exchange Rate Country Eight-quarter Four-quarter Zero Four-quarter Eight-quarter and filter lag lag lag lead lead Chile HP -0.23 -0.55 0.26 0.36 -0.24 BP -0.32 -0.54 -0.41 -0.33 -0.31 Colombia HP -0.12 -0.03 0.06 0.01 -0.30 BP 0.01 -0.19 -0.36 -0.44 -0.46 India HP 0.02 0.47 0.19 -0.23 -0.30 BP -0.10 0.16 0.09 -0.45 -0.61 Korea, Rep. of HP 0.12 -0.47 0.10 0.27 0.06 BP -0.52 -0.62 -0.14 0.41 0.56 Malaysia HP 0.06 0.18 0.10 0.04 -0.49 BP -0.19 0.17 0.19 -0.14 -0.38 Mexico HP -0.47 0.24 0.59 -0.36 -0.39 BP -0.09 0.48 0.47 -0.12 0.01 Morocco HP -0.06 0.08 -0.01 0.01 -0.23 BP -0.10 -0.06 -0.01 0.09 -0.23 Nigeria HP 0.08 -0.06 -0.09 -0.19 -0.12 BP -0.32 -0.46 -0.45 -0.34 0.00 Philippines HP 0.07 -0.56 0.03 0.63 -0.19 BP -0.48 -0.46 0.21 0.61 0.06 Tunisia HP -0.31 -0.17 0.30 0.44 0.02 BP -0.27 0.11 0.67 0.76 0.22 Turkey HP 0.19 -0.13 0.30 -0.23 -0.22 BP 0.44 0.33 -0.15 -0.35 -0.21 Uruguay HP -0.17 -0.16 0.22 0.47 -0.05 BP -0.53 -0.25 0.16 0.31 -0.02 Note: HP and BP refer to the stationary components derived using the modifed Hodrick-Prescott and band-pass filters, respectively. The correlations reported are between the contemporaneous values of domestic output and the jth lag or lead of the real effective exchange rate, with both variables detrended using the same filter. The data series and sources are described in the appendix. Source: Authors' calculations based on IMF data. rates. This finding is in line with a substantial body of research showing that, for industrial countries, nominal and real exchange rates are strongly positively cor- related at business cycle frequencies (see, for instance, Mussa 1986 and Taylor 1995). Indeed, we find that the contemporaneous correlations between real and nominal effective exchange rates are strongly positive for all countries in our sample, irrespective of the filter used. 280 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 IV. SUMMARY OF THE FINDINGS In this section we summarize the main findings of the paper. As noted in the discussion thus far, some of these results have previously been reported by other authors, using different data sets. * Output volatility, as measured by the standard deviations of the filtered cyclical components of industrial production, varies substantially across developing countries. However, on average, it is much higher than the level typically observed in industrial countries. Developing countries also show considerable persistence in output fluctuations. * Activity in industrial countries has a significantly positive influence on output in most developing countries. Real interest rates in industrial countries tend to be positively associated with output fluctuations in our sample of middle- income countries. * Government expenditures are countercyclical. Government revenue is acyclical in some countries and significantly countercyclical in others-a phenomenon that is difficult to explain. The fiscal impulse (defined as the ratio of government spending to government revenue) is negatively correlated with the business cycle. * The cyclical behavior of nominal wages varies markedly across countries and is not robust across filters. In contrast, the evidence strongly supports the assumption of procyclical real wages. * There is no consistent relationship between the stationary components of the levels of output and prices and the levels of output and inflation. Variations in the price level and inflation are countercyclical in some countries and procyclical in a few. Contemporaneous correlations between money (measured by various monetary aggregates) and output are broadly positive, but not very strong- in contrast to the evidence for many industrial countries. * The contemporaneous correlations between the velocity of broad money and industrial output are strongly negative for almost all the countries in our sample. This result is in contrast to the weakly procyclical behavior of velocity observed in most industrial countries. * Domestic credit and industrial output are positively associated in some countries. However, the strength of the relationship is not always robust to the choice of detrending procedure. Some countries show a negative correlation between these two variables. * There is no robust correlation between merchandise trade movements (as measured by the ratio of exports to imports) and output. For some countries the contemporaneous correlations are negative (irrespective of the filter used), whereas for others the contemporaneous correlations are strongly positive. The positive relationship may indicate that fluctuations in industrial output Agenor, McDermott, and Prasad 281 are driven by export demand and that merchandise imports are not as sensitive to domestic demand fluctuations as they are in industrial countries. * Cyclical movements in the terms of trade are strongly positively correlated with output fluctuations. * There are no systematic patterns in the contemporaneous correlations between nominal effective exchange rates and industrial output; the results are similar for real effective exchange rates. Fluctuations in real and nominal effective exchange rates are strongly positively correlated for the developing countries in our sample. V. CONCLUDING REMARKS In this article we studied the cyclical properties of a large number of (season- ally adjusted) macroeconomic time series for a group of 12 (mostly middle- income) developing countries, using two univariate detrending methods. We dis- cussed the cross-correlation patterns between output and the macroeconomic time series and attempted to identify a set of relatively robust regularities that can be used as a benchmark to guide theoretical research in development macro- economics. We also highlighted similarities and differences between our results and other studies on business cycle fluctuations in industrial and developing countries. We can make several remarks on the methodological and analytical implica- tions of our analysis. First, our results suggest that, although the correlations derived from different filters were often very similar, several quantitative (as well as qualitative) features of the data are not robust across detrending methods. This result is similar to that of Blackburn and Ravn (1991) and Canova (1998), among other authors. Because generally we cannot know ex ante when results will vary across filters, considering systematically an array of detrending meth- ods remains an important test of robustness in empirical research on business cycle regularities. Second, the unconditional correlations between different variables (such as exchange rates or prices) and domestic output may be small because they average the effects of different types of shocks. It is, therefore, important to develop and estimate structural models, along the lines, for instance, of Ahmed and Park (1994), Rogers and Wang (1995), Hoffmaister and Rold6s (1997), and Prasad (1999), that attempt to separate out the effects of different types of macroeconomic shocks on variables such as prices, output, foreign trade, and exchange rates in develop- ing countries. However, existing methods remain controversial; we do not yet have models that convincingly isolate different types of shocks. Third, the analysis in this article ignores the possible effects of measurement errors in the raw data. This is a potentially serious problem. For instance, in our analysis of the correlations between domestic output and foreign interest rate shocks, we do not account for the risk premium that borrowers from developing 282 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 countries typically face on world capital markets. However, there is considerable evidence that such premiums can be large on average (particularly for countries with a high ratio of external debt to output) and could change unpredictably in the short run as a result of sudden shifts in market sentiment. This measurement problem, which has not been adequately addressed in other studies, suggests that we should exercise caution in judging the strength and direction of correlations between domestic output and a measure of world interest rates that does not capture movements in country-specific risk premiums. Finally, at the analytical level, a natural step forward is to build stochastic general equilibrium simulation models of small, open developing economies in order to assess if such models (properly calibrated) can reproduce the stylized facts highlighted here. Some of the correlations established (such as the countercyclical behavior of government spending) can indeed be explained within existing theoretical constructs. Building more general quantitative models that are capable of accounting for the other types of business cycle regularities high- lighted here could prove important for the design of stabilization policies and macroeconomic management in developing countries. APPENDIX: DATA SOURCES AND UNIT ROOT TESTS The primary sources of data used in this study are the International Monetary Fund's International Financial Statistics (IFS) and Information Notice System, supplemented by other sources. This appendix describes the series, together with their IFS codes. All data are available on request. * Real output is the industrial production index (series 66) for Mexico, Korea, India, Malaysia, and Tunisia and the manufacturing production index (se- ries 66ey) for Chile, Morocco, Nigeria, the Philippines, and Uruguay. We obtained the industrial production index from the International Monetary Fund (IMF) desk economist for Colombia and from the Organisation for Economic Co-operation and Development (OECD) database for Turkey. IMF desk economists also provided partial information for Turkey, Tunisia, and Uruguay. * The cpi is series 64 for all countries. The IMF desk economist for Tunisia filled in the data (for that country) missing from the IFS. * The nominal wage index is series 65 for Mexico, Chile, and the Philippines and series 65ey for Korea. We obtained data for Colombia from the IMF desk economist. We obtained data for Turkey from the OECD database and the IMF desk economist. We calculate the real wage index by deflating the nominal wage series by the cPi. * The monetary base (or reserve money) is series 14 for all countries. Narrow money is series 34, and broad money is the sum of series 34 and 35, again for all countries. We calculate velocity for each monetary indicator by first transforming the monetary aggregate into an index and then dividing by Agenor, McDermott, and Prasad 283 the product of the cPI and the real output index, which is used as a proxy for nominal output. * Private sector credit is series 32d for all countries. We calculate the real credit variable by deflating the nominal aggregate by the cPi. The IMF desk economist for Tunisia filled in the data (for that country) missing from IFS. * Government expenditures in nominal terms is series 82 for Mexico, Korea, and the Philippines. We obtained data for Chile from Chile's Ministry of Finance. We derive the expenditure index by first transforming the nominal series into an index and then dividing by the same proxy for nominal output used to derive velocity indicators. * Government revenue in nominal terms is series 81. We derive the revenue index in the same way as the expenditure index. * We derive the fiscal impulse measure by dividing series 82 by series 81. * The trade ratio is the ratio of merchandise exports at current prices (series 70) to merchandise imports at current prices (series 71), with both variables measured in U.S. dollar terms. * We obtain trade-weighted measures of nominal and real effective exchange rates from the IMF's Information Notice System. * The terms of trade are the ratio of export unit values (series 74) to import unit values (series 75) for Colombia and Korea. For Mexico we obtain export and import price indexes from the OECD database. * World output is proxied by the industrial production index for industrial countries (series 66, code 110). The world real interest rate is proxied by the difference between the nominal euro-dollar rate in London (series 60d, country code 112) and the rate of inflation in consumer prices in industrial countries (series 64, code 110). We performed a set of standard unit root tests, including augmented Dickey- Fuller tests and Phillips-Perron tests, on our raw data series (all of which were converted into logarithms for the empirical work, except for the world real inter- est rate). These tests indicated that virtually all of the series were nonstationary in levels over the relevant sample period and that, therefore, computing correla- tions using the raw data would not be appropriate. We also used similar unit root tests to confirm that the cyclical components obtained with the filters em- ployed in this article were indeed stationary. In addition, we found that the infla- tion rate (measured as the four-quarter change in the price level) did not appear to be stationary in levels for several countries in our sample. To conserve space, the results of these unit root tests are not reported here but are available on request. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. 284 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Abraham, Katharine G., and John C. Haltiwanger. 1995. "Real Wages and the Business Cycle." Journal of Economic Literature 33(September):1215-64. Agenor, Pierre-Richard. 1997. "Capital-Market Imperfections and the Macroeconomic Dynamics of Small Indebted Economies." Princeton Studies in International Finance 82. Princeton University, Princeton, N.J. Processed. Agenor, Pierre-Richard, C. John McDermott, and Eswar S. Prasad. 1998. "Macroeco- nomic Fluctuations in Developing Countries: Some Stylized Facts." Economic Devel- opment Institute, World Bank, Washington, D.C. Processed. Online at http:// www.worldbank.org/html/edi/ediwp.htm. Ag6nor, Pierre-Richard, and Peter J. Montiel. 1996. Development Macroeconomics. Princeton, N.J.: Princeton University Press. Ahmed, Shaghil, and Kae H. Park. 1994. "Sources of Macroeconomic Fluctuations in Small Open Economies." Journal of Macroeconomics 16(Spring):1-36. Backus, David K., and Patrick J. Kehoe. 1992. "International Evidence on the Historical Properties of Business Cycles." American Economic Review 82(September):864-88. Baxter, Marianne, and Robert G. King. 1995. "Approximate Band-Pass Filters for Eco- nomic Time Series." NBER Working Paper 5022. National Bureau of Economic Re- search, Cambridge, Mass. Processed. Blackburn, Keith, and Morten 0. Ravn. 1991. "Univariate Detrending of Macroeco- nomic Time Series." Working Paper 22. Department of Economics, Aarhus Univer- sity, Aarhus, Denmark. Processed. Canova, Fabio. 1998. "Detrending and Business Cycle Facts." Journal of Monetary Eco- nomics 41(June):475-512. Chadha, Bankim, and Eswar Prasad. 1994. "Are Prices Countercyclical? Evidence from the G7." Journal of Monetary Economics 34(October):239-57. Cooley, Thomas F., and Lee E. Ohanian. 1991. "The Cyclical Behavior of Prices." Jour- nal of Monetary Economics 28(August):25-60. Deaton, Angus S., and Ronald I. Miller. 1995. "International Commodity Prices, Macro- economic Performance, and Politics in Sub-Saharan Africa." Princeton Essays in Inter- national Finance 79. Princeton University, Princeton, N.J. Processed. Fiorito, Riccardo, and Tryphon Kollintzas. 1994. "Stylized Facts of Business Cycles in the G7 from a Real Business Cycles Perspective." European Economic Review 38(Feb- ruary):235-69. Hodrick, Robert J., and Edward C. Prescott. 1997. "Postwar U.S. Business Cycles: An Empirical Investigation." Journal of Money, Credit, and Banking 29(February):1-16. Hoffmaister, Alexander W., and Jorge E. Rold6s. 1997. "Are Business Cycles Different in Asia and Latin America?" Working Paper 97/9. International Monetary Fund, Wash- ington, D.C. Processed. Judd, John P., and Bharat Trehan. 1995. "The Cyclical Behavior of Prices: Interpreting the Evidence." Journal of Money, Credit, and Banking 27(August):789-97. King, Robert G., and Charles I. Plosser. 1984. "Money, Credit, and Prices in a Real Business Cycle." American Economic Review 74(June):363-80. Kose, M. Ayhan, and Raymond Riezman. 1998. "Trade Shocks and Macroeconomic Fluctuations in Africa." Department of Economics, Brandeis University, Waltham, Mass. Processed. Agenor, McDermott, and Prasad 28S Kouparitsas, Michael A. 1996. "North-South Business Cycles." Working Paper 96-9. Federal Reserve Bank of Chicago. Processed. .1997. "North-South Terms of Trade: An Empirical Investigation." Working Pa- per 97-5. Federal Reserve Bank of Chicago. Processed. Kydland, Finn E., and Edward C. Prescott. 1994. "Business Cycles: Real Facts and a Monetary Myth." In Preston J. Miller, ed., The Rational Expectations Revolution: Readings from the Front Line. Cambridge, Mass.: MIT Press. Kydland, Finn E., and Carlos Zarazaga. 1997. "Is the Business Cycle of Argentina Differ- ent?" Federal Reserve Bank of Dallas Economic Review (4th quarter):21-36. Lumsdaine, Robin L., and Eswar S. Prasad. 1997. "Identifying the Common Component in International Economic Fluctuations." NBER Working Paper 5984. National Bureau of Economic Research, Cambridge, Mass. Processed. McDermott, C. John. 1997. "An Automatic Method for Choosing the Smoothing Param- eter in the HP Filter." International Monetary Fund, Washington, D.C. Processed. Mendoza, Enrique. 1995. "The Terms of Trade, the Real Exchange Rate, and Economic Fluctuations." International Economic Review 36(February):101-37. Mussa, Michael. 1986. "Nominal Exchange Rate Regimes and the Behavior of Real Ex- change Rates: Evidence and Implications." Carnegie-Rochester Conference on Public Policy 25(Autumn):117-213. Prasad, Eswar S. 1999. "International Trade and the Business Cycle." Economic Journal 109(October):58 8-606. Prasad, Eswar S., and Jeffery A. Gable. 1998. "International Evidence on the Determi- nants of Trade Dynamics." IMF Staff Papers 45(September):401-39. Rodriguez-Mata, Margarita. 1997. "Cyclical Patterns of the Costa Rican Economy." Central Bank of Costa Rica. Processed. Rogers, John H., and Ping Wang. 1995. "Output, Inflation, and Stabilization in a Small Open Economy: Evidence from Mexico." Journal of Development Economics 46(April):271-93. Rotemberg, Julio J., and Michael Woodford. 1991. "Markups and the Business Cycle." In Olivier J. Blanchard and Stanley Fischer, eds., NBER Macroeconomics Annual: 1991. Cambridge, Mass.: MIT Press. Taylor, Mark P. 1995. "The Economics of Exchange Rates." Journal of Economic Lit- erature 33(March):13-47. van Els, Peter J. 1995. "Real Business Cycle Models and Money: A Survey of Theories and Stylized Facts." Weltwirtschafliches Archiv 131 (June):223-64. THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2: 287-307 Monitoring Banking Sector Fragility: A Multivariate Logit Approach Asli Demirgii-Kunt and Enrica Detragiache This article explores how a multivariate logit model of the probability of a banking crisis can be used to monitor banking sector fragility. The proposed approach relies on readily available data, and the fragility assessment has a clear interpretation based on in- sample statistics. The model has better in-sample performance than currently available alternatives, and the monitoring system can be tailored to fit the preferences of decisionmakers regarding type I and type If errors. The framework can be useful as a preliminary screen to economize on precautionary costs. The past two decades have seen a proliferation of systemic banking crises, as documented by Lindgren, Garcia, and Saal (1996) and Caprio and Klingebiel (1996), among other comprehensive studies. The spread of banking sector prob- lems and the difficulty of anticipating their outbreak have highlighted the need to improve monitoring capabilities at both the national and supranational levels and raised the issue of using statistical studies of past banking crises to develop a set of indicators of the likelihood of future problems. In our previous work we developed an empirical model of the determinants of systemic banking crises for a large panel of countries (Demirgiiu-Kunt and Detragiache 1998, 1999). That research revealed a group of variables, including macroeconomic variables, characteristics of the banking sector, and structural characteristics of the country, that are robustly correlated with the emergence of banking sector crises. In this article we explore how we can use the information contained in that empirical relationship to monitor banking sector fragility.1 The basic idea is to estimate a specification of the multivariate logit model used in our previous work that relies mainly on explanatory variables whose future values are routinely forecasted by professional forecasters, the Interna- tional Monetary Fund (IMF), or the World Bank. We then compute the probabili- 1. Other studies using limited dependent variable econometric models to estimate the probabilities of banking crises are Eichengreen and Rose (1998) and Hardy and Pazarbasioglu (1998). These studies do not address issues of forecasting. Ash Demirgu,c-Kunt is lead economist with the Development Research Group at the World Bank, and Enrica Detragiache is senior economist with the Research Department at the International Monetary Fund. Their e-mail addresses are ademirguckunt@worldbank.org and edetragiache@imf org. The authors wish to thank Anqing Shi for capable research assistance. C 2000 The International Bank for Reconstruction and Development! THE WORLD BANK 287 288 THE WORLD BANK ECONOMIC REVIEW, VOL.. 14, NO. 2 ties of out-of-sample banking crises using the estimated coefficients and fore- casted values of the explanatory variables. Along with the results of in-sample estimations, we use these forecasted probabilities to make a quantitative assess- ment of fragility. We examine two monitoring frameworks. In the first the monitor wants to know whether the forecasted probabilities are high enough to trigger a response. Taking no action when a crisis is nearing is costly, but so is taking action when a crisis is not impending. The decisionmaker chooses a probability threshold that minimizes a loss function reflecting both types of cost. In the second framework the monitor is simply interested in rating the fragility of the banking system. Depending on the rating, several courses of action may follow, but these are not explicitly modeled. In this framework it is desirable for a rating to have a clear interpretation in terms of the probability of a crisis, so that different ratings can be compared. We examine one such example. To illustrate the monitoring procedures developed in the first part of the ar- ticle, we then conduct a limited out-of-sample forecasting exercise in the second part. We construct forecasted probabilities for the six banking crises that oc- curred in 1996-97, namely the Jamaican crisis in 1996 and the five East Asian crises in 1997. I. THE LITERATURE An extensive literature reviews banking crises around the world, examining the developments leading up to the crises as well as policy responses. This body of work does not directly identify leading indicators of banking sector problems, pointing instead to a number of variables that display "anomalous" behavior in the periods preceding the crises. For instance, Gavin and Hausman (1996) and Sachs, Tornell, and Velasco (1996) suggest that credit growth be used as an indi- cator of impending troubles, as crises tend to be preceded by lending booms. Mishkin (1996) highlights equity price declines, while Calvo (1996), in his analy- sis of Mexico's 1995 crisis, suggests that monitoring the ratio of broad money to foreign exchange reserves may be useful in evaluating the banking sector's vul- nerability to a currency crisis. Honohan (1997) evaluates alternative indicators more systematically. Using a sample of 18 countries that experienced banking crises and 6 that did not, he divides the crisis countries into three groups (of equal size) according to the type of crisis (macroeconomic, microeconomic, or related to the behavior of the gov- ernment). He then compares the average values of seven indicators for the crisis countries with the averages for the control group. This exercise shows that bank- ing crises arising from macroeconomic problems are associated with high loan- to-deposit ratios, high foreign borrowing-to-deposit ratios, and high growth rates of credit. Similarly, crises stemming from government interventions are associ- ated with high levels of government borrowing and central bank lending to the banking system. However, banking crises originating from microeconomic pres- Demirguii-Kunt and Detragiache 289 sures do not appear to be associated with abnormal behavior on the part of the indicators examined in the study. Rojas-Suarez (1998) proposes an approach based on bank-level indicators, similar in spirit to the CAMEL system used by U.S. regulators to identify problem banks. She argues that in emerging markets (particularly those in Latin America) CAMEL indicators are not good signals of bank strength and that more informa- tion can be obtained by monitoring the deposit interest rate, the spread between the lending and deposit rates, the growth rate of credit, and the growth rate of interbank debt. Because these variables are measured against banking system averages, however, this approach appears more adequate for identifying weak- nesses specific to individual banks than for identifying systemic fragility. The approach also requires bank-level information, which often is not readily avail- able in developing countries. To date, Kaminsky and Reinhart (1999) have made the most comprehensive effort to develop a set of early warning indicators for banking crises (and cur- rency crises). The methodology is refined in Kaminsky (1998). These studies ex- amine the behavior of 15 macroeconomic indicators for a sample of 20 countries that experienced banking crises during 1970-95.2 The authors compare the be- havior of each indicator in the 24 months prior to the crisis with the behavior during tranquil times. A variable is deemed to signal a crisis if it crosses a particu- lar threshold at any time. If that signal is followed by a crisis within the next 24 months, then it is considered correct; otherwise it is considered noise. The thresh- old for each variable is chosen to minimize the in-sample noise-to-signal ratio. The authors then compare the performance of different indicators based on the associated type I and type II errors, the noise-to-signal ratio, and the probability of a crisis occurring conditional on a signal being issued.3 The indicator with the lowest noise-to-signal ratio and the highest probability of crisis conditional on the signal is the real exchange rate, followed by equity prices and the money multiplier. These three indicators, however, have a high incidence of type I errors, as they fail to issue a signal in 73-79 percent of the observations in the 24 months preceding a crisis. The incidence of type II errors, in contrast, is much lower, ranging from 8 to 9 percent. The variable with the lowest type I error is the real interest rate, which issues a signal in 30 percent of the observations preceding a crisis. The high incidence of type I errors relative to type II errors may not be a desirable feature of a warning system if the costs of raising a false alarm are small relative to the costs of failing to anticipate a crisis. Since, presumably, the likelihood of a crisis is greater when several indicators signal simultaneously, Kaminsky (1998) develops composite indexes. These in- clude the number of indicators that cross the threshold at any given time or a weighted variant of that index in which each indicator is weighted by its signal- to-noise ratio so that more informative indicators receive more weight. The best 2. For a study of early warning indicators of currency crises, see also IMF (1998). 3. The authors use an adjusted version of the noise-to-signal ratio, computed as the ratio of the probability of a type 11 error to 1 minus the probability of a type I error. 290 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 composite indicator outperforms the real exchange rate in predicting crises in the sample, but it is worse at predicting observations of no crisis.4 The approach we develop here will allow policymakers to choose a warning system that reflects the relative cost of type I and type II errors, and it will offer a natural way of measuring the combined effect of various economic forces on banking sector vulnerability. By making better use of all available information, the system will produce lower overall in-sample forecasting errors than would individual indicators. We also examine a problem not addressed by Kaminsky and Reinhart (1999), that of a monitor who wishes to use information contained in the statistical analysis of past crises not just to anticipate a crisis but also to make a more nuanced assessment of banking sector fragility. II. ESTIMATING THE PROBABILITIES OF IN-SAMPLE BANKING CRISES IN A MULTIVARIATE LOGIT FRAMEWORK The starting point of our analysis is an econometric model of the probability of a systemic banking crisis. In Demirgiiu-Kunt and Detragiache (1998, 1999) we estimate alternative specifications of a logit regression for a large sample of de- veloping and industrial countries, including countries that experienced banking crises and those that did not. Details on sample selection, the construction of the banking crisis variable, and the choice of explanatory variables can be found there. To form the basis of an easy-to-use monitoring system, we estimate a specifi- cation of our empirical model that includes only variables that are available from the IMF'S International Financial Statistics or other publicly available databases and that are routinely forecasted by the IMF in its biannual World Economic Outlook or by professional forecasters. As it turns out, this is not the specifica- tion that best fits the data. We estimate the regression using a panel of 766 obser- vations for 65 countries during 1980-95.5 In this panel we identify 36 systemic banking crises, so that crisis observations make up 4.7 percent of the sample (table 1). The set of explanatory variables capturing macroeconomic conditions includes the growth rate of real gross domestic product (GDP), the change in the terms of trade, the rate of depreciation of the exchange rate (relative to the U.S. dollar), the rate of inflation, and the fiscal surplus as a share of GDP. The explana- tory variables capturing characteristics of the financial sector are the ratio of broad money to foreign exchange reserves and the growth rate of bank credit lagged two periods. Finally, GDP per capita proxies for structural characteristics of the economy. 4. Kaminsky (1998) finds that the probability of a crisis computed by taking into account the number of indicators signaling a crisis increased substantially before the 1997 crises in the Philippines, Malaysia, and Thailand, but not in Indonesia. The Republic of Korea was not part of the sample. 5. Because of missing data or breaks in the series, part of the sample period may be excluded for some countries. Years in which banking crises are ongoing also are excluded from the sample. Demirgui,-Kunt and Detragiache 291 Table 1. Banking Crises and Estimated Probabilities of Crises Country Crisis year Estimated probability Chile 1981 0.231 Colombia 1982 0.066 Ecuador 1995 0.439 El Salvador 1989 0.055 Finland 1991 0.066 Guyana 1993 0.007 India 1991 0.069 Indonesia 1992 0.107 Israel 1983 0.999 Italy 1990 0.015 Japan 1992 0.037 Jordan 1989 0.334 Kenya 1993 0.361 Malaysia 1985 0.067 Mali 1987 0.035 Mexico 1982 0.527 Mexico 1994 0.099 Nepal 1988 0.018 Nigeria 1991 0.011 Norway 1987 0.036 Panama 1988 0.539 Papua New Guinea 1989 0.121 Peru 1983 0.244 Philippines 1981 0.035 Portugal 1986 0.064 South Africa 1985 0.196 Sri Lanka 1989 0.036 Swaziland 1995 0.633 Sweden 1990 0.036 Tanzania 1988 0.035 Thailand 1983 0.027 Turkey 1991 0.158 Turkey 1994 0.482 United States 1980 0.238 Uruguay 1981 0.329 Venezuela 1993 0.494 Source: Authors' calculations. The estimated coefficients of the logit regression reveal that low GDP growth, a high real interest rate, high inflation, strong growth of bank credit in the past, and a high ratio of broad money to reserves are all associated with a high prob- ability of a banking crisis (table 2). Exchange rate depreciation, the terms of trade, the fiscal surplus, and GDP per capita are not significant. The estimated probability of a crisis for the 36 episodes included in the sample ranges from a low of 1.1 percent for Nigeria to a high of 99.9 percent for Israel (see table 1). About 70 percent of the episodes have an estimated probability of 4 percent or more, while only 17 percent have an estimated probability of more than 50 percent. 292 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 2. Logit Regression of the Probability of a Banking Crisis Explanatory variable Estimated coefficient GDP Growth -0.172* (0.034) Change in terms of trade -0.021 (0.018) Depreciation 0.007 (0.006) Real interest rate 0.065* (0.016) Inflation 0.020** (0.010) Ratio of fiscal surplus to CDP 0.066 (0.036) Ratio of M2 to reserves 0.013* (0.005) Credit growth,_ 2 0.015** (0.008) GDP per capita -0.039 (0.033) Number of crises 36 Number of observations 766 Model x2 61.46* AICa 249 'Significant at the 1 percent level. * *Significant at the 5 percent level. Note: Standard errors are in parentheses. a. Akaike's Information Criterion. Source: Authors' calculations. Sources of Fragility: The 1994 Mexican Crisis One of the advantages of the multivariate logit model is that we can easily identify the sources of fragility by calculating the contribution of each explana- tory variable to a change in the estimated probability of a crisis. As an illustra- tion, we analyze the factors that contributed to the sharp increase in the esti- mated probability of a crisis in Mexico in 1993, just before the actual crisis occurred in 1994 (table 3). In 1993 high past credit growth, high real interest rates, and high inflation were the main factors underlying the high probability of a crisis in Mexico. Be- cause the logit is nonlinear, the sum of the contribution of each variable to the change in probability does not always add up to the total change (see the last column of table 3). Looking at macroeconomic factors, we see that Mexico had a negative growth shock that significantly raised the probability of a crisis. Real interest rates also rose significantly, and there was a minor terms-of-trade shock. At the same time, appreciation of the exchange rate, lower inflation, and a lower budget surplus offset some of this increase. Financial sector variables played a less important role in explaining the overall increase in probability, slightly offsetting the impact of the macroeconomic fac- Demirgfii-Kunt and Detragiache 293 Table 3. Decomposition of the Estimated Probability of a Banking Crisis, Mexico, 1992-93 Percentage Contribution to change in Change in percentage variable, Weight Weight in weight, change in Explanatory variable 1992-93 in 1993a 1992a 1992-93 probability GDP growth -125 0.154 -0.624 0.778 105 Change in terms of trade -16 -0.034 -0.041 0.007 1 Depreciation -119 -0.002 0.010 -0.012 -1 Real interest rate 386 0.327 0.067 0.259 28 Inflation -31 0.202 0.295 -0.093 -8 Ratio of fiscal surplus to GDP -79 0.022 0.102 -0.080 -7 Ratio of M2 to reserves -16 0.057 0.068 -0.011 -1 Credit growth1_2 -4 0.498 0.517 -0.019 -2 GDP per capita -1 -0.070 -0.070 0.000 0 Estimated probability of a crisis 1992 0.054 1993 0.116 a. Weights are obtained by multiplying the estimated regression coefficient of each variable by the value of the variable. A negative weight indicates that the variable reduced the estimated probability of a crisis. Source: Authors' calculations. tors. The vulnerability of the financial system to capital outflows-measured by M2 divided by the reserves ratio-fell slightly, leading to a 1 percent decrease in the probability of a crisis. Credit growth slowed, reducing the probability by 2 percent. Finally, GDP per capita-which we use as a proxy of institutional devel- opment-did not change significantly in this period. Thus decomposing the prob- ability of a crisis helps us to understand which factors played a role in bringing about the crisis, at least according to the empirical model. Out-of-Sample Probability Forecasts Because the purpose of monitoring is to assess future fragility, the next step is to forecast the probability of a banking crisis. Let ,B be a 1 x N vector containing the N estimated coefficients of the logit regression reported in table 1, and let zit be an N x 1 vector of out-of-sample values of the explanatory variables for coun- try i at date t. These values can be true forecasts, estimates of past values, data for countries or time periods not included in the sample, or ranges of values to con- struct alternative scenarios. Then the out-of-sample probability of a banking cri- sis for country i at date t is (1) Pit = exp[l zit] 1+ expIjp zi,] Once we compute out-of-sample probabilities, the question arises of how to interpret them. Is a 10 percent probability of a crisis high or low? Should a 294 TliE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 policymaker take preventive actions when faced with such a probability? Should a surveillance agency issue a warning? In the next section we address such questions III. BUILDING AN EARLY WARNING SYSTEM USING ESTIMATED CRISIS PROBABILITIES The first monitoring framework that we consider is one in which the decisionmaker must decide whether the forecasted probability is large enough to issue a warning. This is the framework implicit in Kaminsky and Reinhart (1999). Issuing a warning will lead to some sort of preventive action. For instance, the decisionmaker may invest in gathering further information, such as acquiring bank-level balance sheet data or holding discussions with senior bank managers, bank supervisory agencies, or other market participants. Alternatively, the decisionmaker may use the monitoring system to decide whether to take preven- tive policy measures, such as tightening prudential capital or liquidity require- ments for banks or reducing interest rates to ease pressures on bank balance sheets. For a warning system to be useful, preventive measures must substantially reduce the costs of a crisis. We assume that this is the case. Also a useful warning system should minimize false alarms, since preventive measures are usually costly. Tighter prudential requirements may cause banks to cut credit, perhaps leading to a credit crunch; looser monetary policy may lead to higher inflation. The choice of the threshold for issuing a warning will generally depend on three factors. The first is the probability of type I and type II errors associated with the threshold, which, assuming that the sample of past crises is representa- tive of future crises, can be assessed on the basis of the in-sample frequency of the two errors. Clearly, the higher is the threshold that forecasted probabilities must cross before a warning is issued, the higher will be the probability of a type I error and the lower will be the probability of a type II error (and vice versa). The second parameter on which the choice of the threshold depends is the unconditional probability of a banking crisis, which can also be assessed based on the in-sample frequency of crisis observations. If crises tend to be rare events, then the overall likelihood of making a type I error is relatively small (and vice versa). Finally, the third factor is the cost to the decisionmaker of taking preven- tive actions relative to the cost of failing to anticipate a banking crisis. In general, these costs to the decisionmaker are themselves forecasts of the true costs, and making a good decision requires having good forecasts. A policymaker who tends to underestimate the cost of a crisis or to overestimate the cost of taking preven- tive actions will be too conservative in choosing a warning threshold (and vice versa).6 A Loss Function for the Decisionmaker Based on these considerations, we can develop a more formal analysis of the decision process behind the choice of a warning system. Let T be the threshold 6. For estimates of the fiscal costs of recent banking crises, see Caprio and Klingebiel (1996). Demirgiu-Kunt and Detragiache 295 chosen by the decisionmaker, so that if the forecasted probability of a crisis for country i at time t exceeds T, the system will issue a warning. Let p(T) denote the probability that the system will issue a warning, and let e(T) be the joint prob- ability that a crisis will occur and the system will not issue a warning. Further, let cl be the cost of taking preventive actions as a result of having received a warn- ing, and let c2 be the additional cost of a banking crisis if it is not anticipated (if anticipating a crisis can prevent it altogether, then c2 is the entire cost of the crisis). Presumably, c1 is substantially smaller than c2 if further information gath- ering will be useful and if the knowledge that a crisis is impending will allow policymakers to take effective preventive measures. Then we can define a simple linear expected loss function for the decisionmaker as (2) L(T) =_ p(T)c, + e(T)c2. Let a(T) be the type I error associated with threshold T (the probability of not receiving any warning conditional on a crisis occurring), and let b(T) be the prob- ability of a type II error (the probability of receiving a warning conditional on no crisis taking place). Also let w denote the (unconditional) probability of a crisis. Then we can rewrite the loss function of the decisionmaker as (3) L(T) = c[(l - a(T))w + b(T)(1 - w)] + c2a(T)w =WC,[ 1+ CC )a(T) + b(T{ Iw ) The second part of the equality shows that the higher is the cost of missing a crisis relative to the cost of taking preventive action (the larger is c2 relative to c1), the more concerned will the decisionmaker be about a type I error relative to a type II error (and vice versa). Also the higher is the unconditional probability of a banking crisis (measured by the parameter w), the more weight will the decisionmaker place on type II errors, as the frequency of false alarms is greater when crises tend to be rare events.7 Using in-sample frequencies as estimates of the true parameters, w should equal the frequency of banking crises in the sample, namely 0.047 (see table 1). We can obtain the functions a(T) and b(T), which trace how error probabilities change with the threshold for issuing warnings, from the in-sample estimations as fol- lows. Given a threshold of, say, T = 0.05, we can derive a(0.05), that is, the associated probability of a type I error, as the percentage of banking crises in the sample with an estimated probability below 0.05. Similarly, b(O.05), the prob- ability of issuing a warning when no crisis occurs, is the percentage of observa- tions in which no crisis occurs when the estimated probability of a crisis is above 0.05. For T e [0, 11, a(T) is increasing, since the probability of not issuing a 7. A risk-averse decisionmaker would place greater weight on minimnizing type I errors than on minimizing type II errors, since type I errors are more costly. We are indebted to a referee for suggesting this point. 296 THE WORLD BANK ECONOMIC REVIEW, VOL. 14. NO. 2 Figure 1. Crisis Thresbold anzd In-Sample Classification Accuracy Probability of error 0.9 0.8 0.7 o |~~ ~ ~ ~~~ e;rro 0.6 0.5 o.4 0.3 0.2 0.1 Type 11 error 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Probabilitv threshold Source. Authors calculations. warning when a crisis occurs increases as the threshold rises, while b(T) is de- creasing (figure 1). The two functions cross at T = 0.036, at which the probabil- ity of either type of error is about 30 percent. Figure 1 also shows that probabilities estimated through our multivariate logit framework can provide a more accurate basis for an early warning system than the indicators developed by Kaminsky and Reinhart (1999). The indicator asso- ciated with the lowest type I error in the Kaminsky-Reinhart framework is the real interest rate, with a type I error of 70 percent and a type II error of 19 percent. In our model a threshold for a type I error of 72 percent comes at the cost of a type II error of only 1.2 percent. Similarly, the best indicator of banking crises according to Kaminsky and Reinhart is the real exchange rate, with a type I error of 73 percent and a type II error of 8 percent (resulting in an adjusted noise-to-signal ratio of 0.30). With our model we can obtain a type II error of 7.4 percent by choosing a probability threshold of 0.09, which is associated with a type I error of only 53 percent. The adjusted noise-to-signal ratio is 0.25. The better performance of the multivariate logit model likely stems from the fact that it combines into one number (the estimated probability of a crisis) all of the information provided by the economic variables monitored.8 8. The logit parameters are estimated using maximum likelihood, and the likelihood ftunction does not take into account the different costs of type I and type II errors. One way to improve the warning system could be to choose parameters that minimize the decisionmaker's loss functions. Demirgii4-Kunt and Detragiache 297 Figure 2. Loss Functionsfor Varyinlg Cost Parameters Loss function |~~~~~~~~~~~~~C - C7C- 20. - 10 C2- C1 = 5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Probability threshold Note: The value c2 - cl measures the cost to the decisionmaker of failing to identify a crisis relative to the cost of taking precautionary mrieasures. Source: Authors calculations. Choosing the Optimal Threshold To illustrate, we compute loss functions for three configurations of the decisionmaker's cost parameters (figure 2). We normalize the parameter cl to 1 in all three scenarios and give c2 - cl the values 20, 10, and 5. The values of the warning threshold that minimize the loss functions are, respectively, T = 0.034, 0.09, and 0.20. In other words, a decisionmaker whose cost of missing a crisis is, for example, 10 times the cost of taking precautionary measures will issue an alarm every time the forecasted probability of a crisis exceeds 9 percent. Thus, as expected, as the cost of missing a crisis increases relative to the cost of taking preventive action, the optimal threshold falls, resulting in a warning system with fewer type I errors and more type II errors. For values of c2 between 40 and 15, keeping cl constant at 1, the optimal probability threshold for issuing a warning is T = 0.034 (figure 3). With this criterion the probability of not issuing a warning when a crisis occurs is about 14 percent, while the probability of mistakenly issuing a warning is 31 percent. As c2 falls below 15, the threshold increases to 0.09 (type I error of 50 percent and type II error of 7.4 percent) and remains there until c2 reaches 8. At that point the threshold jumps to 0.20, as the decisionmaker becomes very concerned about false alarms. Finally, if the cost of missing a crisis is as low as two to three times 298 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Figure 3. Optimal Probability Thresholds for Varying Cost Parameters Probability threshold 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0 5 10 15 20 25 30 35 40 C2 - Cl Note: The value C2 - measures the cost to the decisionmaker of failing to identify a crisis relative to the cost of taking precautionarv measures. Source: Authors' calculations. that of issuing a false warning, then the optimal threshold is 0.30, corresponding to a type I error as high as 72.2 percent and a type II error as low as 1.2 percent. To fully appreciate the nature of the warning system, it is worth pointing out that the probability of a type I error is not the probability of missing a crisis. To obtain the probability of missing a crisis, we must multiply the probability of a type I error by the unconditional probability of a crisis, which in our sample is 0.047. Similarly, the probability of issuing a false warning is the size of a type II error multiplied by the frequency of noncrisis observations. With a threshold of T 0.09, the probability of missing a crisis is, therefore, only 2.3 percent, since crises occur rarely. In contrast, the probability of issuing a false alarm is 7.1 percent, because observations of no crisis tend to be the majority. Thus warning systems associated with a relatively low incidence of type I errors (below 15 percent) give rise to many false alarms, in part because criscs are infrequent events. If the system is used as a preliminary screen, and further information gathering can help to sort out cases in which the banking system is sufficiently sound, then the decisionmaker will accept the high incidence of type II errors. In some cases what the model considers to be a false alarm may actually be a useful signal. To illustrate this point, we examine the false alarms generated in- sample by a threshold of 0.047. As it turns out, in 21 cases the false alarm oc- curred in the two years immediately preceding a crisis, suggesting that the condi- tions that eventually led to a full-fledged crisis were in place (and were detectable) a few years in advance. In other cases the false alarms may have corresponded to episodes of fragility that were not sufficiently severe to be classified as full-fledged crises in our empirical study. Or they may have corresponded to episodes in Demirgiu-Kunt and Detragiache 299 which a crisis was prevented by a prompt policy response. Thus assessing the accuracy of the warning system based on the accuracy of in-sample classification may exaggerate the incidence of type II errors. However, out-of-sample predic- tions are subject to additional sources of error relative to in-sample predictions: the forecasted values of the explanatory variables include forecast errors, and there may be structural breaks in the relationship between banking sector fragil- ity and the explanatory variables, making predictions based on past behavior inadequate. Also, despite the large size of our panel, the number of systemic banking crises (36) is still relatively small, so that small-sample problems may affect the estimation results. As more data become available and the size of the panel is extended, this problem should become less severe. Comparing the Loss Function with the Noise-to-Signal Ratio It is of interest to compare the optimal threshold derived from minimizing the loss function proposed here with the optimal threshold that would result from minimizing the (adjusted) noise-to-signal ratio, the criterion used by Kaminsky and Reinhart (1999). Define the noise-to-signal ratio as (4) NS(T) _ a (T) 1 1-b(T) Then the loss function can be rewritten as (5) L(T) = wc1 + w(c2 - cl)a(T) + (1- w)c1NS(T)[1 - a(T)], and the first-order condition for the minimization of the loss function is (6) dL(T) [w(c2 - c1) - (1- w)c1NS(T)]a'(T) + (1 - w)c[11 - a(T)]N'S(T) = 0. dT Suppose T", is the threshold that minimizes the noise-to-signal ratio. Then, at T= T;S, NS'(T) = 0, and the derivative of the loss function is (7) dT )T=T"' Mw(c2- c) - (1- w)c,NS(T)]a'(T). For a convex loss function a positive (negative) sign for equation 7 means that the threshold TPs is too large (small) relative to the threshold that would mini- mize the loss function. Accordingly, if equation 7 is positive (negative), by mini- mizing the noise-to-signal ratio, the decisionmaker will make too many type I (type II) errors relative to the threshold that minimizes the loss function. Since a'(T) > 0 (the probability of a type I error is increasing in the threshold), equation 7 has the sign of the term in square brackets. This term is more likely to be positive the larger is the cost of a type I error (c2 - cl) relative to the cost of a type II error (c1) and the larger is the unconditional probability of a crisis. Thus if banking crises tend to be rare, and the cost of missing a crisis is high relative to 300 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 the cost of raising a false alarm, minimizing the noise-to-signal ratio is likely to yield a choice criterion that results in too many missed crises for a decisionmaker whose preferences are captured by the linear loss function of equation 5. IV. CONSTRUCTING A SYSTEM FOR RATING BANK FRAGILITY In this section we consider the problem of a monitor who must rate the fragil- ity of a given banking system. Other agents will then use the rating to decide on a possible policy response, but the monitor is not necessarily aware of the costs and benefits of such policy actions. Another rationale for using fragility classes instead of a critical threshold as a monitoring device is that small changes in the critical threshold may lead to substantial differences in type I and type II errors, as seen in figure 1. In constructing fragility classes, the classification criterion should have a clear interpretation in terms of type I and type II errors. This has two advantages: first, agents who learn the rating can make their own cost- benefit calculations when they decide whether or not to take action, and, second, the fragility of two systems that are assigned different ratings can be compared based on a clear metric. The starting point is once again the set of forecasted crisis probabilities ob- tained using the coefficients estimated in the multivariate logit regression. Clearly, a country with a forecasted probability of x should be deemed more fragile than one with an estimated probability of y < x. To establish fragility classes, we can partition the interval [0, 1], which is the set of possible forecasted probabilities, into a number of subintervals and assign a rating to all estimated probabilities within a given class. There are no objective criteria for choosing one partition over another, but a number of considerations help to narrow the choices. First, because the frequency of crises in the sample is low, choosing a fine partition would give misleading results, because many classes would have no observed crises. For instance, in our sample there are no episodes with an estimated crisis probability between 4 and 5 percent, whereas there are episodes with an estimated probability between 3 and 4 percent (see figure 1). If we choose the intervals [0.04-0.05] and [0.03- 0.04] as two of the classes, then it would appear that fragility decreases with the estimated probability of a crisis-an obviously misleading conclusion. Another caveat is that the empirical distribution of the estimated probabilities is strongly skewed toward zero: only 8.5 percent of the observations have prob- abilities higher than 10 percent, and more than 45 percent are in the 0-2 percent range. Thus partitioning the unit interval by subsets of the same size&would as- sign an uneven number of observations to each class, with very few observations in the highest probability intervals. Based on these considerations, we construct a rating system with four fragility classes (table 4). We choose the upper bounds of each of the four classes so that the type I error associated with the bounds are 10, 30, 50, and 100 percent, respectively. According to this criterion, observations with forecasted probabili- Demirgui-Kuntand Detragiache 301 ties below 1.8 percent belong to the lowest fragility class. Observations with probabilities between 1.8 and 3.6 percent are in the second class, up to 7 percent are in the third class, and above 7 percent are in the highest class. The values of the type II error associated with the upper bound of each class are (about) 60, 30, 12, and 0 percent, respectively. To illustrate the meaning of the fragility groupings, consider that if all obser- vations with forecasted probabilities in classes higher than the most fragile (that is, observations with probabilities higher than 1.8 percent) were treated as crises, the likelihood of missing a crisis (given that one takes place) would be less than 10 percent. However, the probability of falsely predicting a crisis would be higher than 60 percent. Another way to put it is that 90 percent of the crisis observa- tions in the sample have a probability higher than the probabilities in the lowest fragility class. Similarly, if one were to classify as crises only observations with forecasted probabilities in the two highest fragility classes, then the probability of missing a crisis would rise to 30 percent, and the probability of a false alarm would fall to 30 percent. As an additional measure of the degree of fragility associated with each class, we compute the fraction of sample observations in each class that corresponds to an actual banking crisis. This measure ranges from 1.5 percent for the lowest fragility class to 16.8 percent for the highest. Thus the likelihood that an obser- vation in the highest fragility class is a crisis is 16.8 percent; this figure may seem low, but it should be compared to the unconditional probability of a crisis: 4.7 percent (the sample frequency of crises). To put it another way, finding that the probability of a crisis falls in the highest fragility class tells the analyst that the observation is 3.5 times more likely to correspond to a crisis than the average observation. Clearly, these rating systems are just examples of many possible alternatives, and depending on the purposes of the monitor, one alternative may be preferred to another. What is important is that potential users understand the meaning of the fragility score and the criteria used in rating. V. APPLYING THE SYSTEM TO THE BANKING CRISES OF 1996-97 To gauge the performance of our monitoring mechanisms, we consider how accurately they would have predicted the six banking crises that took place in Table 4. A Rating System for Banking Sector Fragility Probability, Type I Type II Number of Crisis per Class interval error error observations observation l 0.000-0.018 0.00-0.10 1.00-0.60 291 0.01 II 0.018-0.036 0.10-0.30 0.60-0.30 232 0.03 III 0.036-0.070 0.30-0.50 0.30-0.12 136 0.05 IV 0.070-1.000 0.50-1.00 0.12-0.00 107 0.17 Note: Class I is the lowest fragility class and class IV is the highest. Source: Authors' calculations. 302 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 1996-97, that is, after the end of the sample period used in the estimation exer- cise above. The banking crises occurred in Jamaica in 1996 and in Indonesia, the Republic of Korea, Malaysia, the Philippines, and Thailand in 1997. Early ac- counts and analyses of the events surrounding the five Asian crises can be found, for instance, in IMF (1997), Radelet and Sachs (1998), and Goldstein and Hawkins (1998). To compute the probabilities of out-of-sample banking crises for the six coun- tries, we use two sets of values for the explanatory variables. The first set consists of actual realizations. The out-of-sample probabilities obtained in this way are not true forecasts, of course. In particular, for the five Asian countries these fig- ures capture the large exchange rate depreciations that took place in the second half of 1997 and their immediate consequences. It is of interest to try to assess whether signs of increasing banking sector fragility would have been apparent before the depreciations took place, since they were largely unanticipated by ob- servers. To this end, and, more generally, to assess the performance of the moni- toring system when true forecasts are used, we also compute out-of-sample prob- abilities using forecasts of the explanatory variables as of April-May 1997. Comparing the two forecasts will reveal the extent to which errors in forecasting the explanatory variables would have clouded the fragility assessment based on our model. We take the forecasted values of the explanatory variables, where available, from the Financial Times' Currency Forecaster, and from Consensus Forecasts. These works survey several prominent private sector forecasters and publish the means of their forecasts. For the five Asian countries the growth rate of real GDP, inflation, exchange rate depreciation, and the real interest rate are from the Cur- rency Forecaster, and broad money is from Consensus Forecasts. The remaining values (and all of the values for Jamaica) are from the May 1997 round of the IMF's semiannual World Economic Outlook.9 To compute the probabilities of out-of-sample crises using realized values of the explanatory variables, we use numbers from the International Financial Statistics when available and February 1998 numbers from the World Economic Outlook otherwise. Based on forecasts as of April-May 1997, the estimated probabilities of crises were relatively low for the five Asian countries, whereas Jamaica was well into the highest fragility zone as early as 1995 (figure 4). This is not surprising, since all the Asian countries had very good macroeconomic performances in the years up to 1996-performances that, by and large, were expected to continue. In Jamaica the forecasted probability of a crisis was 14 percent in 1995 and 13 percent in 1996. The two main factors contributing to the increase in the prob- ability of a crisis were high real interest rates and high inflation. Strong past 9. There are two exceptions. For the Philippines, broad money comes from the World Economic Outlook. For Korea, no forecast of reserves was available, so we arbitrarily assumed that reserves returned to their 1995 value in 1997. Figure 4. Actual and Forecasted Crisis Probabilities in Five Asiani Countries and Jamaica, 1990-97 Jamaica' Indonesia Korea Percent Percent Percent 18 18 - 18 - 18 i4 ~~~~~~~~~~~14 t14 12 12 12 10 M i~~ ~~~~~~~~~ Actual 10 6 ~~~~~~~~Actual 4 4 2 2Frest2 0 0 0F 1990 1991 1992 1993 1994 1995 1996 1990 1991 1992 1993 1994 1995 1996 1997 1990 1991 1992 1993 1994 1995 1996 1997 o = ._ __ _- - , _ -__- _ --_ Malaysia Philippines Thailand Plercent Percent Percent 18 - 18 - __ _ _ _ _ _ _ __ _ _ _ _ _ _ _ 16 16i 1 I 4 14 12 12 12 10 10 lo-- 8 8 . Actual 6 6 Actua 2 ~~~~~~~Actual; 22 0 ~~~~~~~. ..m- Forecast 0ForecasI 1990 1991 1992 1993 1994 1995 1996 1997 1990 1991 1992 1993 1994 1995 1996 1997 1990 1991 1992 1993 1994 1995 1996 1997 a. Jamaican data run through 1996, the year of that colntry's crisis. Source: Authors' calculations. 304 THE WORLI) BANK ECONOMIC REVIEW, VOL. 14, NO. 2 credit growth and a favorable fiscal position also contributed to fragility in 1995, but not in 1996. The two most fragile Asian countries were Thailand and the Philippines, both having a forecasted crisis probability of about 3.5 percent in 1997. This prob- ability would have placed the two countries on the border between the second and third fragility zones based on our rating system. In Thailand the main factor contributing to bank fragility both in 1996 and in 1997 was the high real interest rate; strong past credit growth was also a factor. But in contrast with Jamaica, where GDP growth was lackluster, Thailand had a large predicted GDP growth rate, which worked as an offsetting factor, keeping the overall probability of a crisis relatively low. In the Philippines the predicted probability increased more than 20 percent between 1996 and 1997, mainly because of the high growth rate of credit two years earlier. The real interest rate was lower than that in Thailand, but so was GDP growth. Indonesia, Malaysia, and Korea all had forecasted crisis probabilities below 3 percent in 1996 and in 1997, and would have been placed in the second fragility class (actually, Malaysia would have received the lowest fragility rating in 1996). As in Thailand and the Philippines the expectation that the exchange rate would remain stable and, especially, that GDP growth would continue to be strong more than offset the prospect of fragility coming from high real interest rates (except in Korea) and strong past credit expansion. Indonesia's high rate of inflation also tended to increase bank fragility. Not surprisingly, the picture obtained by estimating the probabilities of crises using the latest available data would have been quite different for the five Asian countries, but not for Jamaica. The estimated probabilities of crises are in the highest fragility class for Indonesia and Thailand and in the second highest for the other three Asian countries. Malaysia, with a probability of 3.7 percent, ap- pears to have been the least fragile.10 Decomposing the probability tells some interesting stories. Of course, the ex- change rate depreciation directly affected fragility in all five countries. However, in 1997 inflation was not much higher than forecasted, so it was not among the main factors contributing to greater banking system vulnerability. In all five coun- tries except Korea lower-than-forecasted GDP growth was one of the main con- tributing factors, as was the higher-than-expected real interest rate (except in Thailand). To summarize, an analysis of banking system fragility using the methods de- veloped in this article would have clearly indicated an impending banking crisis in Jamaica. But although signs of fragility were present in Thailand and the Phil- ippines, the overall image of the five Asian economies would have been fairly reassuring, as expectations of continued strong economic growth and stable ex- change rates would have offset the negative impact of relatively high real interest rates and strong past credit expansion. 10. Of the five Asian countries, Malaysia is the only one without an IMF program. Demirga -Kunt and Detragiache 305 VI. CONCLUSIONS Econometric analysis of systemic banking crises is a relatively new field of study, and the development and evaluation of monitoring and forecasting tools based on the results of such analyses are at an embryonic stage at best. The purpose of this article has been not so much to propose one or more "ready-to- use" procedures for decisionmakers, but rather to highlight which elements must be evaluated in developing such procedures and to explore some possible av- enues. Specifically, we have developed two monitoring tools using forecasted probabilities obtained from a multivariate logit model of banking crises. The first is an early-warning system that issues a signal when the probability of a fore- casted crisis exceeds a certain threshold. The appropriate threshold for issuing a warning can be chosen based on the costs of missing a crisis and the benefits of avoiding a false alarm. The second monitoring tool is a system for rating bank fragility. Both monitoring tools can be used to economize on precautionary costs by pointing to cases of high fragility that warrant more in-depth monitoring. Evaluating banking sector fragility along these lines is subject to several poten- tial errors common to all exercises based on forecasts. First, the regression coef- ficients used to compute the forecasted probability of a crisis are only estimates of the true parameters. Second, new crises may be of a different nature than those experienced in the past, so that the coefficients derived from in-sample esti- mation may be of limited use out-of-sample. This problem may be particularly severe since banking crises tend to be rare events, and, even though the panel used for in-sample estimation is large (766 observations), crisis episodes only number 36. Third, forecasts of the explanatory variables are likely to incorporate errors, as vividly illustrated by the example of the five recent Asian crises. Large forecast errors, in turn, may severely distort the assessment of fragility.1" One way to reduce the impact of forecast errors is to develop alternative scenarios for the explanatory variables and to examine banking sector fragility in the context of such scenarios. This would be particularly useful, because in many cases banking crises are triggered by extreme behavior in one or more explanatory variables (a currency collapse, a bout of inflation, a drastic deterioration in the terms of trade) in a context in which other elements also contribute to overall fragility. Routine forecasts of economic variables rarely capture extreme events of this sort, which instead tend to be discussed as "risk elements" of the overall picture.12 The frame- 11. One direction in which this work can be extended is to explore alternative model specifications and compare them from the point of view of their usefulness for forecasting (see, for instance, Diebold 1997). Here we have used a specification developed in our previous work after eliminating explanatory variables for which forecasts were not readily available. It could be that an even more parsimonious specification is more suitable for forecasting purposes. We leave this issue to future extensions. 12. This is certainly true of IMF forecasts, which often tend to be excessively optimistic (Mussa and Savastano 1999). For the Asian countries we computed crisis probabilities using the most pessimistic forecasts from the Consensus Forecasts group, but this did not lead to a substantial increase in forecasted crisis probabilities. 306 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 work developed here would lend itself easily to the evaluation of fragility in alter- native scenarios, since it allows us to isolate the contribution of each explanatory variable to the forecasted crisis probability. Another important caveat is that, although aggregate variables can convey information about the general economic conditions that tend to be associated with banking sector fragility, they are silent about the situation of individual banks or specific segments of the banking sector. So they would not detect crises that may develop from specific weaknesses in some market segments and spread through contagion. Also informed observers who are familiar with a particular country are likely to be in a better position to detect signs of incoming trouble, so the information generated by a quantitative approach such as ours should comple- ment, not replace, other sources of information. A final message from this exercise is that, to be useful, a monitoring system must be designed to fit the preferences of the decisionmaker. Thus the develop- ment of a system must be the outcome of an interactive process that involves both econometricians and policymakers. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Calvo, Guillermo A. 1996. "Capital Flows and Macroeconomic Management: Tequila Lessons." International Journal of Finance and Economics 1(3):207-24. Caprio, Gerard, and Daniela Klingebiel. 1996. "Dealing with Bank Insolvencies: Cross- Country Experience." World Bank Policy Research Working Paper 1620. Washing- ton, D.C.: World Bank. Consensus Economics. Various years. Consensus Forecasts. London. Demirgiuc-Kunt, Asli, and Enrica Detragiache. 1998. "The Determinants of Banking Cri- ses in Developing and Developed Countries." IMF Staff Papers 45(1):81-109. . 1999. "Financial Liberalization and Financial Fragility." In Boris Pleskovic and Joseph E. Stiglitz, eds., Annual World Bank Conference on Development Economics 1998. Washington, D.C.: World Bank. Diebold, Francis. 1997. Elements of Forecasting. Cincinnati: South-Western College Pub- lishing. Eichengreen, Barry, and Andrew K. Rose. 1998. "Staying Afloat While the Wind Shifts: External Factors and Emerging-Market Banking Crises." NBER Working Paper 6370. National Bureau of Economic Research, Cambridge, Mass. Processed. Financial Times. Various years. Currency Forecaster. Alexandria, Virginia. Gavin, Michael, and Ricardo Hausman. 1996. "The Roots of Banking Crises: The Macro- economic Context." In Ricardo I lausman and Liliana Rojas-Suarez, eds., Volatile Capi- tal Flows: Taming Their Impact on Latin America. Washington, D.C.: Inter-American Development Bank. Goldstein, Morris, and John Hawkins. 1998. "The Origins of the Asian Financial Tur- moil." Reserve Bank of Australia, Sydney. Processed. Demirgui-Kunt and Detragiache 307 Hardy, Daniel, and Ceyla Pazarbasioglu. 1998. "Leading Indicators of Banking Crises: Was Asia Different?" IMF Working Paper 91. International Monetary Fund, Washing- ton, D.C. Processed. Honohan, Patrick. 1997. "Banking System Failures in Developing and Transition Coun- tries: Diagnosis and Prediction." BIS Working Paper 39. Bank for International Settle- ments, Basel. Processed. IMF (International Monetary Fund). 1997. "World Economic Outlook. Interim Assess- ment." Washington, D.C. Processed. -. 1998. World Economic Outlook. Washington, D.C. Kaminsky, Graciela L. 1998. "Currency and Banking Crises: The Early Warnings of Dis- tress." International Finance Discussion Paper 629. Board of Governors of the Federal Reserve System, Washington, D.C. Kaminsky, Graciela, and Carmen M. Reinhart. 1999. "The Twin Crises: The Causes of Banking and Balance of Payments Problems." American Economic Review 89(3):473- 500. Lindgren, Carl J., Gillian Garcia, and Michael Saal. 1996. Bank Soundness and Macro- economic Policy. Washington, D.C.: International Monetary Fund. Mishkin, Frederic S. 1996. "Understanding Financial Crises: A Developing Country Per- spective." NBER Working Paper 5600. National Bureau of Economic Research, Cam- bridge, Mass. Mussa, Michael, and Miguel Savastano. 1999. "The IMF Approach to Stabilization." NBER Macroeconomic Annual 14. Cambridge, Mass.: MIT Press. Radelet, Steven, and Jeffrey Sachs. 1998. "The Onset of the Asian Financial Crises." In Paul Krugman, ed., Currency Crises. Chicago: University of Chicago Press. Rojas-Suarez, Liliana. 1998. "Early Warning Indicators of Banking Crises: What Works for Developing Countries?" Research Department, Inter-American Development Bank, Washington, D.C. Processed. Sachs, Jeffrey, Aaron Tornell, and Andres Velasco. 1996. "Financial Crises in Emerging Markets: The Lessons from 1995." Brookings Papers on Economic Activity 1(1996): 147-98. THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2: 309-30 Trade Reform Dynamics and Technical Efficiency: The Peruvian Experience Ila M. Semenick Alam and Andrew R. Morrison Markets around the world are becoming more competitive because of cbanging operat- ing and regulatory environments. One such change-the loosening of trade restrictions- is a macroeconomic policy shift that should have a microeconomic impact on industrial efficiency. Specifically, competitive pressure should discipline or eliminate inefficient prodncers. This article explores whether or not there is such a dynamic link. It uses a previously unexploited data set to gauge the impact of the 1990 Peruvian reform on plant-level technical efficiency. The results support the argument that the degree of pro- tection and the level of efficiency are inversely related. The chill winds of competition should have favorable effects on industrial effi- ciency according to the neoclassical paradigm. Leibenstein (1966) was the first to state explicitly that "proper motivations" should discipline firms, forcing them to become more efficient or perish. Many industries-such as trucking, air travel, and banking-have faced one such motivational factor: reduced regulation. In a study of the U.S. airline industry, for example, Alam and Sickles (2000) find support for the hypothesis that resource utilization in the industry became more efficient as market forces compelled the airlines to economize after the 1978 deregulation. Similarly, Alam (2000) finds that the U.S. banking industry, which underwent substantial deregulation and notable financial innovations during the 1980s, experienced sustained productivity growth during most of that decade. In the international arena one motivational factor is the heightened competi- tion arising because countries have adopted trade liberalization strategies, the most conspicuous examples being the North American Free Trade Agreement (NAFTA) and the European Economic Community (EEC). In the case of developing countries several empirical studies have confirmed a positive link between trade reform and efficiency, yet many researchers continue to have doubts about the Ila M. Semenick Alam is with the Department of Economics at Tulane University, and Andrew R. Morrison is with the Social Development Division at the Inter-American Development Bank. Their e-mail addresses are ila.alam@tulane.edu and andrewm@iadb.org. The authors thank Peru's Ministry of Industry, Tourism, Integration, and International Trade Negotiations, particularly Jaime Garcia and Julia Hernandez, for supplying the data used in this study. They also thank Fatima Ponce, Jaime Saavedra, and Jorge Vega for their invaluable help and Gerald Granderson, Myriam Quispe-Agnoli, Mark Johnson, and the anonymous referees for useful comments on an earlier version of this article. (C 2000 The International Bank for Reconstruction and Development / THE WORLD BANK 309 310 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 impact of trade liberalization on performance.' Despite the intuitive appeal of the efficiency hypothesis, empirical studies have been limited in part by the scar- city of plant-level data from developing countries. This article contributes to the ongoing debate by examining the relationship between trade reform and industrial efficiency in Peru, focusing on the impact of the reform and liberalization program initiated in 1990 after many years of im- port substitution industrialization. Our article is unique in that we use a new panel data set and introduce to the literature a nonparametric, mathematical programming methodology that allows us to estimate time-varying, producer- specific efficiency levels. It is important to study efficiency dynamics because our data set spans a period in which Peru's policies toward protectionism changed dramatically. We study Peru because the empirical evidence concerning the rela- tionship between trade reform and industrial efficiency in developing countries is not definitive; most studies confirm a positive link, but some have failed to detect such a connection. Peru should provide a valuable test about the generalizability of the well-studied Chilean case (see, for example, Tybout, de Melo, and Corbo 1991 and Liu 1993). 1. THE DISMANTLING OF PROTECTIONISM IN PERU Peru was one of the last Latin American countries to abandon import substitu- tion industrialization as a development strategy. Import substitution was designed to protect domestic "infant" manufacturing industries from imports through the use of protective instruments, such as tariffs, quotas, exchange rate controls, and price and wage controls. The death knell for this policy sounded with the presi- dential election of Alberto Fujimori, who took office in July 1990 and imple- mented a far-reaching neoliberal reform package. Other leaders of Peru had attempted to introduce neoliberal reforms, but never completed those reforms. The Belaiinde administration (1980-85) substantially reduced tariffs on imports.2 In January 1981 the maximum tariff fell from 60 to 35 percent and then to 32 percent. By the end of 1981, 98 percent of all registered items could be imported without a duty, up from only 38 percent in 1978. The government relaxed regulation of foreign investment and announced plans to priva- tize state-owned enterprises. The impact on domestic industry was severe: manu- facturing output fell nearly 20 percent between 1980 and 1983, and idle capacity in manufacturing rose to more than 54 percent. Bowing to strong pressure from Peruvian industrialists, the Belaiinde administration abandoned the reform pack- age and restored nominal tariff rates to their levels before the reform. The Garcia government (1985-90) was characterized by populist policymaking, heterodox stabilization, and a continuation of import substitution industrializa- 1. Most notable among them are Pack (1988), Krugman (1994), and Rodrik (1995). This issue is discussed later in the article. 2. The discussion of trade reform under the Belaunde administration is drawn from Conaghan, Malloy, and Abugattas (1990) and Pastor and Wise (1992). Alam and Morrison 311 tion. Saavedra Chanduvi (1996:2) provides a concise and accurate description of the Peruvian industrial sector at the end of the Garcia government: The Peruvian manufacturing sector developed for three decades [1960-90] sheltered by a set of tariff and nontariff barriers that permitted it to enjoy very high-and in some cases infinite-levels of protection . .. The pattern of trade and industrial production were so distant from that dictated by comparative advantage that Peru came to produce automobiles and com- puters for the internal market. The Fujimori government took the first step toward dismantling import sub- stitution industrialization in October 1990 by consolidating tariff categories. The number of tariff categories dropped from 56 to 3, with rates of 15, 25, and 50 percent. In March 1991 these rates were reduced further to 5, 15, and 25 per- cent, with 82 percent of all goods subject to the 15 percent rate. The magnitude of the reform is revealed by the decline in average tariff rates, which fell from 66 percent in July 1990 to 17 percent in March 1991 (Escobal 1992 and Saavedra Chanduvi 1996). The mean effective rate of protection fell from more than 90 percent in July 1990 to 36 percent in December 1990 and then to less than 30 percent in March 1991-the month that additional structural reforms were added to the reform package (table 1). Disaggregated by three-digit Standard Industrial Classification (sic) codes, the percentage declines in effective rates of protection between July 1990 and March 1991 are impressive, falling almost 98 percent for nonferrous metals, 81 percent for clothing, 79 percent for other chemical products, and 76 percent for furniture. The smallest decline was in iron and steel production, al- though the drop was still 22 percent. The mean decline was 63 percent. Also notable, the standard deviation of the effective rates of protection fell from more than 70 percent in July 1990 to less than 20 percent in March 1991. Although tariff reform was the major departure from import substitution, other key elements of the 1990 reform package included removal of wage and price controls, increases in the prices of and elimination of subsidies to public services, reduction in public sector employment, unification of a multiple exchange rate system, efforts to increase tax collection, elimination of restrictions on capital flows, and liberalization of interest rates (Quijandria 1995). In 1991 and 1992 many of these reforms were deepened: the government removed interest rate ceil- ings on dollar-denominated deposits and loans, instituted a private pension sys- tem, and created an agency to regulate the behavior of private firms and protect consumers' rights. II. THEORETICAL LINKS BETWEEN TRADE REGIME AND INDUSTRIAL EFFICIENCY There are many arguments explaining why more open trade regimes lead to more efficient industrial production. Perhaps the most basic is that returns to 312 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Table 1. Effective Rates of Protection, by Three-Digit Standard Industrial Classification, Peru, 1990-91 (percent) SIC code and name July 1990 December 1990 March 1991 311-12 Food products' 132.7 42.8 34.5 313 Beverages and tobacco 197.2 62.6 49.5 321 Textilesb -42.2 -26.1 -24.0 322 Clothing 261.6 87.6 48.5 323 Leather products 151.2 45.1 53.8 331 Furniture 132.0 46.0 31.9 341 Paper products 63.9 40.4 35.7 342 Publishing 85.1 33.1 21.8 351 Basic chemicals 66.5 20.8 30.1 352 Other chemical products 131.0 34.4 27.3 355 Rubber products 88.7 35.9 30.3 356 Plastic products 88.7 35.9 30.3 369 Nonmetallic mineral products 83.1 34.0 27.2 371 Iron and steel production 42.0 21.2 32.8 372 Nonferrous metalsb -45.1 -1.5 -1.1 381 Miscellaneous metal products 89.3 50.2 35.3 382 Nonelectric machinery 30.2 21.0 20.8 383 Electric machineryc 91.8 40.8 34.6 390 Other manufactured products 73.1 50.9 36.9 Mean 90.6 35.5 29.3 Standard deviation 71.9 23.6 17.5 Note: The effective rate of protection measures the percentage by which value added can increase over the free-trade level as a consequence of a tariff structure. This rate for a sector is (VT - V,)/V,, where V. is value added under trade policies and V, is value added at world prices (Krugman and Obstfeld 1994). The effective rate of protection captures protection of intermediate and final goods. A negative rate implies that input industries are particularly favored. This table includes only those sectors for which price deflators are available. Thus SiC 324 (footwear) and sic 384 (transport equipment) are not included. a. The average of milk products, flour and bakery products, and other food products. b. The negative rates indicate higher tariffs on input imports than on final goods. The increases in these rates between July 1990 and March 1991 are the result of tariff reductions on inputs due to trade liberalization. c. The average of machinery and equipment for industrial and professional use and home appliances and devices. Source: Banco Central de Reserva del Peri. entrepreneurial effort increase as exposure to foreign competition rises (Corden 1974, Martin and Page 1983, and Tybout 1992a). A second argument is that increasing returns to scale imply lower costs per unit as output increases (Pack 1988 and Tybout 1992a). For this argument to be complete, however, a reduc- tion in protectionism must be accompanied by an increase in domestic output- a conclusion that is far from certain, since increased competition may force pro- ducers to exit instead of expand (Tybout 1992a). Several authors (Pack 1988, Grossman and Helpman 1991, and Edwards 1992) argue that greater openness may accelerate developing countries' adoption of technological innovations originating in industrial countries. From the viewpoint of the new growth theory the creation of larger markets through trade liberaliza- Alam and Morrison 313 tion (and market-based exchange rates) will raise demand for products, leading to more investment in product development and innovation (Tybout 1992a).3 Two other effects may be important: share effects and residual effects. If more efficient plants gain market share as a result of exposure to foreign competition, industrywide efficiency should rise, even if no scale economies are present (Bond 1986, Roberts and Tybout 1991, and Tybout and Westbrook 1995). Tybout and Westbrook (1995) coin the term "residual" effects, noting that this "catchall category" includes capacity utilization, externalities, learning-by-doing, and mana- gerial effort.4 III. EMPIRICAL LINKS BETWEEN TRADE REGIME AND EFFICIENCY Two approaches, mirroring two different techniques for measuring productiv- ity, have been used to test empirically for a relationship between the type of trade regime and industrial productivity: studies of total factor productivity (TFP), usu- ally based on secondary data sources available at the one- or two-digit SIC level, and estimations of production functions using plant-level data. Total Factor Productivity Studies Several studies have estimated TFP and linked its evolution to changes in trade regime. One of the earliest is by Michaely (1975), who finds a correlation be- tween the reduction in quantitative restrictions and increased TFP in Israel in the mid-1950s. Nishimizu and Robinson (1986) decompose TFP growth into the shares accounted for by the expansion of domestic demand, the expansion of exports, and import substitution. They use data from Japan, the Republic of Korea, Tur- key, and Yugoslavia for subperiods between 1955 and 1973. Their results are intriguing: although domestic demand accounts for the largest share of TFP growth for all countries in all subperiods, export expansion frequently plays an impor- tant role. Import substitution, in contrast, contributes negatively to growth in most cases. More recently, Edwards (1994) calculates differences in TFP growth between 1987-91 and 1978-82 for manufacturing sectors in six Latin American coun- tries. Although he warns against inferring causal relationships, he does remark that Chile and Costa Rica, the two countries that began trade reform earliest among the six, had the largest increases in the rate of TFP growth.5 Haddad, de Melo, and Horton (1996) study the productivity of Moroccan manufacturing sectors following a modest and gradual reform process aimed at heightening com- petition and improving technical efficiency. Despite data limitations, the authors find evidence suggesting that TFP grew as a result of the reform program. In an 3. However, entrepreneurs may be less likely to develop new products because trade liberalization makes available a wider variety of imported substitutes (Tybout 1992a). 4. They also include technological innovation in this class, but we follow Edwards (1992) and Grossman and Helpman (1991), who include it as a separate class. 5. The other four countries in the study are Argentina, Bolivia, Mexico, and Uruguay. 314 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 effort to determine the robustness of the positive link between trade liberaliza- tion and economic performance, Edwards (1997) uses several estimation tech- niques, time periods, functional forms, and measures of openness. He extends his analysis to include 93 countries and finds that countries with greater trade barri- ers experienced slower productivity growth. As Edwards (1994) points out, one drawback of TFP studies is their aggregate nature, which may obscure diverse sectoral responses to trade policy.6 Further- more, as Pack (1988) and Havrylyshyn (1990) observe, TFP studies have pre- sented mixed results. Havrylyshyn (1990:10) sums up this problem: The evidence [on the relationship between trade reform and efficiency] from studies of TFP is weak and ambiguous. Some evidence of positive links between trade policy and productivity growth certainly exists . . . But many cases ... are ambiguous, and some suggest a negative relation. Production Function Studies The second major approach to investigating the relationship between trade policy and industrial efficiency calls for estimating plant-level production func- tions, deriving estimates of efficiency from those functions, and examining the links between the efficiency estimates and trade policy. Tybout, de Melo, and Corbo (1991) estimate deterministic production functions at the three-digit sic level for Chile for the years 1967 and 1979. From the production function pa- rameters, they create sic-level indexes of efficiency levels and dispersion of effi- ciency levels with 1967 as the base year. They find that reductions in tariff pro- tection are correlated with increases in efficiency and decreases in the variance of efficiency scores. The authors conclude that additional plant-level studies-using panel data and other trade liberalization events-must be performed to help verify their findings. Tybout and Westbrook (1995) use a similar technique to analyze reform in Mexico. They estimate production and cost functions for 1984 (before trade reforms were implemented) and for 1990 (after reforms were implemented) and then use correlation analysis to determine whether changes in efficiency at the three-digit sic level are correlated with changes in various measures of trade policy. In addition to deterministic measures of efficiency, stochastic frontier produc- tion functions have been a popular technique for estimating production tech- nologies. Early efforts use cross-sectional data collected before and after liberal- ization to examine changes in efficiency levels. Handoussa, Nishimizu, and Page (1986), for example, find significant increases in technical efficiency among Egyp- tian public sector firms during a period of trade liberalization beginning in 1973. More recent work has taken advantage of the availability of firm-level panel data. Liu (1993), for example, assumes a Cobb-Douglas production function and, using 6. This is not an insurmountable problem. Nishimizu and Robinson (1986), for example, examine determinants of TFP growvth by country at the sectoral level. Although they do not include explicit trade policies in their list of determinants, such an analysis is possible. Alam and Morrison 315 panel data for 1979-86 from Chile, finds that the mean efficiency levels of sur- viving plants tend to be higher than those of exiting plants.7 By and large, pro- duction function approaches have produced empirical results confirming that trade liberalization improves efficiency (Havrylyshyn 1990).Y IV. DATA The data we use to estimate plant-level efficiency were provided by the Peru- vian Ministry of Industry, Tourism, Integration, and International Trade Nego- tiations (MITINCI), which conducts an annual survey of manufacturing plants with more than 20 employees.9 Our data cover 1988-92-two years before and two years after the implementation of economic reform in July 1990. The coverage of the survey is reasonably good since completion was obligatory. Our output variable is the value of total production. We use six inputs as explanatory variables: capital stock, raw materials, electricity, payments for in- dustrial services, and blue- and white-collar workers.10 All of the data are re- corded in value terms with the exception of the labor variables. Capital stock is the value reported for the end of the previous calendar year, and workers are measured in physical (not efficiency) units. Tybout (1992b) points out that firm- level data on capital stocks of developing countries may be subject to measure- ment error. To check the consistency of the capital measure in our data set, we calculate capital-to-output ratios (KIQ) for all establishments. There were some unrealistic ratios, so we dropped the observations corresponding to the largest and smallest 10 percent of all KIQ values from the sample.'" 7. Liu's classification of exiting, entering, and surviving firms is based on "intertemporal patterns of missing values for each plant" (Liu 1993:220). Thus if a plant exits the sample and does not reappear, it is classified as an exiting plant. However, this plant just may have stopped filing information while continuing to produce. Similarly, entering plants may be survivors that did not file in preceding years. Still, there is no superior panel data set for Latin American industries. 8. Pack (1988:372) dissents from this view. 9. MtriNCI, Estadistica manufacturera, datos a nivel de establecimiento (various years). To our knowledge, we are the first researchers to be given access to these data. 10. We do not use other input categories reported in the data set, such as fuel, replacement parts and accessories, and containers, because there are many missing values. We deflate all of the variables, except for labor, using data obtained from MITINCi, Estadistica industrial mensual (various years). The MITINCI deflators were available for 1988, 1989, 1990, and 1993; deflators for 1991 and 1992 had not been processed so we have interpolated to obtain them. Since price deflators are not available at the firm level, some of the measured cross-firm variation in productivity is capturing firm-specific variation in prices; without firm-level price indexes, we cannot address this problem. (This observation was contributed by an anonymous referee.) 11. Tybout (1992b) provides a more elegant, econometric solution to this measurement error problem. He uses indirect least squares to instrument for capital. Applying this method to five Chilean manufacturing industries, he finds that, although the relative sizes of the input coefficients change, returns to scale are not affected. Unfortunately, Tybout's solution is not feasible in the nonparametric, mathematical programming framework we use here. However, by retaining the middle 80 percent of observations, we attempt to reduce the influence of outliers without introducing bias. This approach may affect parameter estimates and the level of efficiency but should not influence temporal patterns, the main focus of our investigation. To test this observation, we also perform our analysis using all observations and find that each dependent variable maintained the same (direct or inverse) relationship with efficiency as is observed with the filtered data (see section VI). 316 rHE WORLD BANK ECONOMIC RI.i.\Ni', VOL. 14, NO. 2 V. METHODOLOGY To measure the level of technical efficiency in Peruvian industries, we use the linear programming method of data envelopment analysis. This method com- pares an entity's observed level of performance with its theoretically possible level of performance. This best-practice level of performance is determined by creating a production frontier based on the firms that produce the largest amount of output(s) for a given level of input(s) or, conversely, those that minimize the amount of input(s) needed to produce given levels of output(s). Efficiency Measurement The production technology, S, is defined as all feasible combinations of inputs and outputs, feasible meaning that the combination of inputs is able to produce the levels of outputs. For each point in time t = 1, . . ., T, there are n = 1, . . ., N firms, each consuming j = 1, . . ., J inputs to produce k = 1, . . ., K outputs. Thus xj,, is the amount of input j used by firm n in period t, and Yknt is the level of output k produced by firm n in period t. All input and output observations are positive. Assuming a contemporaneous production set, for each time period t, input and output observations from only that time period are used; separate production frontiers are calculated for each industry. As an example, assume that there is only one time period. In a one-input, one- output activity, S may be illustrated as in figure 1. Efficiency measures are calcu- lated as the distance, X, from each point to the efficiency frontier. An output- based distance function, OD, is defined as:12 Figure 1. Production Technology in Input-Output Space Production Output fiontie f B A a 0 Input Note: Letters in input-output space represent firms with different input-output combinations. 12. Similarly, an input-based distance function, ID, is written as: ID(x, y) = maxtX I (x/X, y) e S}. Note that under constant returns to scale the values obtained from the output-oriented and input-oriented approaches are simply reciprocals. Alam and Morrison 317 (1) OD(x,y) = min{X I (x, yAl) E SI. Holding the input vector x constant, this expression expands the output vector y as much as possible without exceeding the boundaries of S. If a firm is output- efficient, it has a value of 1 for this expression, whereas if it is output-inefficient, the value is less than 1. Data Envelopment Analysis Charnes, Cooper, and Rhodes (1978) first introduced data envelopment analy- sis to the economics literature; it has since found multiple applications. One reason that these studies have proliferated is that linear programming methods, in general, do not require price information. This is an empirical advantage since often the only data available are physical units of inputs and outputs. It also has widespread appeal because it requires neither the assumption of cost minimization or profit maximization nor the specification of a production func- tion. Since data envelopment analysis is nonparametric, it does not confound the effects of inefficiency with misspecification of the functional form, a signifi- cant problem of parametric production function approaches.13 Furthermore, it is able to compute the relative efficiency of each firm under study, which may have multiple inputs and outputs, with any software program that has linear programming capabilities. Data envelopment analysis, as its name suggests, envelopes observed pro- duction points. It creates a flexible piecewise linear approximation to model the best-practice reference technology. It is flexible in that constraints can be placed on the linear program to account for constant, decreasing, increasing, or variable returns to scale. Radial measures of levels of technical efficiency can then be developed for firms that operate inside the convex hull of the data. We obtain the efficiency score in outputs or, equivalently, the value of the output distance function for an observation of input(s) and output(s) for firm m at time t, (xmt, Ymt), from the following linear programming model: (2) [OD(Xmt, Ymt)] = Max Xmt subject to ;,mtYkmt < WnYkntI k K, n Wnxint < j =Xt n W, 20, n =l ., N 13. Data envelopment analysis does, however, have its own weakness: because it is nonstochastic, it cannot distinguish between noise and technical inefficiency, like the stochastic frontiers methodology (Lovell 1993). 318 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 where the w,s are intensity weights that allow the comparison of convex combi- nations of data points with firm m observed in period t. The assumption of a convex polyhedral cone made here implies constant returns to scale.'4 In the simple hypothetical one-period, one-input, one-output, six-firm example, this process creates a production technology frontier (figure 1). Firms b and e define the frontier. They are efficient and have scores of 1. The other firms, which lie beneath the frontier, are inefficient and have scores less than 1. The farther away from the frontier a firm lies, the less efficient it is relative to the best- practice firms and the lower is its technical efficiency score. Firm a's output- based score, for example, is the vertical distance between it and the frontier, which is given by the ratio (OAIOB) < 1. If this number equals 0.4, firm a is only 40 percent efficient relative to the best-practice frontier. Holding the input con- stant, firm a's output could be increased 150 percent. An efficient firm a would produce at a point on the frontier that is a linear combination of processes used by firms b and e. VI. RESULTS OF THE DATA ENVELOPMENT ANALYSIS We carry out the linear programming operation (equation 2) using the total value of production as the output (k = 1) and electricity, raw materials, blue- and white-collar workers, payments for industrial services, and capital stock as in- puts (j = 6). A total of 6,473 linear programming problems are solved. Comparison of Efficiency Levels The industries with the lowest unweighted mean efficiencies were electric ac- cessories and instruments for domestic use (SIc 383) in 1988 and 1989, food products excluding beverages (sIc 311) in 1990 and 1991, and beverages and tobacco (sic 313) in 1992 (table 2). Electric accessories and instruments for do- mestic use reported the lowest unweighted mean score in 1988-the industry was only 39 percent efficient. The lowest unweighted mean score grew to 54 percent in 1989, 57 percent in 1990, and 70 percent in 1991, before falling to 68 percent in 1992. Out of 20 industries, 15 registered increases in their unweighted mean efficiency scores between 1988 and 1992. The weighted mean efficiency scores put more weight on scores from plants with higher output levels. In 1988-90 the sectors that had the lowest weighted mean efficiency scores were the same as those that had the lowest unweighted scores. In 1991, however, plastic products (sic 356) had the lowest weighted mean efficiency score, and in 1992 food products excluding beverages (sic 311) had the lowest score. The lowest weighted mean efficiency score was 43 percent in 1988, 62 percent in 1989, 75 percent in 1990, 77 percent in 1991, and 81 percent in 14. Various restrictions on the sum of the w,s result in nonincreasing (E w. < 1), nondecreasing (Y.w 2 1), or variable = 1) returns to scale. For further details see Seiford and Thrall (1990) and Fare, Grosskopf, and Lovell (198'). We assume constant returns to scale in this study, since this restriction was tested on this data set and not rejected. Table 2. Unweighted and Output-Weighted Mean Efficiency Scores by Standard Industrial Classification, Peru, 1988-92 1988 1989 1990 1991 1992 Number Un- Output- Number Un- Output- Number Un- Output- Number Un- Output- Number Un- Output- SIC of weighted weighted of weighted weighted of weighted weighted of weighted weighted of weighted weighted code plants mean mean plants mean mean plants mean mean plants mean mean plants mean mean 311 235 0.59 0.72 175 0.75 0.86 171 0.57 0.75 174 0.70 0.78 158 0.70 0.81 312 52 0.84 0.92 34 0.82 0.87 31 0.84 0.94 34 0.83 0.94 32 0.78 0.84 313 80 0.72 0.80 66 0.72 0.90 64 0.58 0.84 63 0.71 0.89 65 0.68 0.89 321 214 0.72 0.78 166 0.75 0.80 144 0.71 0.79 141 0.80 0.84 131 0.78 0.83 322 135 0.74 0.89 79 0.84 0.92 68 0.80 0.91 65 0.83 0.92 57 0.80 0.91 323 53 0.89 0.92 39 0.87 0.94 35 0.92 0.92 32 0.90 0.93 31 0.87 0.88 331 41 0.86 0.96 27 0.86 0.95 14 0.87 0.95 30 0.81 0.87 30 0.88 0.90 341 46 0.81 0.85 34 0.89 0.98 33 0.89 0.96 30 0.91 0.90 24 0.97 0.97 342 94 0.79 0.83 49 0.86 0.88 50 0.83 0.92 49 0.86 0.93 41 0.83 0.89 351 81 0.67 0.70 62 0.79 0.81 63 0.78 0.89 57 0.82 0.85 54 0.85 0.95 352 158 0.69 0.78 114 0.69 0.76 114 0.75 0.78 106 0.71 0.79 88 0.79 0.86 355 26 0.80 0.98 16 0.94 0.99 17 0.95 0.98 18 0.92 0.99 16 0.89 0.99 356 129 0.69 0.82 87 0.82 0.88 81 0.69 0.76 83 0.82 0.77 73 0.89 0.92 369 58 0.78 0.83 36 0.71 0.87 36 0.77 0.93 37 0.78 0.89 32 0.78 0.85 371 22 0.96 0.99 17 0.90 0.99 15 0.95 0.97 14 0.96 0.97 15 0.89 0.95 372 16 0.94 0.91 17 0.80 0.84 15 0.95 0.99 15 0.86 0.99 16 0.95 0.99 381 157 0.66 0.79 100 0.74 0.89 90 0.67 0.80 85 0.74 0.82 80 0.76 0.86 382 80 0.64 0.76 45 0.80 0.91 44 0.81 0.93 46 0.76 0.92 34 0.89 0.97 383 88 0.39 0.43 69 0.54 0.63 55 0.72 0.85 56 0.79 0.90 55 0.72 0.87 390 46 0.84 0.94 33 0.87 0.94 34 0.87 0.96 29 0.87 0.96 27 0.86 0.94 Source: Authors' calculations based on data from Peru's Ministry of Industry, Tourism, Integration, and International Trade Negotiations. 320 THE WORiLD BANK ECONOMIC REVIEW, VOl.. 14, NO. 2 1992. As with the unweighted mean efficiency scores, there was a marked trend of increasing efficiency over time. Out of the 20 industries, 16 showed an improve- ment in output-weighted mean efficiency between 1988 and 1992. In almost all cases-the sole exception being nonferrous metals (SIC 372) in 1988-the output-weighted means were higher than the unweighted means for 1988 and 1992. In general, the more efficient firms were producing more of the output and, hence, the output-weighted means were higher than simple averages. Trade liberalization should reduce the dispersion of efficiency scores within a given sector and-to the extent that tariff rates in different sectors converge- among sectors. First, consider the possibility of efficiency scores converging among sectors. The difference between the highest and lowest unweighted mean scores declined substantially between 1988 and 1992. This was also the case for weighted mean scores. In fact, the decline in the spread between highest and lowest scores was identical for weighted and unweighted means: from 57 to 29 percent be- tween 1988 and 1992.15 Second, consider the issue of convergence within sec- tors. The standard deviation of the unweighted efficiency scores declined in 14 of 20 industries (table 3).16 Determinants of Mean Efficiency Scores In addition to comparing efficiency levels among industries and over time, we are also interested in the determinants of these efficiency scores. For our analysis we use mean efficiency scores, as opposed to individual-plant efficiency scores, because the data set does not contain plant-level data other than inputs and out- puts.17 Thus we are limited to an aggregate analysis of the determinants of sectoral efficiency. We explore two possible determinants of efficiency at the three-digit sic level: commercial policy and industrial structure. Commercial policy is mea- sured by rates of effective protection; these data are produced by the Peruvian Central Bank. Industrial structure is measured by the Herfindahl index of indus- trial concentration, calculated from the same data set used to produce the data envelopment analysis efficiency scores. Before presenting our econometric results, a word of caution is in order. Other important macroeconomic events in Peru between 1988 and 1992-exchange rate overvaluation, hyperinflation, changes in the real interest rate-could over- whelm the effects of changes in trade policy and market structure. In addition, our data have potential econometric problems: price indexes may be biased over time because of hyperinflation, and some variables, especially capital stock, may be measurcd with error. These econometric problems further complicate the at- 15. Page (1980), for example, has a more detailed data source and is able to use experience of entrepreneurs, age of the plant, and education level of the plant's labor force as regressors. 16. This decline was due almost entirely to the rise in the mean of the least-efficient industry, since the highest mean score was already close to 100 percent. 17. We do not report standard deviations calculated on the basis of output-weighted efficiency scores in table 3, because shifts in output to more efficient firms may increase the standard deviation rather than lower it. In other words, because of the weighting procedure, the standard deviation now reflects two factors: the standard deviation per se and the distribution of output. Only the former is of interest as a test of convergence. Table 3. Standard Deviations of Unweighted Efficiency Scores by Standard Industrial Classification, Peru, 1988-92 1988 1989 1990 1991 1992 Number Number Number Number Number SIC of Standard of Standard of Standard of Standard of Standard code plants deviation plants deviation plants deviation plants deviation plants deviation 311 235 0.23 17S 0.20 171 0.27 174 0.22 158 0.21 312 52 0.18 34 0.20 31 0.17 34 0.20 32 0.21 313 80 0.27 66 0.25 64 0.29 63 0.24 65 0.25 321 214 0.21 166 0.20 144 0.22 141 0.17 131 0.19 322 135 0.22 79 0.18 68 0.23 65 0.19 57 0.21 323 53 0.14 39 0.18 35 0.14 32 0.15 31 0.14 331 41 0.17 27 0.17 14 0.15 30 0.21 30 0.18 341 46 0.19 34 0.16 33 0.17 30 0.12 24 0.05 342 94 0.22 49 0.18 50 0.19 49 0.16 41 0.17 351 81 0.28 62 0.23 63 0.22 57 0.20 54 0.21 352 158 0.23 114 0.23 114 0.22 106 0.22 88 0.21 355 26 0.22 16 0.15 17 0.11 18 0.14 16 0.18 356 129 0.23 87 0.18 81 0.21 83 0.18 73 0.14 369 58 0.22 36 0.26 36 0.24 37 0.20 32 0.23 371 22 0.07 17 0.16 15 0.11 14 0.07 15 0.15 372 16 0.10 17 0.23 15 0.08 15 0.18 16 0.09 381 157 0.26 100 0.22 90 0.23 85 0.20 80 0.20 382 80 0.25 45 0.21 44 0.22 46 0.24 34 0.19 383 88 0.25 69 0.28 55 0.27 56 0.19 55 0.24 390 46 0.19 33 0.20 34 0.17 29 0.18 27 0.18 Source: Authors' calculations based on data from Peru's Ministry of Industry, Tourism, Integration, and International Trade Negotiations. 322 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 tempt to uncover the links among trade policy, market structure, and industrial efficiency. Despite these observations, however, it should be noted that, although all industrial sectors faced the same macroeconomic events, they were subject to different changes in trade policy and market structure. Thus although it is diffi- cult to identify macroeconomic determinants of productivity change, since trade policy and industrial concentration are more sector-specific, their influence on productivity is easier to capture."8 Industries with high effective rates of protection before stabilization still had high rates after stabilization (p = 0.82; table 4). This pattern also holds for industrial concentration (p = 0.83) and mean efficiency levels (p = 0.80). Both contemporane- ous and lagged Herfindahl indexes are significantly positively correlated with mean efficiency levels (0.50 < p < 0.66). Correlations between Herfindahl scores and effec- tive rates of protection are negative but generally insignificant; the same is true for correlations between effective rates of protection and mean efficiency levels. Correla- tions between effective rates of protection and standard deviations are positive (as expected) for the unweighted case and negative for the standard deviations calcu- lated on the basis of output-weighted means (see note 16). To test whether there is a relationship among industrial concentration, effec- tive rates of protection, and efficiency at the industry level, we perform several regressions. We hypothesize that higher levels of industrial concentration and effective rates of protection are associated with lower efficiency scores. The logic behind the first inverse relationship is that the smaller is the Herfindahl index, the less concentrated and more competitive is the industry, which should lead to greater efficiency. The reasoning behind the second inverse relationship is that, as effective rates of protection decline with trade liberalization, increased compe- tition compels firms to become more efficient. We also include a third covariate: the square root of the number of plants in each sector, \,,V. Caves and Barton (1990) show a link between the number of firms used to estimate technical efficiency and the resulting efficiency score. The expected relationship is an inverse one, which at first appears counterintuitive, since one might expect that, as the number of firms increases, competition in- creases, driving average efficiency upward. What is actually occurring, however, is a purely statistical phenomenon: as the number of observations drawn from a distribution increases, so do the number of extreme values (both high and low). Caves and Barton (1990:60) argue that the relationship between estimated tech- nical efficiency and the number of observations may be similarly linked. In other words, the more draws taken from a distribution, the more likely the researcher is to encounter a highly efficient plant, which makes all other plants in the sample less efficient in comparison. Caves and Barton observe that the range of effi- ciency values increases at a rate close to v'N. Thus we estimate our equations with and without this regressor. 18. Tybout, de Melo, and Corbo (1991) offer this justification for examining the effect of trade policy on industrial efficiency in Chile between 1967 and 1979. Table 4. Pearson Correlations between Mean Efficiency Levels, Effective Rates of Protection, and Herfindabl Indexes of Industrial Concentration, Peru, 1990-92 OUTPUT OUTPUT OUTPUT OUTPUT MEAN MEAN STD STD WGTD WGTD WGTD WGTD STD ERP90 ERP92 HERF90 HERF92 EFF 90 EFF 92 DEV 90 DEV 92 EFF 90 EFF 92 STD DEV 90 DEV 92 ERP90 1 ERP91 0.82*** 1 (<0.01) HERF90 -0.29 -0.18 1 (0.21) (0.45) HERF92 -0.42* -0.25 0.83*** 1 (0.06) (0.28) (<0.01) MEAN -0.28 -0.05 0.66"** 0.63*** 1 EFF 90 (0.24) (0.84) (<0.01) (<0.01) MEAN -0.43* -0.16 0.51 * * 0.50** 0.80*** 1 EFF 92 (0.06) (0.50) (0.02) (0.03) (<0.01) STD 0.38* 0.15 -0.73*** -0.70*** -0.91*** -0.80*** 1 DEV 90 (0.10) (0.54) (<0.01) (<0.01) (<0.01) (<0.01) STD 0.45** 0.20 -0.34 -0.39* -0.58*** -0.85"' 0.69*** 1 DEV 92 (0.05) (0.40) (0.14) (0.09) (<0.01) (<0.01) (<0.01) OUTPUT -0.16 0.07 0.72*** 0.67*** 0.87*** 0.63*** -0.71*** -0.39* 1 WGTD EFF 90 (0.49) (0.78) (<0.01) (<0.01) (<0.01) (<0.01) (<0.01) (0.09) OUTPUT -0.30 -0.01 0.61*** 0.64*** 0.65*** 0.80*** -0.60*** -0.56*** 0.64*** 1 WGTD EFF 92 (0.20) (0.96) (<0.01) (<0.01) (<0.01) <0.01) (0.01) (0.01) (<0.01) OUTPUT -0.30 -0.16 0.98*** 0.83*** 0.71*** 0.56* -0.77*** -0.38* 0.76*** 0.61*** 1 WGTD STD DEV 90 (0.20) (0.50) (<0.01) (<0.01) (<0.01) (0.01) (<0.01) (0.10) (<0.01) (<0.01) OUTPUT -0.41* -0.20 0.85*** 0.98*** 0.71*** 0.56*** -0.76*** -0.42* 0.73*** 0.69*** 0.86*** 1 WGTD STD DEV 92 (0.07) (0.41) (<0.01) (<0.01) (<0.01) (0.01) (<0.01) (0.06) (<0.01) (<0.01) (<0.01) *Significant at the 10 percent level. "Significant at the 5 percent level. * *Significant at the 1 percent level. Note: p-values are in parentheses. ERP90 is the effective rate of protection in July 1990; ERP91 is the effective rate of protection in March 1991. HERF90 is the Herfindahl index in 1990; HERF92 is the Herfindahl index in 1992. MEAN EFF 90 is the mean efficiency level in 1990; MEAN EFF 92 is the mean efficiency level in 1992. STD DEV 90 is the standard deviation of efficiency in 1990; STD DEV 92 is the standard deviation of efficiency in 1992. OUTPUT WGTD EFF 90 is the output- weighted mean efficiency in 1990; OUTPUT WGTD EFF 92 is the output-weighted mean efficiency in 1992. OUTPUT WGTD STD DEV 90 is the standard deviation of output-weighted efficiency in 1990; OUTPUT WGTD STD DEV 92 is the standard deviation of output-weighted efficiency in 1992. Source: Effective rates of protection come from unpublished data from the Peruvian Central Bank. Herfindahl indexes and mean efficiency levels are computed by the authors using data from Peru's Ministry of Industry, Tourism, Integration, and International Trade Negotiations. 324 THE WORLD RANK ECONOMIC REVIEW, VOL. 14, NO. 2 We are able to calculate efficiency scores for plants in 25 sectors, but three-wood and cork products (sic 332), glass and glass products (sIc 362), and the manufacture of professional, scientific, and measurement and control equipment including photo- graphic equipment and optics (sic 385)-do not have data on effective rates of pro- tection. We dropped two other industries-footwear (SIC 324) and transport equip- ment (sic 384)-because we were unable to obtain deflators for these classifications. Thus there are 20 observations for each year. Also the earliest year for which we could obtain effective rates of protection is 1990; therefore our regressions cannot go back to 1988 or 1989. We also do not have effective rates of protection for 1992, so we must use 1991 as the postreform year. Clearly, this will tend to understate the efficiency gains resulting from liberalization, since firms will not have adjusted com- pletely to the change in regime by the end of 1991.19 REGRESSION RESULTS. In the absence of the ,WT regressor, the results of the regression using 1990 data (table 5, column 1) provide no support for our hypothesis. Although the coefficient on trade protection is negative (indicating that industries with higher levels of protection have lower mean efficiency scores), it is not statistically significant. The coefficient on industrial concentration, as measured by the Herfindahl index, is statistically significant, but its sign is the opposite of what we would expect: more concentrated industries have higher efficiency scores. However, once we control for the statistical factor identified by Caves and Barton, the results do support our hypothesis (table 5, column 2). All three coef- ficients have the expected negative sign, the effective rate of protection is signifi- cant at the 10 percent level, and ,, is significant at less than 1 percent. The coefficient on the Herfindahl index, although not significant at standard levels, is suggestive of an inverse relationship. Note that V[ and the Herfindahl index are significantly, inversely related (p = -0.61 and -0.66 for 1990 and 1991, respec- tively). This high degree of collinearity makes it difficult to isolate the link be- tween market concentration and efficiency and, hence, may explain why the Herfindahl coefficient has a higher p-value. The results for 1991 tell a similar, although less striking, story. The signs on the regressors are negative, but only N'Th is statistically significant (table S, col- umn 4). Again, if \T is not included (table 5, column 3), the effective rate of protection is insignificant, and the Herfindahl index, although significant, has the wrong sign.20 19. In terms of the sign of each parameter estimate, the results are the same whether we use the complete data set or the 80 percent subset. However, with the complete data set the parameter estimates, although similar in magnitude, are typically lower, and most are not significant. By using the 80 percent subset, we thus are able to measure more precisely the relationship between trade reform and efficiency, which is obscured when outliers are present. 20. An alternative interpretation of the Herfindahl index is as a measure of the underlying differences in efficiency within the industry: efficient plants get larger, resulting in higher concentration. This is a possible explanation for its positive sign. Note also that the Herfindahl index can be written as the sum of a "variance equivalent" (or dispersion component that measures the within-industry variance in size) and the inverse of a "numbers equivalent" (or numbers component that equals 1 over the number of firms). Alam and Morrison 325 Table S. Determinants of Mean Efficiency at the Three-Digit Level Independent 1990 data 1991 data Pooled data Disequilibriuma variable (1) (2) (3) (4) (5) (6) (7) (8) Constant 0.022 0.128** -0.062 0.018 -0.021 0.070 -0.020 0.047 (0.77) (0.04) (0.35) (0.74) (0.70) (0.11) (0.77) (0.48) Effective rate of protectionb -0.030 -0.051* -0.003 -0.107 -0.034 -0.056** -0.101 -0.189* (0.46) (0.09) (0.98) (0.24) (0.29) (0.03) (0.40) (0.10) In Herfindahl index 0.104** -0.077 0.060** -0.043 0.083**t -0.054* 0.061** -0.027 (< 0.01) (0.12) (0.01) (0.21) (<0.01) (0.07) (0.02) (0.52) -0.068*** _0.040** * -0.052*** -0.034** (< 0.01) (< 0.01) (0.01) (0.02) Year dummyd 0.023* -0.008* (0.04) (0.03) R2 0.45 0.74 0.31 0.62 0.39 0.67 0.30 0.50 F 6.97*** 15.53** 3.74** 8.60*** 7.97*** 17.99*** 3.70** 5.31* (< 0.01) (< 0.01) (0.05) (< 0.01) (< 0.01) (< 0.01) (0.05) (0.01) Number of observations 20 20 20 20 40 40 20 20 *Significant at the 10 percent level. * Significant at the 5 percent level. * *Significant at the 1 percent level. Note: p-values are in parentheses. Dependent variable is the natural log of the mean efficiency score at the three-digit SIC level. a. Data from 1992 are regressed on data from 1991. b. Estimated coefficient times 100. c. N is the number of plants in each sic (see tables 2 and 3). d. Year dummy = 0 for 1990; 1 for 1991. Source: Authors' calculations. In order to determine if the parameter estimates are insignificant because of the high degree of collinearity, we run a regression pooling the data for both 1990 and 1991 (table 5, columns 5 and 6). As expected, with twice as many observations, all of the coefficients in the pooled regression can be measured more precisely and become significant. They also have the expected sign (when Tis omitted, the point estimates are significant, but, again, Herfindahl has the incorrect sign). These results suggest that the expected relationship among mar- ket structure, trade reform, and efficiency does exist, but it is hard to detect unless there is a sufficient number of observations so that ordinary least squares can separate out the effects of individual regressors on efficiency. The fact that the econometric results are stronger for 1990 than for 1991 should not be surprising. Almost 30 years of import substitution industrializa- tion produced a strong negative relationship between rates of effective protection and efficiency levels. In the year spanning 1990 and 1991, firms may not have adjusted completely to the new tariff structure. There are several reasons for such a delayed response. The first reason is the credibility of the reform-past govern- This would explain why, when we include both N and the Herfindahl index, the coefficient on