WPS6295 Policy Research Working Paper 6295 Are the Benefits of Export Support Durable? Evidence from Tunisia Olivier Cadot Ana M. Fernandes Julien Gourdon Aaditya Mattoo The World Bank Development Research Group Trade and Integration Team December 2012 Policy Research Working Paper 6295 Abstract This paper evaluates the effects of the FAMEX export levels of beneficiaries are not significantly different from promotion program in Tunisia on the performance of those of non-beneficiary firms. Exports of beneficiaries beneficiary firms. While much of the literature assesses do remain more diversified, but the diversification does only the short-term impact of such programs, the paper not translate into lower volatility of exports. The authors considers also the longer-term impact. Propensity-score also did not find evidence that the program produced matching, difference-in-difference, and weighted least spillover benefits for non-beneficiary firms. However, the squares estimates suggest that beneficiaries initially see results on the longer-term impact of export promotion faster export growth and greater diversification across must be interpreted cautiously because the later years of destination markets and products. However, three years the sample period saw a collapse in world trade, which after the intervention, the growth rates and the export may not have affected all firms equally. This paper is a product of the Trade and Integration Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at olivier.cadot@unil.ch, afernandes@worldbank.org, julien.gourdon@cepii.fr and amattoo@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Are the Benefits of Export Support Durable? Evidence from Tunisia # Olivier Cadot* Ana M. Fernandes+ Julien Gourdon§ Aaditya Mattoo++ December 2012 Abstract This paper evaluates the effects of the FAMEX export promotion program in Tunisia on the performance of beneficiary firms. While much of the literature assesses only the short-term impact of such programs, the paper considers also the longer-term impact. Propensity-score matching, difference-in-difference, and weighted least squares estimates suggest that beneficiaries initially see faster export growth and greater diversification across destination markets and products. However, three years after the intervention, the growth rates and the export levels of beneficiaries are not significantly different from those of non- beneficiary firms. Exports of beneficiaries do remain more diversified, but the diversification does not translate into lower volatility of exports. We also did not find evidence that the program produced spillover benefits for non-beneficiary firms. However, the results on the longer-term impact of export promotion must be interpreted cautiously because the later years of the sample period saw a collapse in world trade, which may not have affected all firms equally. JEL classification codes: F13, F14, L15, L25, 017, 024, C23. Keywords: Export promotion, firms, export margins, Tunisia, impact evaluation, propensity- score matching, matching grant. # We are grateful to Hamid Alavi, Daniel Lederman, Jean-Michel Marchat, Mine Senses, and Tani Fukui for useful comments and discussions, to Yasusuki Todo for sharing his matching methodology and to participants at CEPII, the 2011 Conference on Trade and Firms in Aix en Provence, the 2011 World Bank-IGC workshop on ―Trade policy for growth‖, the 2012 International Industrial Organization Conference, the Par is School of Economics, the 2011 Southern Economic Association Meetings, the University of Lausanne, the University of Illinois at Urbana- Champaign, and the World Bank workshop of the Chief Economist of the Middle East and North Africa Region for comments. This study was supported by the governments of Norway, Sweden and the United Kingdom through the Multi-Donor Trust Fund for Trade and Development, and by Switzerland‘s SNF through NCCR work package 6 on impact assessment. The findings expressed in this paper are those of the authors and do not necessarily represent the views of the World Bank. * University of Lausanne, CEPREMAP, CEPR and FERDI. + Trade and Integration Unit, Development Economics Research Group, World Bank. § CEPII. ++ Trade and Integration Unit, Development Economics Research Group, World Bank. 1 1. Introduction Since trade liberalization per se has not always led to improved export performance, the focus of trade policy has shifted in recent years towards targeted interventions to facilitate trade and especially to promote exports. Significant resources are devoted to export-processing zones, exporter assistance programs, and projects aimed at modernizing border management and customs procedures. The paucity of rigorous evaluations of these interventions, however, has made it hard to assess their desirability and to improve their design. This paper provides new evidence on the firm-level impact of a recent export promotion program in Tunisia. It seeks, in particular, to fill an important gap in the literature: whereas evaluations have typically focused on the contemporaneous or short-term effect of interventions, we know relatively little about their longer-term impact.1 The literature assessing the effectiveness of export promotion has developed along two strands. The older strand relies on cross-country evidence and examines effects on aggregate export performance. Thus, Rose (2007) uses a gravity equation to show that diplomatic representations had a positive effect on bilateral trade flows. Lederman, Olarreaga, and Payton (2010) show that export promotion activities, after a long history of failure, in particular in developing countries where they coexisted with import substitution policies and currency overvaluation, have recently had more success in increasing aggregate exports, particularly when management of the activities involved the private sector. A more recent strand of the literature has assessed export promotion using quasi-experimental methods, comparing the export performance of treated firms with that of a control group. Since enrollment into export promotion programs is never random, most papers go to great lengths to control for selection effects through matching, fixed effects, and two-step (IV or Heckman) estimation methods. The first broad finding is that export promotion seems to be more successful at affecting the performance of established exporters than at encouraging non-exporting firms to start exporting (Bernard and Jensen, 2004; Görg, Henry and Strobl, 2008; Girma, Gong, Görg and Yu, 2009). This is in accordance with the literature on heterogeneous firms and trade, which suggests that there are differences between exporters and non-exporters in terms of productivity and a host of other firm characteristics (see, e.g. Bernard, Jensen, Redding and Schott, 2007), which export promotion activities may not be able to offset. The second broad finding is that for exporters, the impact is stronger along the extensive margin than along the intensive one 1 This gap can lead to what Ravallion (2008) has called ―myopia bias‖, whereby evaluation focused on short-term effects may tilt incentives toward development projects that yield quick results. 2 (Alvarez and Crespi, 2000, Volpe and Carballo, 2008).2 Thus assistance may be more successful in helping firms overcome hurdles to break into new markets (product- or destination-wise) than in ramping up export volumes.3 These papers break new ground in terms of rigorous evaluation of trade interventions, but focus primarily on the short-term effects of interventions. The present paper examines both the short-term and the longer-term impact of Tunisia‘s export promotion program, FAMEX, which consists of matching grants provided to Tunisian firms to implement export business plans. We combine several sources of firm-level data—FAMEX program data, National Statistical Institute and Investment Promotion Agency data, and customs transaction data—into a unique, rich dataset on Tunisian exporters. In particular, merging customs data with other sources eliminates the risk of recall bias in outcome variables, which tends to arise when public programs are evaluated ex-post using surveys.4 We estimate FAMEX‘s treatment effects using a menu of estimation methods, including difference-in-differences combined with propensity-score matching (PSM-DID) and difference- in-differences weighted by propensity scores, designated henceforth as weighted least squares regressions (WLS). Our rich dataset allows us to extend the analysis in several directions, including the sustainability of the program‘s effect. We find that, compared to a control group, FAMEX beneficiaries successfully diversify in terms of export destination markets and products, and durably so. However, the beneficiary firms‘ total exports diverges only temporarily from the control group‘s total exports. One year after treatment, the differential in growth rates of total exports is not significant anymore. Three years after treatment, even export levels are no longer significantly different. Even though export destination and product counts remain significantly different throughout the sample period, the treatment group‘s diversification does not seem to translate into reduced export volatility. We also examine the existence of program spillovers by estimating FAMEX‘s indirect impact on the performance of control firms. This is an important—although typically underexplored—part of program impact evaluation, because in the presence of spillovers, the absence of a positive measured treatment effect could reflect a positive true treatment effect transmitted to the control group through positive externalities, which is precisely the combination that would justify 2 Girma, Gong, Görg and Yu (2009) find a positive impact along the intensive margin but they consider the special case of production subsidies. 3 See Rangan and Lawrence (1999) and references therein on the hurdles facing the internationalization of firms. Assistance may have stronger effects for small firms, perhaps because they face relatively greater hurdles, as Volpe and Carballo (2010) find in the case of an export promotion program in Chile. 4 In the case of FAMEX, the World Bank collected firm-level survey data to analyze the impact of the program and the corresponding analysis is described in Gourdon, Marchat, Sharma, and Vishwanath (2011). 3 government intervention.5 Catching-up by control firms due to gradual spillovers could also explain why we find only temporary treatment effects on total exports, although one would also expect imitation in terms of diversification. Under the hypothesis that any spillover effects are stronger for control firms in the same industry and geographical area as treated firms, we find that such effects are generally not significant, and in some cases are negative. The latter adverse effect could reflect, for instance, poaching of good managers and workers by treated firms. Finally, we study heterogeneity in treatment effects as a function of the beneficiaries‘ objectives and use of assistance. We show that a firm‘s objective when requesting FAMEX assistance matters for performance: firms seeking to expand into new markets or to develop new export products benefit more than firms seeking to become more substantive exporters. Also, market prospection and promotion activities correlate more significantly with export outcomes than other components of FAMEX, like firm or product development, suggesting that informational barriers are the most amenable to effective government assistance. These findings support the broad view that firms seeking and using assistance to expand along the extensive margin are less likely to be disappointed with the longer-term outcome than those seeking and using assistance to expand along the intensive margin. The paper is organized as follows. Section 2 describes the export promotion program and Section 3 presents the data. Section 4 discusses estimation issues. Section 5 presents the FAMEX treatment effects and robustness checks. Section 6 focuses extensions, including the diversification and volatility of exports, the presence of externalities, whether firms‘ objectives and use of assistance matter, and estimates of the economic magnitude of the FAMEX effects. Section 7 concludes. 2. Export Promotion in Tunisia The Tunisian government has worked since the mid-1990s to reduce the traditional anti-export bias of Tunisia‘s trade policy (World Bank, 2008). Apart from the elimination of tariffs on imported raw materials, equipment and capital goods in a number of sectors, it expanded its use of export promotion tools. The World Bank supported these efforts through a loan for the Export Development Project (EDP) implemented in two phases: 2000-2004 for the first phase and 2005- 2009 for the second phase. The Export Promotion Centre (CEPEX in its French acronym) was 5 The fact that potential exporters are not fully informed about foreign market opportunities is not sufficient, in itself, to create a market failure if information production is costly but appropriable. By contrast, a market failure could arise in the presence of imperfect appropriability of the information. Indeed, Volpe and Carballo (2008), citing McDermott (1994), note that ―customer lists are the most common target of corporate spies.‖ 4 the key agency under the Ministry of Trade responsible for implementing Tunisia‘s export promotion activities. Our analysis focuses on FAMEX, a major demand-driven program under the second phase of the EDP project—whose aim was to help Tunisian firms overcome barriers to sell in foreign markets and enhance their competitiveness.6 The program‘s rationale was that Tunisian firms were poorly informed about export markets and had difficulty identifying the right target markets, product segments, and sales channels. The program provided firms with matching grants co-financing 50 percent of the cost of export business plans. In terms of firm size, the minimum annual turnover required for FAMEX eligibility was 200,000 Tunisian Dinars (144,000 USD) in manufacturing and 100,000 Tunisian Dinars (71,000 USD) in service and craft sectors.7 In terms of age, only firms that had been in operation for a minimum of two years were eligible for FAMEX, but there were a few exceptions. A firm approaching FAMEX for assistance had to submit an export business plan focused on one of three possible objectives: (i) become a substantive exporter (if the firm had little or no export experience), (ii) to diversify its destination markets, or (iii) to develop new export products.8 While a single main objective had to be provided for each export business plan, firms could request assistance also for other objectives. The export business plan was evaluated by a panel of five local and international experts and, if accepted, the FAMEX team would provide technical assistance—outsourced to specialized consulting firms—to help the firm improve its plan. The panel would draw up, together with the firm, a list of activities eligible for matching grants of up to 50 percent of their cost, with a ceiling of 100,000 Tunisian Dinars (71,000 USD). FAMEX received 1,710 applications and accepted 1,231 from 1,060 firms between 2005 and 2009.9 In terms of the primary objective to request FAMEX assistance, 31 percent had little or no 6 The FAMEX program also helped to build the institutional capacity of local professional organizations (export associations, chambers of commerce, and professional consulting organizations) and to strengthen the export consulting sector in Tunisia. Another component of the second phase of the EDP project focused on trade facilitation, including investments and technical assistance to modernize customs procedures, through a combination of investments in hardware and software and procedural improvements. These components—if effective—are likely to have benefited Tunisian firms broadly and thus do not necessarily contaminate the identification of FAMEX effects. 7 Tunisian Dinars are converted to USD using the exchange rate as of October 10, 2011 (1 USD = 1.463 Tunisian Dinars). 8 While there was no clear rule on which firms were deemed to have little export experience, in interviews with the authors, FAMEX management officers indicated that they included in that category firms that either did not export or exported an amount representing less than 20 percent of their total sales in the recent past. 9 Some firms applied to FAMEX twice and had two export development plans accepted prior to 2009, some firms started a second export development plan in 2009, and some firms dropped one export development plan before re- 5 export experience while 69 percent of the beneficiaries were already exporters and wanted to diversify either by expanding into new destination markets (49 percent) or into new products (20 percent). The program‘s coverage was fairly broad in terms of sectors and locations (see Section 3). FAMEX grants were used mostly to co-finance the cost of technical assistance and marketing services provided by local and foreign experts. Five types of activities were financed: (i) market prospection, (ii) promotion, (iii) product development, (iv) firm development, and (v) foreign subsidiary creation. The amounts disbursed by FAMEX for each type of activity along with a description of the activities are shown in Table 1. In terms of actual disbursements, shares in the program total in the second column add up to 100 percent, but the number of firms in the third column adds up to more than the total number of FAMEX beneficiaries because each firm typically received co-financing to undertake several activities.10 As FAMEX was a matching grant program in which firms contributed half the costs, the program‘s management team expected firms‘ incentives to be aligned with the program, so that funds were unlikely to be misallocated or devoted to low-value services. The fact that FAMEX operated on a reimbursement basis, whereby firms had to present receipts upon implementing the activities in their plan, gives us reasonable confidence that the matching grant funds were used for their intended purpose.11 These features of FAMEX make it a particularly attractive program to evaluate.12 applying to FAMEX. The FAMEX management team did not systematically collect information on firms whose applications to FAMEX were turned down. 10 A different breakdown of the activities in Table 1 provided by FAMEX program data indicates that 25 percent of FAMEX funds covered marketing research costs, 18 percent covered fees from private export-marketing consultants, 15 percent covered the participation in trade fairs, 10 percent went to establishing foreign representations, 10 percent covered printing costs for advertising material, and the rest was scattered over minor items. 11 Moreover, FAMEX beneficiaries were obliged to supply the FAMEX management team with data to allow a general assessment of the project‘s impact on export growth, and supervision teams from the World Bank also had access to that information. However, as in the case of any assistance program, the impact of FAMEX on its stated activities could be reduced by fungibility. That is, $100 given to a firm for a specific activity, even through a matching grant, could still have a ―windfall effect‖ and allow the firm to re-optimize and spend less on the activity than it would have in the absence of the program. In that case, the program‘s money would (at least in part) replace money that the firm would have spent otherwise. 12 Other World Bank-funded programs, for example in education, have been shown to suffer from a misuse of funds (Reinikka and Svensson, 2004). 6 Table 1 FAMEX Program Components Amounts Share in Number Description of activities disbursed (in program total of firms millions of USD) Market prospection Acquisition of information (e.g., through the purchase of data or market studies), firm missions abroad to visit trade fairs 2.665 23.9% 313 and foreign exhibitions, and visits of prospective buyers. Promotion Production of information and marketing including design, production and publication of ads in various media (e.g., newspapers, magazines, TV, radio, web, brochures), sending 4.113 36.9% 319 of mailings and samples, and firm representation (stands) in trade fairs and exhibitions. Product development Product design modifications and production of samples, package design and modifications, and trademark 1.515 13.6% 184 registration. Firm development Training on organizational issues such as setting up a marketing watch, an export cell, or an export-oriented 1.169 10.5% 220 business plan. Foreign subsidiary creation Assistance for the establishment of a facility abroad including legal, consulting, covering rental and salary costs for the 1.688 15.1% 84 first year of establishment. Total 11.150 100.0% Note: Tunisian dinars were converted into U.S. dollars at the October 10, 2011 exchange rate (1.463 Tunisian Dinars per USD). The figures in the table concern the 455 FAMEX beneficiaries for which data was requested to FAMEX‘s management team, as described in Section 3. 3. Data and Descriptive Statistics In order to evaluate rigorously the impact of FAMEX, we need data on beneficiary firms as well as a control group. Our dataset combines three main sources: (i) FAMEX program data, (ii) data from the National Statistical Institute (INS in its French acronym) and the Investment Promotion Agency (API in its French acronym), and (iii) customs transaction data. First, we obtained from FAMEX‘s management team a complete list of the 1,060 beneficiary firms indexed by their tax ID. After dropping 120 firms that dropped out of the program, 175 firms in the services sector for which customs transaction data is not available, and 310 firms whose first export development plan was still ongoing at the end of 2009 (the penultimate year in our sample period), we were left with 455 FAMEX beneficiaries. For these firms, we obtained program data covering the following variables: years in which the firms joined and terminated the program (which in most cases lasted for one year), firm location, sector, employment and total sales when it joined the program and when it left, whether the firm had an in-house export unit prior to joining the program, its objective in applying to FAMEX and its grant use in terms of total disbursements and breakdown across activities. 7 Second, we obtained from the INS a stratified sample of control firms with a structure similar to that of the 455 FAMEX beneficiaries. The stratification was performed based on size, prior exporting status, and sector (within manufacturing), resulting in 48 cells. For each cell we asked INS to provide us with twice as many non-beneficiaries as there were FAMEX beneficiaries, i.e., in total, 910 control firms. To draw the stratified sample of control firms, INS used its 2007 census of firms which includes information on firm location, sector, date of creation, employment, and total sales, with the last two variables being defined in terms of discrete intervals.13 Since INS data was incomplete for a number of firms, we supplemented it with data obtained from the API. API‘s database for 2007 includes employment, sector, date of firm creation, social capital, and status (offshore or common law) for 5,000 firms across all sectors (of which 500 are also in the INS census). We extracted a group of 2,000 manufacturing firms from the API database that were neither in the FAMEX sample nor in the INS sample. Third, we obtained transaction-level export data from Tunisian Customs for the 455 FAMEX beneficiaries, the 910 control firms from INS, and the 2,000 control firms from API. For every year between 2000 and 2010 and every firm (identified by its tax ID) we obtained monthly export transaction values by destination country and product code, the latter using an 11-digit Tunisian nomenclature derived from the Harmonized System (HS).14 We aggregated monthly data to annual export totals for each firm and year. Combining data from all three sources, and after data consistency checks and cleaning, we obtained an unbalanced panel of yearly export activity for 2,746 exporting firms with an average of six years of data per firm over the period 2004-2010.15 Of those, 401 benefitted from FAMEX and 2,346 did not. Among the 2,346 non-beneficiaries, 71 applied to FAMEX but were turned down while 126 dropped out. In some robustness specifications, we will include the 126 dropouts in the treatment group instead of the control group while in others we will eliminate them from the sample. In the combined dataset, firm-level characteristics other than those related to export transactions are time-invariant, being available only for 2007. Our combined dataset has two important features. First, the merger of customs transaction data with other sources of firm-level data ensures that the outcome variables do not suffer from recall bias, as would be the case if the outcome variables were obtained from survey data. Second, the 13 The employment intervals are 1-9 employees, 10-19 employees, 20-49 employees, 50-99 employees, 100-199 employees, and more than 200 employees. The total sales intervals are: under 50,000 Tunisian Dinars, 50,000-1 million Tunisian Dinars, 1 million-2 million Tunisian Dinars, 2 million-5 million Tunisian Dinars. 14 The data was converted to HS 6-digit by keeping the classification‘s first six digits. 15 The merging of the three data sources was possible thanks to the use of unique tax IDs by all Tunisian administrations concerned and their willingness to share the data with us. Some of the data inconsistencies addressed were wrong sectoral classifications in the FAMEX program data which were corrected using INS and API data. 8 fact that all control firms are exporting firms (as stratification was based on prior exporting status) improves overall sample homogeneity and the identification of FAMEX effects.16 Table 2 provides descriptive statistics for FAMEX and control firms in terms of sector, location, and employment categories. The sectoral distribution of FAMEX and control firms is quite similar with the exception of the textiles & apparels sector which is more heavily represented in the control group, although it is also the treatment group‘s largest sector, accounting for 31 percent of beneficiaries. Location was also used for stratification, hence the geographical distribution of FAMEX and control firms is fairly similar, although FAMEX firms are more concentrated in Tunis. There are also only minor differences across size categories measured in terms of employment and the same is true in terms of sales (not reported). Table 2 Summary Statistics Panel A. Distribution by Sector Paper, Machine Agro- Textile & Total wood & Chemicals Metals & Electric Sector industry apparels number of furniture (%) (%) equipment (%) (%) (%) firms (%) (%) FAMEX firms 15 31 13 12 8 14 6 401 Control firms 11 43 9 11 7 11 7 2346 Panel B. Distribution by Region Total Grand Central Rest of Location Tunis number of Tunis Sea Tunisia firms FAMEX firms 22 48 28 2 401 Control firms 10 46 37 8 2346 Panel C. Distribution by Employment Category Total Employment [1,9] [10,19] [20,49] [50,99] [100,199] >=200 number of firms FAMEX firms 11 9 29 19 16 16 401 Control firms 5 12 31 23 17 12 2346 Source: Authors‘ calculations using the combined dataset. Table 3 shows export trends between 2003 and 2010 for FAMEX and control firms, as well as for Tunisian manufacturing exports as a whole. There is no prima-facie evidence of superior 16 By ―exporters‖, we mean firms having a customs code and having conducted at least one export transaction during the sample period 2004-2010. 9 performance by FAMEX firms. In fact, considering the sample period as a whole, they perform on average worse than control firms, and both groups perform substantially worse than the universe of Tunisian manufacturing exporters (represented by aggregate trade flows). Only in 2004-2005 was the performance of FAMEX firms superior, but this may be an exceptional period since almost half of FAMEX firms were treated in 2005. Interestingly, FAMEX firms seem to have been hit by the global financial crisis in 2007-2008 earlier than control firms, recording a six percent drop in total exports while control firms still had positive export growth. During the recovery, FAMEX firms were growing at a slightly slower pace than control ones. This might indicate treated firms exposed themselves more to destination markets that contracted most at the outset of the crisis and that experienced a slower recovery. We will return to this conjecture later on in the paper. Also worth noting is that our overall sample, including FAMEX and control firms, accounts for a substantial share of total Tunisian exports.17 Table 3 Growth in Tunisia‘s Exports 2003-2004 2004-2005 2005-2006 2006-2007 2007-2008 2008-2009 2009-2010 2003-2010 Growth in total exports of: FAMEX firms 16% 27% 3% 12% -6% -12% 2% 42% Control firms 24% 6% 7% 18% 3% -16% 4% 51% Tunisia 21% 8% 13% 25% 21% -21% 8% 95% Share of exports by FAMEX and 59% 60% 61% 57% 50% 49% 55% 53% control firms in Tunisia total exports Source: Authors‘ calculations using the combined dataset and data from COMTRADE. Note: The row ‗Tunisia Total‘ shows the growth in exports for the country as a whole excluding phosphates based on COMTRADE data. 4. Estimation Issues The main identification problem in evaluating the impact of FAMEX on firm-level export outcomes is that program assignment is non-random, so FAMEX beneficiaries may differ from other firms in characteristics that affect both participation decisions and outcomes. This classical 17 Exports by FAMEX firms account, on average, for about 10 percent of Tunisia‘s total exports over the sample period. 10 problem of non-experimental impact evaluation methods requires estimation approaches that control for selection bias.18 As a first approach, we use a propensity-score matching difference-in-differences (PSM-DID) estimator which has been widely used in program evaluation, in particular for export promotion (Görg, Henry and Strobl, 2008; Volpe and Carballo, 2008).19 The PSM-DID method controls for selection bias by comparing changes in an outcome for program beneficiaries and for ‗observationally similar‘ control firms.20 It is based on the twin assumptions that (i) assignment to treatment (or the decision to undertake it) is independent of potential outcomes, conditional on observed pre-treatment covariates (Hirano, Imbens, and Ridder, 2003); and (ii) there is sufficient overlap in the distribution of propensity scores between the treatment and control groups to find matches for all or most treated firms. By relying on a comparison of changes in an outcome, the PSM-DID estimator also controls for unobserved time-invariant pre-program differences across firms, which could lead to self-selection into the program and influence outcomes (Blundell and Costa Dias, 2009). In our study, ‗observationally similar‘ firms in the control group will be defined using a rich set of observable firm covariates. Formally, let T and C be the treatment and control groups, respectively, and S be their common support.21 Indexing firms by i and years by t, let be an export outcome variable (say, total export sales, a destination count, or a product count), t(i) the year in which firm i enrolled in the one-year FAMEX program, and a binary treatment indicator defined as: 1 if i  T and t  t (i ) Dit   (1) 0 otherwise. Throughout, we will transform our outcome variables by taking log-differences: � ln  yit   ln  yit   ln  yi ,t 1  . (2) Suppose first, for simplicity, that all firms underwent treatment in the same year τ, and let and be firm i‘s outcome in the year before treatment and in the treatment year, respectively. Using log-changes, the PSM-DID estimator is given by: 18 This is the fundamental problem of causal inference defined by Holland (1986). 19 The seminal study is Heckman, Ichimura, and Todd (1997). 20 The rationale underlying PSM-DID is the idea of reproducing the treatment group among the control group and thus reestablishing ―the experimental conditions in a non-experimental setting‖ (Blundell and Costa Dias, 2009). The matching assumptions ensure that the only remaining difference between the groups is program participation. 21 The common support is the range of estimated propensity scores (i.e. the probabilities of receiving treatment as defined below) for which the frequency of both treated and control firms is non-zero. 11 (3) where j designates control firms and the weights are determined by a propensity-score matching algorithm. A PSM algorithm matches each treated firm with the set of control firms that are ‗most similar‘, that is, those with the closest propensity score. The latter is the estimated probability obtained from a regression of the probability of receiving the FAMEX treatment in any sample year on a set of firm covariates likely to be associated with selection into the program.22 One concern with the treatment indicator as defined in Eq. (1) is that it would allow the PSM algorithm to match a treated firm with a treatment group firm (possibly itself) after the treatment is over, in essence considering treated firms after treatment as if they were control firms. The treatment indicator in Eq. (1) considers a treated firm symmetrically before the treatment and after it is over, thus building in an assumption that the treatment effect is transient. An alternative coding of as equal to one for all years after treatment would build in the opposite assumption, namely that the treatment effect is permanent. In order to get around this, we follow standard practice and recode for treated firms as equal to a missing value for all t > t(i) and all t < t(i) – 1. That is, letting a period (―.‖) stand for a missing value, 1 if i  T  S and t  t  i   Dit   . if i  T  S and t  t  i  (4) 0 if i C  S  The recoded treatment indicator Dit is the dependent variable in the propensity score regression and will be a regressor in the treatment-effect equation. While we present estimates from the PSM-DID estimator in Section 5, a complication arises in our setup because the treatment year is not the same for all firms: some firms joined FAMEX in 2005 and others in each subsequent year up to 2009. Designating as before by t(i) firm i‘s treatment year, the before-after difference in outcomes is thus:      � ln yi ,t i   ln yi ,t i   ln yi ,t i 1  (5) 22 Depending on the algorithm, there can be for each treated firm either one matched control firm – the case of nearest-neighbor matching that we do not detail here - or several control firms with a weighting scheme as in Eq. (3). We consider kernel matching which uses a weighted average of all control-group firms in the common support as a match for each treated firm i, with weights being higher for control firms with a propensity score that is closer to the propensity-score of the treated firm. See Caliendo and Kopeinig (2008) for additional details on matching. 12 instead of Eq. (2). This expression is well defined for treated firms but not for control firms, for which there is no treatment year. In standard statistical packages for propensity-score matching estimation (such as psmatch2 in STATA), treated firms are matched with control firms in any year, which may be problematic if calendar time matters for performance. We address this issue by using a weighted least squares (WLS) estimator shown by Hirano, Imbens, and Ridder (2003) (henceforth HIR) to be a good alternative to PSM-DID.23 The HIR estimator uses the estimated propensity scores to construct weights for the observations used for ˆ i be the estimated a difference-in-differences treatment-effect regression. Formally, let p ˆ i  its estimated odds. ˆ i / 1  p ˆi  p propensity score of firm i (be it a treated or control one) and r The HIR estimator‘s regression-weighing scheme is given by:  1 if i  T  S �i   rˆi if i  C  S . ˆi to each That is, the scheme assigns a unit weight to all treated firms and a weight equal to r control firm, weighing more heavily those with a higher propensity score (the more ―treatable‖ firms among the untreated). The advantage of a regression framework is its flexibility in including covariates and controls, in particular year fixed effects which matter for us given that our sample period includes large macroeconomic swings. Formally, our baseline treatment-effect equation can be written as a weighted difference-in- differences (DID) regression estimated using the weights defined as above: � ln  yit   �  � Dit  Xit γ  � t  uit (6) where δt controls for calendar year effects (macroeconomic cycles), and X it is a vector of firm covariates including age and its square, a categorical variable for firm employment, a dummy variable identifying whether the firm exports 100 percent of its output, lagged total export value, the lagged number of destinations served, lagged number of export products, and location and sector fixed effects.24 23 See DiNardo, Fortin and Lemieux (1996) and Van de Walle and Mu (2007) for applications of the method. HIR show that the WLS regression estimator of the average treatment effect is actually more efficient than the PSM-DID estimator. 24 As Bernard, Redding, and Schott (2010) show, more productive firms tend to export more products to more destinations, therefore controlling for the lagged number of products and destinations is a way of controlling 13 A limitation of the regression in Eq. (6) is that it restricts the measured treatment effect to its short-term impact on . In order to measure the persistence of the FAMEX treatment effect on the growth and on the levels of the outcome variables, we recode either the treatment variable or the outcome variable as follows. First, in order to test for the persistence of treatment effects on outcome growth, we lag the treatment variable by k years, with k = 0,…,5, rewriting Eq. (6) as: � ln  yit   �  � Di ,t k  Xit γ  � t  uit . (7) This can be thought of equivalently as ―forwarding‖ the LHS variable by k periods. We will use this to explore whether a firm treated in, say, 2005 experienced higher growth in the outcome than control firms between 2005 and 2006 (k = 1), between 2006 and 2007 (k = 2), and so on. The regression in Eq. (7) will be estimated separately for each k.25 Second, in order to test for the persistence of treatment effects on outcome levels, we take ―long‖ differences in the outcome variable relative to the year prior to treatment for each of the subsequent k years, with k = 0,…,5, � k ln  yit   ln  yit   ln  yi ,t k  (8) and estimate the modified regression: �k ln  yit   �  � Dit  Xit γ  � t  uit (9) This will test for cumulative treatment effects, i.e., whether a firm treated in 2005 was still ahead of control firms in 2006 (k = 1), in 2007 (k = 2), and so on, (always relative to 2004, its pre- treatment year). The regression in Eq. (9) will be estimated separately for each k. 26 indirectly for TFP, for which we have no data. On this, see the discussion in Footnote 24 of Volpe and Carballo (2008). 25 The estimating sample for each regression includes all available one-year differences for control firms, but for treated firms the one-year differences included differ across regressions. For k=1, the regression sample includes the one-year difference (t(i)+1) - t(i) for treated firm i, for k=2, the regression sample includes the one-year difference (t(i)+2) - (t(i)+1) for treated firm i , and so on. 26 For each k, the regression sample includes all available long differences of k length for control firms but for treated firms the long differences included differ across regressions. For k=1, the regression includes the long difference (t(i)+1) - (t(i)-1) for treated firm i, for k=2, the regression includes the long difference ( t(i)+2) - (t(i)-1) for treated firm i , and so on. 14 For both Eqs. (7) and (9), sample size is reduced as k rises, since for firms treated in 2009 lagged treatment or long differences in the outcome variables are not defined; for firms treated in 2008 only k = 1 lagged treatment or long differences are defined, and so on. Only firms treated in 2005 have a full set of lagged treatments and long differences defined. So as k rises the treatment group shrinks and for k = 5 lagged treatments and long differences are defined only for firms treated in 2005. However, the sample shrinkage is limited as later FAMEX cohorts were relatively small compared to the 2005 cohort. Moreover, since the WLS regressions condition on the year of treatment through year fixed effects, this is not a major problem. Nevertheless, we will address this issue in a robustness check in Section 5.3. Finally, as an alternative approach to deal with differences in treatment year across firms, we will use as a robustness check a PSM-DID estimator that restricts control firms matched to treated firm i to be taken in year t(i), following Todo (2011), as discussed in Section 5.3.27 Before turning to estimation results, a caveat is necessary. While Glazerman, Levy, and Myers (2003) show that PSM successfully reduces selection bias in impact evaluation, particularly when combined with DID or WLS regression (these methods are indeed pervasive in the evaluation of public programs), it has well-known limitations. In particular, estimates may still be biased if unobserved time-varying firm characteristics affect both participation and outcomes. In a non-experimental study, selection bias on time-varying unobservables (say, management changes) can never by fully ruled out. 5. FAMEX Effects: Short-Term Impact and Persistence This section presents and discusses estimates from the first-stage propensity score matching as well as treatment-effect estimates, along the intensive and extensive margins, using both PSM- DID and WLS regressions. We conclude it with an assessment of the robustness of the estimates. 5.1. First-Stage PSM Estimates Propensity scores are retrieved from a cross-sectional probit regression of the probability of receiving a FAMEX grant in any year between 2005 and 2009 on the same set of firm covariates included in the treatment-effect equation listed above.28 The probability of treatment correlates positively with location in Tunis and the number of export products and destinations as shown in 27 Todo (2011) uses PSM-DID estimators to evaluate the impact of Japanese aid-funded technical assistance programs on Indonesian foundry firms, based on propensity score matching done year by year. 28 We refrain from matching on sector because the small size of sector sub-samples would drastically reduce the matching‘s quality. 15 Appendix Table A.1. It correlates negatively with total exports and 100-percent exporter status. Sector dummies are all insignificant, suggesting no sectoral targeting. Appendix Figure B.1 shows the propensity-scores distributions for the treatment and control groups. They have a large common support, which includes 401 FAMEX beneficiaries and 2,346 control firms. They are also fairly similar in shape, although not completely overlapping, highlighting the importance of matching.29 Following Rosenbaum and Rubin (1983), Dehejia and Wahba (2002) and Smith and Todd (2005), we run balancing tests to assess whether matching corrects for significant differences in the distribution of pre-treatment covariates between the treatment and control groups (Caliendo and Kopeining, 2008). They are all satisfied (see Appendix B). 5.2. Treatment-Effect Estimates Table 4 shows PSM-DID estimates using kernel matching30 and WLS regression estimates using HIR weights for three firm-level outcomes: (i) total exports in Tunisian dinars, (ii) the number of export destinations, and (iii) the number of export products, all in log-differences, and pooling together firms treated in different years. PSM-DID standard errors are based on the asymptotic variance estimator, as bootstrapped standard errors can be invalid in this context because their asymptotic properties are not known (Abadie and Imbens, 2006). However we obtain qualitatively similar findings when using bootstrapped standard errors. WLS regressions use robust White-corrected t-statistics. The first two columns show the short-term impact effect, i.e., the effect in the treatment year t(i) (henceforth designated as ‗TY‘), using PSM-DID in column (1a) and WLS in column (1b). Columns (2)-(6) show estimates of the persistence of treatment effect on growth using Eq. (7). The short-term impact effects in columns (1a) and (1b) are very similar and significant at the one percent level for all three outcome variables. Considering total exports (first row of Table 4), the estimate in column (1b) indicates that growth in total exports (measured by the log-difference) is 0.511 higher for FAMEX beneficiaries than for the control group. Short-term treatment effects at TY are also positive and highly significant for destination and product counts. 29 The importance of a large common support and similarity in the distribution of covariates/propensity scores across treated and control groups for unbiased matching DID estimators is discussed by Heckman, Ishimura and Todd (1997). 30 We also used five-nearest neighbors matching with caliper 0.1. Results, available upon request, are very similar. 16 Columns (2)-(6) show different patterns of persistence for treatment effects depending on the outcome. For total exports (first row), treatment effects become insignificant as early as the year after treatment. For destination counts (second row) and product counts (third row), by contrast, growth rates remain significantly different between the two groups of firms throughout the sample period, suggesting that once FAMEX firms learnt how to break into new markets, they kept on adding more and more markets and retrenched less than control firms during the global financial crisis.31 Table 4 Year-to-Year Effects of FAMEX on Export Outcomes Difference TY-(TY-1) TY-(TY-1) (TY+1)-TY (TY+2)-(TY+1) (TY+3)-(TY+2) (TY+4)-(TY+3) (TY+5)-(TY+4) Estimator PSM-DID WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1a) (1b) (2) (3) (4) (5) (6) Outcome Total exports 0.496** 0.511*** 0.251 -0.042 -0.157 -0.240 0.025 [2.66] [3.08] [1.55] [-0.26] [-0.83] [-1.06] [0.11] R-squared 0.17 0.14 0.11 0.09 0.11 0.11 Nb. destinations 0.144*** 0.150*** 0.086*** 0.052** 0.021 0.036 0.059** [5.52] [6.10] [3.70] [2.10] [0.84] [1.11] [2.07] R-squared 0.15 0.12 0.08 0.12 0.12 0.08 Nb. products 0.145*** 0.147*** 0.071** 0.049 0.008 0.060 0.097*** [4.33] [4.68] [2.22] [1.59] [0.23] [1.59] [2.58] R-squared 0.15 0.13 0.13 0.12 0.13 0.13 Observations 12,263 12,214 9,803 7,401 4,975 2,607 Notes: T-statistics based on robust standard errors in brackets; *: significant at 10%; **: significant at 5%; ***: significant at 1%. The sample includes treated and control firms in the common support. PSM-DID estimates are based on propensity scores obtained using kernel matching. The WLS regressions include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. Table 5 shows cumulative treatment effects estimated from Eq. (9), using long differences in log-outcomes as defined in Eq. (8). Thus, whereas the question in columns (2)-(6) of Table 4 was ―for how long do growth trajectories diverge‖, the question in Table 5 is ―for how long do outcome levels remain different after a temporary growth surge‖. Column (1) reproduces column (1b) of Table 4. Columns (2)-(6), which rely on increasingly longer differences, show again different patterns depending on the outcome. In the first row (export values), columns (4)-(6) show that cumulative effects on levels disappear three years after treatment. In the second and 31 The latter part of the statement reflects the fact that column (3) onwards captures the years of the global financial crisis. In columns (3) and (4) the treated firms in the estimating sample are those receiving FAMEX in 2005 or in 2006 for which the TY+3 and TY+4 years are, respectively, 2008 and 2009 or 2009 and 2010. In column (5) the treated firms in the estimating sample are those receiving FAMEX in 2005 for which the TY+5 year is 2010, the last year in our sample. 17 third rows (destination and product counts respectively), again cumulative effects persists, which is to be expected if growth rates keep on diverging. Table 5 Cumulative Effects of FAMEX on Export Outcomes Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) Outcome Total exports 0.511*** 0.723*** 0.571** 0.272 0.043 0.200 [3.08] [3.59] [2.57] [1.02] [0.13] [0.52] R-squared 0.17 0.22 0.23 0.22 0.23 0.25 Nb. destinations 0.150*** 0.191*** 0.190*** 0.151*** 0.143*** 0.177*** [6.10] [6.93] [5.91] [4.18] [3.03] [3.22] R-squared 0.15 0.20 0.20 0.24 0.29 0.30 Nb. products 0.147*** 0.175*** 0.178*** 0.117** 0.156*** 0.219*** [4.68] [4.70] [4.42] [2.51] [2.66] [3.37] R-squared 0.15 0.20 0.23 0.26 0.27 0.30 Observations 12,263 12,124 9,664 7,238 4,839 2,524 Notes: T-statistics based on robust standard errors in brackets; *: significant at 10%; **: significant at 5%; ***: significant at 1%. The sample includes treated and control firms in the common support. The WLS regressions include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. Figure 1 summarizes in a bar chart the findings of Tables 4 and 5: the temporary effect on the intensive margin and the durable effect on the extensive margin. Figure 1 Year-to-Year and Cumulative Effects of FAMEX a. Year-to-Year Effects (Table 4) b. Cumulative Effects (Table 5) 80% 60% 50% 70% Total exports Total exports Nb. destinations 40% 60% Nb. destinations Nb. products 30% 50% Nb. products 20% 40% 10% 30% 0% 20% TY TY+1 TY+2 TY+3 TY+4 TY+5 -10% 10% -20% 0% -30% TY TY+1 TY+2 TY+3 TY+4 TY+5 Note: Bar heights show the point estimates of the coefficients in Table 4 and Table 5. Insignificant effects are shown as dotted bars. 18 All in all, results so far suggest a contrast between a large but transient treatment effect on total exports and moderate but persistent effects on destination and product counts. We also explore whether new destinations and new products added by FAMEX firms to their portfolios survive better than new destinations and products launched in the same years by control firms. Our estimates from WLS Tobit regressions shown in Appendix C provide evidence that this is indeed the case for new destinations but not for new products. 5.3 Robustness The treatment effects shown in Tables 4 and 5 are based on a sample which includes 126 FAMEX program dropouts in the control group. If those are more similar to FAMEX recipients than other control firms, their inclusion in the control group should improve the quality of the matching and the accuracy of the estimates of the treatment effect. 32 However, as a robustness check we re-estimate Eqs. (7) and (9) using two alternative sample constructions: (i) eliminating dropouts from the sample or (ii) including them in the treatment group. Results are shown in Appendix Table D.1. In accordance with intuition, the elimination of dropouts gives a slightly larger treatment effect, whereas the opposite is true when they are included in the treatment group. That is, dropouts performed better than other control firms but worse than the FAMEX firms that took their export business plan to full completion. Second, the sample size shrinks across columns (2)-(6) of Tables 4 and 5 due to the fact that several lagged treatments and long differences in outcomes are not defined for firms that enrolled in the FAMEX program later in the sample period (as discussed in Section 4) and due to the unbalanced nature of the panel for control firms (as well as for treated firms).33 We re-estimate the specifications in Tables 4 and 5 using a fixed sample across columns that is restricted to include only firms enrolled in FAMEX in 2005 and control firms (in the common support) operating continuously in export markets between 2004 and 2010. The results are qualitatively maintained (see Appendix Table D.2). Third, as an alternative to the HIR procedure, we address the problem of ‗time-wise mismatch‘, i.e., matching a firm treated in t(i) with a control firm at time t '  t  i  , by using a procedure suggested by Todo (2011).34 Todo‘s procedure pairs each treated firm with control firms 32 An ideal specification would consider a control group including only dropouts. However, such a specification cannot be estimated given the small number of dropouts relative to the number of FAMEX firms. 33 We are grateful to Beata Javorcik for attracting our attention to this issue. 34 We are grateful to Yasusuki Todo for sharing his matching routine. See also Arnold and Javorcik (2009) for a discussion of time-wise mismatch. 19 observed in i‘s treatment year t(i), generating, for each firm i treated in t(i), a fictitious composite control with an outcome calculated as a weighted average of the outcomes observed in that year for control firms with propensity scores close to i‘s. This results in a new dataset whose size is just twice that of the treatment group and where each treated firm is matched with a unique composite control. We pool across all treatment years and re-estimate the equivalent of Eq. (9) by OLS. Results are shown in Appendix Table D.3 and are very similar to those of Table 5. Finally, we re-estimate Eq. (9) using the number of HS 6-digit products exported instead of the number of 11-digit products (in Tunisia‘s nomenclature) exported, with results that are very similar to those in Table 5 (see Appendix Table D.4). 6. Extensions 6.1 Did FAMEX Assistance Encourage Risk-Taking or Facilitate Hedging? The experimental economics literature has long noticed that individuals tend to be more willing to take risks out of windfall gains than out of regular earnings, a phenomenon called the ―house money effect‖ (a term borrowed from gambling).35 We noted in Table 3 that FAMEX firms performed worse in terms of export growth than control firms in the early stages of the global financial crisis. Although FAMEX was a matching-grant program rather than a pure subsidy— precisely in order to limit moral hazard—could it be nevertheless that export promotion encouraged beneficiary firms to take more risks? As a preliminary, it is worth noting that the treatment-induced growth at the extensive margin (destinations and products) observed in Tables 4 and 5 did translate into lower concentration. To verify this, we re-estimate Eq. (9), our cumulative treatment-effect equation, using, as new outcome variables, Herfindahl and Theil concentration indices across product-destination cells at the firm level. The results are reported in Table 6 and suggest that concentration indeed declines for treatment firms relative to control ones, with an effect that seems to have been persistent. 35 For instance, Thaler and Johnson (1990) asked two groups of individuals to choose between two lotteries with the same expected value, one riskier than the other. The treatment group was given a prior endowment (the windfall) while the control group was given no endowment. The y report that 77 percent of the treatment group‘s individuals chose the risky lottery against only 44 percent for the control group. 20 Table 6 Effects of FAMEX on Concentration and Volatility Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Before-After 2005 Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) (7) Outcome Herfindahl index -0.131*** -0.116*** -0.151*** -0.183*** -0.172*** -0.303*** [-4.35] [-2.91] [-3.61] [-3.93] [-3.03] [-4.26] R-squared 0.08 0.11 0.09 0.11 0.11 0.14 Theil index -0.021*** -0.021*** -0.027*** -0.030*** -0.031*** -0.052*** [-4.82] [-3.57] [-4.35] [-4.33] [-3.68] [-4.82] R-squared 0.10 0.14 0.12 0.14 0.16 0.17 Observations 7,743 7,308 5,627 4,059 2,629 1,326 Coeff. variation of exports -0.023 [0.51] R-squared 0.13 Observations 1,198 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%; *** significant at 1%. The sample in columns (1)-(6) includes all treated firms and control firms in the common support. The sample in column (7) includes only treated firms in 2005 and control firms operating in export markets continuously from 2000 to 2009 (one observation per firm). The WLS regressions in columns (1)-(7) include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. However, a crude measure of export volatility at the firm level suggests that increased diversification failed to translate into lower risk. We use the time-wise coefficient of variation of total export turnover at the firm level before versus after FAMEX, considering only firms that enrolled in 2005, for which five years are observed both before and after treatment. For those firms as well as for a set of control firms which exported continuously between 2000 and 2009, we obtain additional customs data going back to 2000 to calculate the coefficient of variation over the five years prior to 2005. We then re-estimate Eq. (7) by WLS using as dependent variable the before 2005-after 2005 difference in the coefficient of variation. The results are shown in column (7) of Table 6.36 In spite of enhanced diversification, the evidence does not show significant benefits from FAMEX in reducing export volatility for beneficiary firms relative to control firms. Several reasons might lie behind this. Relative to control firms, FAMEX beneficiaries might have experimented with new destinations/products and as a result failed with higher probability; or they might have ventured into inherently more risky markets; or, finally, they might have expanded into markets where returns (in terms of export sales) were 36 The sample size in column (7) is smaller than in the other columns due to the constraint of considering only firms with uninterrupted exports between 2000 and 2009. In calculating the coefficient of variation for each firm, data for the year 2010 is ignored in order for the two periods considered - 2000-2004 and 2005-2009 - to have the same number of years. 21 correlated with those in their existing markets, thus failing to diversify their portfolio in a way that reduced risks. 6.2. Did FAMEX Generate Spillover Benefits? Estimated treatment effects can be biased in the presence of general equilibrium effects —not very likely in the case of an assistance program of limited scale like FAMEX—or by externalities ―polluting‖ the control group‘s outcomes.37 In particular, program benefits will be underestimated if the control group‘s export outcomes improve as a result of spillovers from beneficiaries. Thus, if one does not control for externalities, the absence of treatment effects has two possible interpretations: Either the program was ineffective, or it was effective but its benefits were not appropriable. This is not just a technical issue. Non-appropriability of benefits from information production (say, marketing research) in the presence of externalities can be seen as the market failure which provides the reason for subsidized intervention by the Tunisian government and the World Bank.38 In our case, the central result—lack of persistence of treatment effects on export values—might reflect catching up by control firms rather than vanishing benefits for treated firms, although catching up through imitation should apply equally—perhaps even more—to the extensive margin, for which we do observe permanent divergence. For instance, FAMEX beneficiaries‘ actions, such as participating in trade fairs or hiring export-marketing consultants, could have been visible to and easily imitable by other firms in their sector or location. Information acquired by FAMEX beneficiaries might even have been shared voluntarily with other firms, as exporters from the same country do not necessarily see themselves as competitors on foreign markets.39 The difficulty in investigating this issue explicitly is that the measurement of spillovers is elusive, especially when the transmission channel is unknown. Our exposure variable is a time- variant count of the number of FAMEX beneficiaries in each sector-region-year cell, assuming that firms are more likely to benefit from externalities if they produce similar goods in the same 37 Formally, treatment effects are measured under the assumption of ―stable unit treatment value assumption‖ (Rubin 1980), which means that a treated individual‘s outcome is independent of the treatment‘s mode of administration and of the status of other individuals (treated or not). 38 Credit rationing is another market failure that could justify government intervention. However, if the government were to provide export credit services in lieu of deficient financial markets, the benefits would be appropriable and the services should be extended on a full-cost recovery basis rather than as a matching grant. 39 Cadot, Iacovone, Pierola, and Rauch (2011) show that, for African exporters, expected survival rises with the number of firms from the same country exporting the same product to the same destination. Whether export entrepreneurship creates externalities that need to be supported by public action, as argued e.g. in Hausmann and Rodrik (2003), is still largely an open question. 22 region.40 We then regress control firm export outcomes on exposure to FAMEX beneficiaries.41 That is, � ln  y jsrt   �  k �k n jsr ,t k  � j  � st  � jsrt (10) where s and r designate, respectively, sectors and regions and n jsr ,t k is the number of FAMEX beneficiaries in control firm j ‘s sector-region in year t  k . Firm fixed effects � j account for unobserved firm heterogeneity in growth of export outcomes. The equation also includes sector- year fixed effects � st and region-year fixed effects � rt to control for shocks that could affect both outcomes and the number of firms receiving FAMEX support in a sector or region. The exposure variable enters with various lags to mitigate endogeneity and, more importantly, to allow for the slow diffusion of externalities. Estimates for Eq. (10) shown in Table 7 fail to suggest any positive externalities; indeed, the only instances of significant coefficients for total exports in column (1) and for the number of products in column (12) are negative. We re-estimate a variant of Eq. (10) with a sample including all firms, both FAMEX beneficiaries and control firms, and again find no evidence of externalities (see Appendix Table D.5). One might argue that spillovers to control firms are more likely to emerge from FAMEX firms whose objective was to expand into new destinations or to export new products. To address this possibility, we re-estimate Eq. (10) using a variant of the exposure variable that counts only FAMEX firms whose objective was to reach more export destinations or export more products. Not even this type of firms generated externalities to control firms (see Appendix Table C.5). 40 Our spillover proxy follows standard practice. For instance, Krautheim (2012) develops a trade model with heterogeneous firms including a spillover effect from the number of exporters to the fixed costs of exporting. Empirical studies on export spillovers such as Aitken, Hanson, and Harrison (1997), Bernard and Jensen (2004) and Kneller and Pisu (2007) use the presence of exporters in an industry and location to capture spillovers. Recent studies using customs data such as Koenig, Mayneris, and Poncet (2010) and Mayneris and Poncet (2010) also test for the presence of spillovers using the numbers of exporters in the same region exporting similar products and/or to similar destinations. We also test for spillovers from the exporting activity itself and find no significant effect. 41 One weakness with this approach is that we are unable to measure spillovers on Tunisian firms other than the control firms because the data agencies did not allow access to data on the universe of Tunisian manufacturing firms. Another weakness is that we cannot explore spillovers on outcomes other than exports, again because we do not have the necessary data. 23 Table 7 Effect of Exposure to FAMEX Firms on Control Firms‘ Export Outcomes Estimator Within reg. Within reg. Within reg. Difference t-(t-1) t-(t-1) t-(t-1) Outcome Total exports Nb. destinations Nb. products Sample of control firms only Sample of control firms only Sample of control firms only (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Exposure to FAMEX benef. t-1 -0.052* -0.050 -0.016 -0.122 -0.003 -0.004 0.004 -0.000 -0.006 -0.006 -0.000 -0.022* [-1.79] [-1.64] [-0.39] [-1.39] [-1.04] [-1.27] [0.87] [-0.03] [-1.49] [-1.56] [-0.03] [-1.95] Exposure to FAMEX benef. t-2 0.004 0.037 -0.019 -0.002 0.005 -0.005 -0.001 0.005 -0.020 [0.14] [0.85] [-0.18] [-0.75] [1.25] [-0.47] [-0.33] [0.83] [-1.44] Exposure to FAMEX benef. t-3 0.012 -0.028 0.005 -0.004 0.006 -0.015 [0.31] [-0.28] [1.39] [-0.43] [1.12] [-1.14] Exposure to FAMEX benef. t-4 -0.060 -0.008 -0.022** [-0.76] [-1.11] [-2.05] Number of firms 2,620 2,620 2,618 2,618 2,620 2,620 2,618 2,618 2,620 2,620 2,618 2,618 Firm fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Location-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.02 0.02 0.02 0.02 0.01 0.01 0.01 0.01 0.02 0.02 0.02 0.02 Observations 12,785 12,785 10,316 7,802 12,785 12,785 10,316 7,802 12,785 12,785 10,316 7,802 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%. The sample includes only control firms in the common support or outside. 6.3. Did Firms’ Objectives and Use of FAMEX Support Matter? The detailed information available on FAMEX allows us to exploit two dimensions potentially affecting the treatment effect: (i) the objective of the individual project supported by FAMEX, and (ii) the specific activities supported by the FAMEX grant. To our knowledge, this type of analysis is novel in the export promotion literature. Tunisian firms had to state an objective when applying for FAMEX assistance, whether they wanted to: (i) become a significant exporter, (ii) export to a new destination market, or (iii) export a new product. Given the way the FAMEX application packages were structured, firms could state only one of these three objectives. Hence the project objectives partition the 401 treated firms into three non-overlapping groups: 95 came to FAMEX to become a more substantive exporter (in terms of the exports to sales ratio), 194 came to FAMEX to export to a new destination, and 112 came to FAMEX to export a new product. We re-estimate cumulative treatment effects of FAMEX allowing the treatment effect in Eq. (9) to differ across objectives and show the results in Appendix Table E.1. Firms that approached FAMEX with the objective of expanding into either new destinations or new products, saw a significant and sustained expansion in terms of the both the number of destinations and products, and a significant but temporary expansion in total exports. Firms which had the objective of expanding total exports saw less significant benefits along any dimension. Together, these results 24 suggest that export promotion is most effective in helping firms break into new markets and new products. Consider now the types of activities supported by the FAMEX program financed listed in Table 1. We replace the binary treatment variable used so far with a vector of continuous variables measuring, for each firm, the amount of FAMEX funding earmarked under each type of activity (the equivalent of the first column of Table 1, but at the firm level) which is available for 328 FAMEX beneficiaries.42 The results from re-estimating Eq. (9) results from re-estimating Eq. (9) are shown in Appendix Table E.2 show that market prospection activities and promotion activities have a significant effect on all outcome variables up to four years after treatment, with both activities exhibiting similar marginal returns on the dinar. Firm development has a significant positive effect on all outcomes but only four or five years after treatment, which may be due to the longer gestation period needed for such activities bring export benefits. By contrast, the other two types of activities have insignificant returns. Again, we see a broad correspondence with our earlier results: activities that support diversification have a clear positive return. A large number of FAMEX firms—200—had an in-house export unit prior to the start of FAMEX assistance. It is natural to expect such firms to exhibit stronger performance since their in-house dedicated export units may enable them to make better use of the FAMEX assistance. To address this possibility we allow the treatment effect to differ according to whether the firm had an in-house export unit. Results in Appendix Table E.3 suggest that FAMEX beneficiaries with a dedicated in-house export unit exhibit stronger growth in total exports and in the number of destinations served, relative to other FAMEX beneficiaries.43 6.5. A Tentative Cost-Benefit Assessment Our baseline results suggest that FAMEX had a large and positive—albeit short-lived—effect on total exports of treated firms. We turn in this section to a tentative cost-benefit calculation to 42 Using the vector of amounts instead of bins for the different types of activities allows us to avoid the problems of multicollinearity across the individual components of the treatment that would arise due to their large overlap. For each activity, the amount entering in the WLS regressions is the amount co-financed at 50 percent by FAMEX. The total amount spent by the firm in that activity is twice as large. In this exercise we control for selection into FAMEX through the usual propensity score matching weighting scheme, but we do not control for selection into particular levels of support for each activity. 43 We also examine whether firm size affects the magnitude of the impact of FAMEX assistance, classifying firms into two broad size categories: below 50 workers (196 firms) and above 50 workers (205 firms). We re-estimate Eq. (8) by WLS, allowing the FAMEX effect to differ across size categories. The estimates suggest that FAMEX has a strong effect in the year of treatment for both small and large firms, and the magnitudes of the effects are quite similar across size categories for all export outcomes (see Appendix Table E.3). 25 estimate the rate of return of the FAMEX program per firm, laying out clearly at each step the assumptions made. The details of the cost-benefit calculations are provided in Appendix F. We first consider FAMEX benefits based on the estimated effect of FAMEX on total export growth in the year of treatment: 0.511 (in Tables 4 and 5). This implies that FAMEX beneficiaries had 66.7 percentage points higher export growth than control firms.44 Since the average annual total export growth for control firms in the 2004-2008 period was 8.35 percent, the estimated annual growth in total exports for a FAMEX beneficiary in the year of treatment is 13.9 percent.45 Given average total exports per firm in 2004 (prior to FAMEX) of 2,308 thousand Tunisian Dinars, the growth rates above imply that, in the year of treatment, exports for a typical FAMEX beneficiary would have increased to 2,629 thousand Tunisian Dinars against only 2,501 thousand Tunisian Dinars for a control firm, a gain of 129 thousand Tunisian Dinars. After two years, the difference would have peaked at 483 thousand Tunisian Dinars, and after three years it would become insignificant. The implied trajectory of exports for a typical FAMEX beneficiary and a typical control firm are illustrated in Figure 2. Figure 2 Evolution of Total Exports for Typical FAMEX Beneficiary and Control Firm No FAMEX FAMEX 4000 3500 Average Total Exports per Firm (thousands of Tnusian Dinars) 3000 2500 2000 1500 1000 500 0 TY-1 TY TY+1 TY+2 TY+3 TY+4 TY+5 Note: The figure is based on the numbers presented in Appendix Table F.1. The average grant amount disbursed by the FAMEX program per firm was 21.7 thousand Tunisian Dinars. Thus, on impact in the treatment year, the implied rate of return on public funds would be almost 6 Tunisian Dinars of additional exports per Tunisian Dinar of grant. Over the three years where exports of FAMEX beneficiaries were significantly higher than those of 44 This figure is obtained as exp(0.511)-1. 45 We calculate the average annual total export growth over the period 2004-2008 to avoid including the years after the onset of the global financial crisis. The figure of 13.9 percent is obtained as 8.35% * (1+66.7%). 26 control firms, the additional total exports per Tunisian firm generated per Tunisian Dinar of publicly-funded grant would be 22. Although consistent with the emerging body of empirical results on the impact of export promotion, these rates of return are surprisingly high. It is important to note that they are an upper bound, first because they are based on the grant-component cost of FAMEX and do not take into account the overhead administrative costs of the FAMEX program for which we have no information. Second, a more meaningful cost-benefit analysis would focus on the increase in producers‘ surplus generated by FAMEX, rather than simply the increase in aggregate exports but again we do not have the necessary data. 7. Concluding Remarks Trade promotion policies are increasingly popular, but evidence on their impact is limited, in particular over the long run. Our paper adds to the small existing literature by taking a longer view of the impact of interventions. This longer-term perspective enabled us to ask three questions: (i) Is any export expansion along the intensive and extensive margins durable? (ii) If there is indeed diversification of exports, does it reduce the longer-term volatility of exports? (iii) Are there spillover benefits for other firms, especially over time? We find the Tunisian export promotion program –FAMEX– had a durable impact along the extensive margin in terms of export destinations and products but provided only a temporary boost to exports along the intensive margin. The results are robust and supported by evidence on the specific ingredients of the program: market prospection and promotional activities prove to be the most fruitful. However, our results on the long-term impact of export promotion must be interpreted cautiously given that the later years of our sample period were characterized by the collapse of world trade, which may not have affected all firms equally. In particular, treated firms may have ventured into riskier markets. Our evidence shows that the increased diversification of beneficiary firms did not result in reduced volatility of exports. In terms of policy implications, it is important to recognize that treatment effects, which are generally construed as favorable to policy intervention when they are significant, only reflect appropriable benefits. They give no indication of the presence of a market failure. Indeed, if anything, it is the absence of treatment effects that is consistent with the non-appropriability of benefits. We attempt to address this problem by testing directly for the presence of externalities but find no evidence of positive externalities from treated firms to control ones. 27 All in all, whereas our results do suggest, like previous findings, that export promotion has a strong contemporaneous effect on the export performance of treated firms, the effects we identify are not sustained and do not seem to spill over to untreated firms, two aspects had so far been largely overlooked in the literature. 28 References Abadie, A. and G. Imbens (2006). ―Large Sample Properties of Matching Estimators for Average Treatment Effects‖; Econometrica 74, 235-267. Alvarez, R., and G. Crespi (2000). ―Exporter Performance and promotion Instruments: Chilean Emprical Evidence‖; Estudios de Economia 27, Universidad de Chile. Aitken, B., Hanson, G., and A. Harrison (1997). ―Spillovers, Foreign Investment and Export Behavior‖; Journal of International Economics 43, 103-132. Arnold, J. and B. Javorcik (2009). ―Gifted Kids or Pushy Parents? Foreign Direct Investment and Plant Productivity in Indonesia‖; Journal of International Economics 79, 42-53. Bernard, A. and B. Jensen (2004). ―Why Some Firms Export‖; Review of Economics and Statistics 86, 561-569. Bernard, A., Redding, S., and P. Schott (2007). ―Firms in International Trade‖; Journal of Economic Perspectives 21, 105-130. Bernard, A., Redding, S., and P. Schott (2010). ―Multi-product Firms and Trade Liberalization‖; Quarterly Journal of Economics 126, 1271-1318. Blundell, R. and M. Costa Dias (2009). ―Alternative Approaches to Evaluation in Empirical Microeconomics‖; Journal of Human Resources 44, 565-640. Caliendo, M. and S. Kopeinig (2008). ―Some Practical Guidance for the Implementation of Propensity Score Matching‖; Journal of Economic Surveys, 22, 1, pp. 31-72. Cadot, O., Iacovone, L., Pierola, M., and F. Rauch (2011). ―Success and Failure of African Exporters‖; World Bank Policy Research Working Paper 5657. Dehejia, R. and S. Wahba (2002). ―Propensity Score Matching Methods for Nonexperimental Causal Studies‖; Review of Economics and Statistics 84, 151-161. DiNardo, J., Fortin, N., and T. Lemieux (1996). ―Labor Market Institutions and the Distribution of Wages, 1973-1992: A Semiparametric Approach‖; Econometrica 64, 1001-1044. Girma, S., Gong, Y., Görg, H., and Z. Yu (2009). ―Can Production Subsidies Explain China‘s Export Performance? Evidence from Firm-level Data‖; Scandinavian Journal of Economics 111, 863-891. Glazerman, S. Levy, D., and D. Myers (2003). ―Nonexperimental Versus Experimental Estimates of Earnings Impacts‖; The Annals of the American Academy of Political and Social Science 589, 63-93. Görg, H., Henry, M., and E. Strobl (2008). ―Grant Support and Exporting Activity‖; Review of Economics and Statistics 90, 168-174. Gourdon, J., Marchat, J., Sharma, S. and T. Vishwanath (2011). ―Can Matching Grants Promote Exports? Evidence from Tunisia's FAMEX Program‖; in Cadot, O., Fernandes, A., Gourdon, J. and A. Mattoo (eds.) Where to Spend the Next Million? Applying Impact Evaluation to Trade Assistance; The World Bank and CEPR, pp. 81-106. 29 Hausmann, R. and D. Rodrik (2003). ―Economic Development as Self-Discovery‖; Journal of Development Economics 72, 603-633. Heckman, J., Ichimura, H., and P. Todd (1997). ―Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme‖; Review of Economic Studies 64, 605-654. Hirano, K. and G. Imbens (2004). ―The Propensity Score with Continuous Treatments‖; in A. Gelman and X. Meng (2004), Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives: An Essential Journey with Donald Rubin's Statistical Family; New York, NY: Wiley, pp. 73-84. Hirano, K. Imbens, G. and G. Ridder (2003). ―Efficient Estimation of Average Treatment Effects Using the Estimated Propensity Score‖; Econometrica 71, 1161-1189. Holland, P. (1986). ―Statistics and Causal Inference‖; Journal of the American Statistical Association 81, 945-960. Kneller, R. and M. Pisu (2007). ―Industrial Linkages and Export Spillovers from FDI‖; World Economy 30, 105-134. Koenig, P., Mayneris, F., and S. Poncet (2010). ―Local Export Spillovers in France‖; European Economic Review 54, 622-641. Krautheim, S. (2012). ―Heterogeneous Firms, Exporter Networks and the Effect of Distance on International Trade‖; Journal of International Economics 87, 27-35. Lederman, D., Olarreaga, M., and L. Payton (2010). ―Export Promotion Agencies Revisited‖; Journal of Development Economics 91, 257-265. Mayneris, F. and S. Poncet (2010). ―Export Performance of China‘s Domestic Firms: the Role of Foreign Export Spillovers‖; Document de Travail No. 32, CEPII. McDermott, M. (1994). ―Is International Marketing a Game of Spy vs. Spy?‖; Brandweek 35, June 20. Rangan, S. and R. Lawrence (1999). ―Search and Deliberation in International Exchange: Learning from International Trade about Lags, Distance Effects, and Home Bias‖; NBER working paper 7012. Ravallion, M. (2008), Evaluation in the Practice of Development, World Bank Policy Research Working Paper No. 4547. Reinikka, R. and J. Svensson (2004). ―Local Capture: Evidence From a Central Government Transfer Program in Uganda‖; Quarterly Journal of Economics 119, 679-705. Rose, A. (2007). ―The Foreign Service and Foreign Trade: Embassies as Export Promotion‖; The World Economy 30, 22-38. Rosenbaum, P. and D. Rubin (1983). ―The Central Role of the Propensity Score in Observational Studies for Causal Effects‖; Biometrika 70, 41-55. Rubin, D. (1980). Discussion of ―Randomization Analysis of Experimental Data in the Fisher Randomization Test‖ by D. Basu. Journal of the American Statistical Association 75, 591-593. 30 Smith J. and P. Todd (2005). ―Does Matching Overcome LaLonde‘s Critique of Nonexperimental Estimators? Journal of Econometrics 125, 305-353. Thaler, R. and E. Johnson (1990). ―Gambling with the House Money and Trying to Break Even: The Effects of Prior Outcomes on Risky Choice‖; Management Science 36, 643–660. Todo, Y. (2011). ―Impacts of Aid-Funded Technical Assistance Programs: Firm-Level Evidence from the Indonesian Foundry Industry‖; World Development 39, 351-362. Volpe, C. and J. Carballo (2008). ―Is Export Promotion Effective in Developing Countries? Firm-level Evidence on the Intensive and Extensive Margins of Export Growth‖; Journal of International Economics 76, 89-106. Volpe, C. and J. Carballo (2010). ―Beyond the Average Effects: The Distributional Impacts of Export-Promotion Programs in Developing Countries‖; Journal of Development Economics 92, 201-214. Van de Walle, D. and R. Mu (2007). ―Fungibility and the Flypaper Effect of Project Aid: Micro- Evidence for Vietnam‖; Journal of Development Economics 84, 667-685. World Bank (2008). Tunisia‘s Global Integration: Second Generation of Reforms to Boost Growth and Employment. 31 Appendix Appendix A: Propensity Score Estimation Table A.1 Probit Regression for the Propensity to receive FAMEX treatment FAMEX treatment status Age 0.355 [1.00] Age squared -0.098 [-1.51] Lagged total exports -0.038*** [-4.35] Lagged number of exported products 0.158*** [3.40] Lagged number of export destinations 0.497*** [8.58] 100% exporter -0.341*** [-4.91] 10-19 employees -0.491*** [-4.22] 20-49 employees -0.359*** [-3.61] 50-99 employees -0.393*** [-3.71] 100-199 employees -0.385*** [-3.41] More than 200 employees -0.411*** [-3.37] Textiles and apparels -0.067 [-0.76] Paper, wood, and furniture 0.019 [0.19] Chemicals -0.041 [-0.41] Metals -0.021 [-0.18] Machine and equipment 0.017 [0.17] Electric -0.111 [-0.91] Grand Tunis -0.352*** [-4.86] Central Sea -0.950*** [-6.06] Rest of Tunisia -0.448*** [-5.81] Year fixed effects Yes Observations 12,263 Notes: T-statistics in brackets; * significant at 10%; ** significant at 5%; *** significant at 1%. Unless noted, firm characteristics refer to 2007. The omitted sector is agro-industry, the omitted location is Tunis, and the omitted category in terms of employment is less than 10 workers. 32 Appendix B: Propensity Score Matching Appendix Figure B.1 Densities and histogram of propensity scores, treatment and control groups 8 6 4 2 0 0 .2 .4 .6 .8 0 .2 .4 .6 .8 x Propensity Score kdensity untreated kdensity treated Untreated Treated Source: Authors‘ calculations. To assess the quality of the propensity score matching in balancing adequately the covariates between treatment and control groups, we conduct four types of tests. The first test is the balancing or stratification test proposed by Dehejia and Wahba (2002) which divides observations into strata based on the estimated propensity score and uses t-tests within each strata to test if the distribution of covariates is similar between the treatment and control group. such that there are no statistical differences between the mean of the propensity score in the treatment and control group. Implementing the test in stata as in Becker and Ichino (2002) over 6 strata of the propensity score shows that the balancing property is satisfied for our data. The second set of tests shown in the first columns of Appendix Table B.1 consists in two-sample t-tests for the equality of sample means for all the covariates between treated and matched control groups. The t-tests indicate no significant differences in the means suggesting that the covariates are balanced in the two groups and thus the quality of our matching is high. The third set of tests shown in the last columns of Appendix Table B.1 are the standardized biases for the covariates defined as the corresponding difference in sample means between treated and matched control groups normalized by the square root of the average of sample variances in both groups. The results show that the standardized bias for our covariates is in most cases lower than 5%. Caliendo and Kopeining (2008) suggest that a standardized bias of that magnitude after matching indicates high quality of the matching. 33 The fourth test is based on the comparison of the pseudo-R-squared of the propensity score estimated on the full sample versus on the matched sample, which explains how well the covariates explain the propensity to participate in the program. With a high quality matching, the pseudo-R-squared should be very low after matching because there should be no differences in the distribution of the covariates that can explain the propensity to participate in the program. Indeed, our pseudo-R-squared is 0.208 before matching and 0.006 after matching. Moreover, the associated likelihood-ratio test of the joint insignificance of covariates in the propensity score estimation on the full sample versus on the matched sample should indicate that the covariates are jointly insignificant in explaining participation after matching. Indeed our likelihood-ratio chi-squared test is 733.92 with a p-value of 0 before matching and 6.22 with a p-value of 1 after matching. Table B.1 Balancing Tests Mean in Matched Sample T-test Percentage Percentage Covariates Treatment Control Bias Bias T-statistic P-value Reduction Age 2.707 2.710 -0.06 0.948 -0.5 96.3 Age squared 7.651 7.665 -0.06 0.955 -0.4 96 Lagged total exports 10.038 9.961 0.17 0.862 1.3 93.2 Lagged number of exported products 1.266 1.250 0.22 0.824 1.7 94.7 Lagged number of export destinations 1.048 1.032 0.27 0.79 2.1 96 100% exporter 1.365 1.368 -0.09 0.927 -0.7 97.6 10-19 employees 0.099 0.097 0.09 0.929 0.6 94.3 20-49 employees 0.292 0.302 -0.26 0.798 -1.9 33.3 50-99 employees 0.199 0.203 -0.13 0.901 -0.9 83.3 100-199 employees 0.155 0.163 -0.31 0.758 -2.3 -434.1 More than 200 employees 0.149 0.137 0.47 0.639 3.7 72.7 Textiles and apparels 0.334 0.329 0.15 0.884 1.1 95.1 Paper, wood, and furniture 0.136 0.119 0.64 0.519 5.2 49 Chemicals 0.113 0.125 -0.5 0.619 -3.8 -28.2 Metals 0.075 0.075 -0.02 0.982 -0.2 97.7 Machine and equipment 0.135 0.141 -0.21 0.831 -1.7 81.7 Electric 0.064 0.065 -0.06 0.95 -0.5 88 Grand Tunis 0.478 0.486 -0.21 0.834 -1.6 66.6 Central Sea 0.017 0.027 -0.98 0.33 -5.1 81 Rest of Tunisia 0.301 0.296 0.16 0.874 1.2 93.6 Year 2005 0.616 0.564 1.4 0.162 12 88.8 Year 2006 0.06077 0.08974 -1.46 0.146 -9.3 74.8 Year 2007 0.215 0.21 0.17 0.867 1.4 78.7 Year 2008 0.10773 0.1276 -0.82 0.414 -5.9 68.7 Year 2009 0 0.00837 -1.75 0.081 -3 94.3 34 Appendix C: Effects of FAMEX on Survival of New Destinations and Products We explore whether new destinations and products (at the HS 6-digit level) added by FAMEX firms to their portfolios survive better than new destinations and products launched in the same years by control firms. Let nit be the number of destinations served by firm i in year t (with any of its products so nit  d nidt ), and zi,t+k be the number of those destinations still served (with any product) in year t + k and that have been served uninterruptedly between t and t + k. Let sitk  zi ,t k nit . That is, sit1 is the one-year-forward survival rate, sit2 is the two-year forward survival rate, and so on. For treated firms, we restrict t to be the treatment year and the year before, so our dependent variable is the survival rate of new destinations (or products) introduced in the year before treatment or in the treatment year and surviving uninterruptedly for k years. Our estimable equation is thus: (C.1) This equation is estimated by weighted Tobit using HIR weights with left and right censoring since the dependent variable varies between zero and one. The results are shown in Table C.1 for k = 1,…,5 across columns (1)-(5). There is no evidence of reduced sustainability of new destinations and products introduced in the treatment year, be it because of experimentation or a ‗windfall effect‘ inducing excessive risk-taking. There is improved sustainability of new destinations up to three years forward for treated firms, compared to control firms. Table C.1 Survival of New Destinations and Products Duration TY to TY+1 TY to TY+2 TY to TY+3 TY to TY+4 TY to TY+5 Estimator Weighted Tobit Weighted Tobit Weighted Tobit Weighted Tobit Weighted Tobit (1) (2) (3) (4) (5) Outcome New destination survival rate 0.261 0.262 0.259 0.192 0.359 (2.83)*** (2.76)*** (2.28)** (1.43) (2.37)** R-squared 0.019 0.036 0.041 0.067 0.076 Observations 4,046 3,342 2,578 1,738 956 New HS 6d product survival rate 0.034 0.033 -0.032 0.056 0.135 (0.52) (0.44) (-0.33) (0.42) (0.90) R-squared 0.030 0.040 0.046 0.043 0.059 Observations 5,553 4,569 3,538 2,390 1,278 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%; *** significant at 1%. The sample includes treated firms and control firms in the common support. The dependent variable is a survival rate. The weighted Tobit regressions include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. 35 Appendix D: Robustness of Main Results and Spillover Results Table D.1 Effects of FAMEX on Export Outcomes – Alternatives for Dropouts Panel A. Excluding Dropouts from the Sample Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) Outcome Total exports 0.548*** 0.743*** 0.599*** 0.269 0.051 0.279 [3.26] [3.65] [2.65] [0.99] [0.14] [0.69] R-squared 0.17 0.22 0.22 0.22 0.22 0.24 Nb. destinations 0.154*** 0.191*** 0.192*** 0.145*** 0.143*** 0.183*** [6.20] [6.88] [5.85] [3.93] [2.93] [3.22] R-squared 0.15 0.19 0.20 0.23 0.29 0.31 Nb. products 0.150*** 0.180*** 0.185*** 0.120** 0.167*** 0.238*** [4.74] [4.77] [4.51] [2.52] [2.72] [3.48] R-squared 0.15 0.20 0.23 0.26 0.27 0.30 Observations 11,645 11,506 9,172 6,872 4,598 2,405 Panel B. Including Dropouts in the Treated Group Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. DID reg. DID reg. DID reg. DID reg. (1) (2) (3) (4) (5) (6) Outcome Total exports 0.522*** 0.609*** 0.370* 0.239 -0.060 0.244 [3.32] [3.23] [1.76] [0.95] [-0.18] [0.65] R-squared 0.17 0.22 0.22 0.22 0.23 0.26 Nb. destinations 0.139*** 0.148*** 0.153*** 0.116*** 0.119*** 0.167*** [6.13] [5.78] [5.19] [3.45] [2.60] [3.18] R-squared 0.14 0.18 0.20 0.23 0.29 0.31 Nb. products 0.116*** 0.145*** 0.143*** 0.105** 0.132** 0.232*** [4.18] [4.44] [3.92] [2.49] [2.31] [3.64] R-squared 0.14 0.18 0.23 0.26 0.28 0.32 Observations 11,950 11,735 9,327 6,970 4,653 2,453 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%; *** significant at 1%. TY is the first year when FAMEX support is received. The sample includes treated and control firms in the common support. The WLS regressions include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. In Panel A the sample includes 401 treated firms and 2220 control firms while in Panel B the sample includes 526 treated firms and 2220 control firms. 36 Table D.2 FAMEX Effects for Firms Treated in 2005 Panel A. Year-to-Year Effects of FAMEX on Export Outcomes Difference TY-(TY-1) TY-(TY-1) (TY+1)-TY (TY+2)-(TY+1) (TY+3)-(TY+2) (TY+4)-(TY+3) (TY+5)-(TY+4) Estimator PSM-DID WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1a) (1b) (2) (3) (4) (5) (6) Outcome Total exports 0.425* 0.485** 0.387* -0.007 -0.156 -0.308 0.058 [1.67] [2.32] [1.89] [-0.03] [-0.73] [-1.33] [0.25] R-squared 0.16 0.15 0.11 0.09 0.11 0.11 Nb. destinations 0.148*** 0.156*** 0.094*** 0.061* 0.023 0.042 0.061** [4.30] [4.81] [3.00] [1.76] [0.68] [1.18] [2.13] R-squared 0.14 0.12 0.08 0.13 0.14 0.07 Nb. products 0.162*** 0.156*** 0.073* 0.061 0.006 0.065 0.099*** [3.56] [3.72] [1.65] [1.45] [0.14] [1.58] [2.62] R-squared 0.16 0.17 0.11 0.11 0.13 0.13 Observations 2,524 2,524 2,524 2,524 2,524 2,524 Panel B. Cumulative Effects of FAMEX on Export Outcomes Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) Outcome Total exports 0.485** 0.838*** 0.636** 0.386 0.023 0.200 [2.32] [3.40] [2.29] [1.22] [0.06] [0.52] R-squared 0.16 0.21 0.21 0.21 0.22 0.25 Nb. destinations 0.156*** 0.208*** 0.214*** 0.175*** 0.157*** 0.177*** [4.81] [5.69] [4.64] [3.54] [3.07] [3.22] R-squared 0.14 0.19 0.20 0.26 0.30 0.30 Nb. products 0.156*** 0.190*** 0.200*** 0.158*** 0.177*** 0.219*** [3.72] [3.95] [3.66] [2.61] [2.75] [3.37] R-squared 0.16 0.21 0.22 0.24 0.26 0.30 Observations 2,524 2,524 2,524 2,524 2,524 2,524 Notes: T-statistics based on robust standard errors in brackets; *: significant at 10%; **: significant at 5%; ***: significant at 1%. The sample includes treated firms in 2005 and control firms in the common support that are present in the sample in every year from 2004 to 2010. The PSM-DID estimates are estimated based on propensity scores obtained using kernel matching. The WLS regressions include firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. 37 Table D.3 Effects of FAMEX on Export Outcomes using Year-by-Year Matching Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator DID reg. DID reg. DID reg. DID reg. DID reg. DID reg. (1) (2) (3) (4) (5) (6) Outcome Total exports 0.503*** 0.689*** 0.486** 0.139 -0.25 -0.148 [2.95] [3.18] (2.10)** [0.51] [-0.71] [-0.37] Nb. destinations 0.135*** 0.161*** 0.147*** 0.100*** 0.076 0.100* [5.35] [5.64] [4.51] [2.69] [1.54] [1.81] Nb. products 0.139*** 0.155*** 0.137*** 0.058 0.094 0.130** [4.32] [3.97] [3.27] [1.17] [1.56] [1.99] Observations 802 802 798 716 560 516 Treated 401 401 399 359 280 258 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%; *** significant at 1%. TY is the first year when FAMEX support is received. The DID regressions include treatment year effects. Table D.4 FAMEX Effects on the Number of HS 6-Digit Products Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) Outcome Nb. Products HS 6-digit 0.132*** 0.165*** 0.168*** 0.109** 0.140** 0.208*** [4.44] [4.75] [4.42] [2.50] [2.54] [3.43] R-squared 0.144 0.201 0.232 0.270 0.274 0.309 Observations 12,263 12,124 9,664 7,238 4,839 2,524 Notes: T-statistics based on robust standard errors in brackets; *: significant at 10%; **: significant at 5%; ***: significant at 1%. The sample includes treated and control firms in the common support. The WLS regressions include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. 38 Table D.5 Further Effects of Exposure to FAMEX Firms on Export Outcomes Panel A. Spillovers for Sample of All Firms Estimator Within reg. Within reg. Within reg. Difference t-(t-1) t-(t-1) t-(t-1) Outcome Total exports Nb. destinations Nb. products Sample of all firms Sample of all firms Sample of all firms (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Exposure to FAMEX benef. t-1 -0.050* -0.049* -0.034 -0.114 -0.003 -0.004 0.001 -0.001 -0.004 -0.005 -0.003 -0.018* [-1.90] [-1.76] [-0.92] [-1.52] [-1.14] [-1.22] [0.16] [-0.18] [-1.16] [-1.18] [-0.65] [-1.77] Exposure to FAMEX benef. t-2 0.004 0.017 -0.041 -0.001 0.003 -0.007 -0.001 0.000 -0.024* [0.13] [0.43] [-0.44] [-0.34] [0.74] [-0.79] [-0.18] [0.07] [-1.89] Exposure to FAMEX benef. t-3 -0.004 -0.055 0.001 -0.010 -0.000 -0.023* [-0.12] [-0.62] [0.21] [-1.15] [-0.01] [-1.90] Exposure to FAMEX benef. t-4 -0.072 -0.011 -0.023** [-1.02] [-1.51] [-2.39] Number of firms 3,024 3,024 3,022 3,022 3,024 3,024 3,022 3,022 3,024 3,024 3,022 3,022 Firm fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Location-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.02 0.02 0.02 0.02 0.01 0.01 0.01 0.01 0.02 0.02 0.02 0.02 Observations 14,786 14,786 11,926 9,013 14,786 14,786 11,926 9,013 14,786 14,786 11,926 9,013 Panel B. Spillovers from FAMEX Firms Looking to Increase Export Destinations and Exported Products Estimator Within reg. Within reg. Within reg. Difference t-(t-1) t-(t-1) t-(t-1) Outcome Total exports Nb. destinations Nb. products Sample of control firms only Sample of control firms only Sample of control firms only (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Exposure to FAMEX benef. t-1 -0.052* -0.050 -0.016 -0.122 -0.003 -0.004 0.004 -0.000 -0.006 -0.006 -0.000 -0.022* [-1.79] [-1.64] [-0.39] [-1.39] [-1.04] [-1.27] [0.87] [-0.03] [-1.49] [-1.56] [-0.03] [-1.95] Exposure to FAMEX benef. t-2 0.004 0.037 -0.019 -0.002 0.005 -0.005 -0.001 0.005 -0.020 [0.14] [0.85] [-0.18] [-0.75] [1.25] [-0.47] [-0.33] [0.83] [-1.44] Exposure to FAMEX benef. t-3 0.012 -0.028 0.005 -0.004 0.006 -0.015 [0.31] [-0.28] [1.39] [-0.43] [1.12] [-1.14] Exposure to FAMEX benef. t-4 -0.060 -0.008 -0.022** [-0.76] [-1.11] [-2.05] Number of firms 2,620 2,620 2,618 2,618 2,620 2,620 2,618 2,618 2,620 2,620 2,618 2,618 Firm fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Sector-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Location-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes R-squared 0.02 0.02 0.02 0.02 0.01 0.01 0.01 0.01 0.02 0.02 0.02 0.02 Observations 12,785 12,785 10,316 7,802 12,785 12,785 10,316 7,802 12,785 12,785 10,316 7,802 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%; *** significant at 1%. In Panel A, the sample includes all FAMEX firms and control firms in the common support or outside. In Panel B, the sample includes only control firms inside the common support or outside that are in sector- location cells where FAMEX firms that required assistance with the objective of increasing their destinations or their exported products were present. 39 Appendix E: Extensions Table E.1 Effect of Treatment Interacted with Project Objective Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) Outcome Objective More substantive exporter 0.467 0.898* 0.591 -0.028 -1.650* -0.669 [1.15] [1.68] [1.11] [-0.04] [-1.94] [-0.76] New destinations 0.563** 0.515** 0.215 0.109 0.154 0.314 Total exports [2.48] [2.01] [0.70] [0.31] [0.36] [0.66] New products 0.184 0.650** 0.780*** 0.946** 0.851 0.596 [0.78] [2.29] [2.62] [2.42] [1.60] [0.94] R-squared 0.17 0.22 0.23 0.23 0.24 0.25 Outcome Objective More substantive exporter 0.144*** 0.190*** 0.178** 0.104 -0.065 0.068 [2.66] [2.79] [2.50] [1.34] [-0.62] [0.60] Nb. New destinations 0.171*** 0.190*** 0.176*** 0.183*** 0.173*** 0.223*** destinations [4.95] [5.20] [4.01] [3.74] [2.89] [3.26] New products 0.085** 0.166*** 0.169*** 0.104* 0.194*** 0.147* [2.42] [3.97] [3.48] [1.87] [2.71] [1.78] R-squared 0.16 0.20 0.20 0.24 0.30 0.31 Outcome Objective More substantive exporter 0.130* 0.227*** 0.256*** 0.119 -0.004 0.082 [1.89] [2.62] [2.87] [1.23] [-0.04] [0.67] Nb. products New destinations 0.156*** 0.148*** 0.093* 0.108* 0.164** 0.221*** [3.43] [3.08] [1.71] [1.74] [2.31] [2.84] New products 0.082* 0.148** 0.212*** 0.158** 0.215** 0.278** [1.72] [2.35] [3.29] [2.05] [2.06] [2.35] R-squared 0.15 0.20 0.23 0.26 0.27 0.30 Observations 12,263 12,124 9,664 7,238 4,839 2,524 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%; *** significant at 1%. TY is the first year when FAMEX support is received. The sample includes treated and control firms in the common support. The WLS regressions include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. 40 Table E.2 Effects of FAMEX Program Components Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) Outcome Activity (amounts in TND) Market prospection 0.039** 0.048* 0.072** 0.101*** 0.018 0.027 [2.03] [1.86] [2.49] [2.98] [0.36] [0.53] Promotion 0.028*** 0.039*** 0.029** 0.005 0.027 0.001 [3.06] [3.04] [2.20] [0.31] [1.15] [0.03] Product development -0.014 -0.013 0.003 -0.045 -0.023 -0.023 Total exports [-0.96] [-0.55] [0.18] [-1.41] [-0.67] [-0.64] Firm development -0.022 0.004 0.000 0.048* 0.087*** 0.101** [-1.12] [0.19] [0.02] [1.83] [2.80] [2.33] Foreign subs. creation -0.003 0.026* -0.019 -0.008 -0.047 -0.063* [-0.15] [1.91] [-0.64] [-0.29] [-1.59] [-1.87] R-squared 0.21 0.26 0.26 0.25 0.25 0.26 Outcome Activity (amounts in TND) Market prospection 0.007** 0.013*** 0.013*** 0.009* 0.010 0.008 [2.07] [3.29] [2.95] [1.72] [1.56] [1.08] Promotion 0.006*** 0.006*** 0.011*** 0.006** 0.002 0.004 [3.20] [3.57] [4.70] [2.34] [0.47] [0.79] Nb. Product development 0.000 -0.001 0.003 -0.003 -0.003 -0.002 destinations [0.06] [-0.28] [0.98] [-0.70] [-0.69] [-0.41] Firm development 0.001 0.004 0.007 0.009 0.017*** 0.020*** [0.27] [1.02] [1.39] [1.49] [3.09] [2.73] Foreign subs. creation 0.000 0.001 -0.004 -0.004 -0.001 -0.009 [0.00] [0.68] [-1.05] [-1.23] [-0.30] [-1.58] R-squared 0.17 0.21 0.24 0.25 0.31 0.32 Outcome Activity (amounts in TND) Market prospection 0.009** 0.009* 0.015*** 0.014** 0.012 0.012 [2.05] [1.78] [2.99] [2.32] [1.40] [1.36] Promotion 0.004 0.006** 0.004 0.002 0.002 0.000 [1.11] [2.02] [1.44] [0.54] [0.53] [0.01] Product development -0.003 0.004 0.004 -0.003 0.001 0.006 Nb. products [-1.39] [0.90] [1.09] [-0.53] [0.20] [0.99] Firm development -0.002 0.000 0.003 0.014* 0.025*** 0.024*** [-0.41] [0.07] [0.59] [1.84] [3.34] [2.59] Foreign subs. creation 0.002 0.004 -0.003 -0.006 -0.005 -0.009 [0.49] [1.13] [-0.75] [-1.31] [-0.79] [-1.27] R-squared 0.16 0.22 0.26 0.27 0.29 0.31 Observations 12,157 12,018 9,590 7,188 4,808 2,496 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%; *** significant at 1%. TY is the first year when FAMEX support is received. The sample includes treated and control firms in the common support. The WLS regressions include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. 41 Table E.3 FAMEX Effects, Firm Size and In-House Export Unit Panel A. Firm Size Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) Outcome Firm size FAMEX*Less than 50 workers 0.512* 1.015*** 0.673* 0.398 0.190 0.666 Total [1.84] [3.15] [1.91] [0.99] [0.37] [1.18] exports FAMEX*More than 50 workers 0.511*** 0.444* 0.474* 0.149 -0.074 -0.138 [2.82] [1.83] [1.74] [0.42] [-0.16] [-0.26] R-squared 0.17 0.23 0.23 0.22 0.23 0.25 Outcome Firm size FAMEX*Less than 50 workers 0.151*** 0.223*** 0.182*** 0.149*** 0.099 0.166** Nb. [3.98] [5.41] [3.80] [2.95] [1.54] [2.32] destinations FAMEX*More than 50 workers 0.149*** 0.161*** 0.197*** 0.154*** 0.178*** 0.184** [4.69] [4.40] [4.59] [2.99] [2.63] [2.35] R-squared 0.15 0.20 0.20 0.24 0.29 0.30 Outcome Firm size FAMEX*Less than 50 workers 0.179*** 0.219*** 0.215*** 0.140** 0.153* 0.172** Nb. [3.58] [3.91] [3.57] [2.13] [1.84] [1.98] products FAMEX*More than 50 workers 0.115*** 0.133*** 0.143*** 0.094 0.159* 0.253*** [2.97] [2.68] [2.66] [1.43] [1.95] [2.80] R-squared 0.15 0.20 0.23 0.26 0.27 0.30 Observations 12,263 12,124 9,664 7,238 4,839 2,524 Panel B. In-House Export Unit Difference TY-(TY-1) (TY+1)-(TY-1) (TY+2)-(TY-1) (TY+3)-(TY-1) (TY+4)-(TY-1) (TY+5)-(TY-1) Estimator WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. WLS reg. (1) (2) (3) (4) (5) (6) Outcome Export Unit Status FAMEX*Has in-house export unit 0.532*** 0.893*** 0.767*** 0.553 0.097 0.168 Total [2.94] [3.96] [2.97] [1.56] [0.21] [0.33] exports FAMEX*No in-house export unit 0.491* 0.554* 0.377 -0.003 -0.016 0.235 [1.93] [1.77] [1.12] [-0.01] [-0.03] [0.45] R-squared 0.17 0.23 0.23 0.22 0.23 0.25 Outcome Objective FAMEX*Has in-house export unit 0.162*** 0.225*** 0.248*** 0.208*** 0.197*** 0.222*** Nb. [4.96] [6.32] [6.00] [4.13] [3.02] [2.97] destinations FAMEX*No in-house export unit 0.138*** 0.158*** 0.132*** 0.096** 0.083 0.125* [3.95] [4.00] [2.93] [2.01] [1.41] [1.85] R-squared 0.15 0.20 0.21 0.24 0.29 0.31 Outcome Objective FAMEX*Has in-house export unit 0.134*** 0.200*** 0.192*** 0.132** 0.211*** 0.288*** Nb. [3.42] [4.11] [3.71] [2.08] [2.58] [3.20] products FAMEX*No in-house export unit 0.159*** 0.150*** 0.163*** 0.103 0.097 0.141* [3.47] [2.86] [2.89] [1.63] [1.31] [1.75] R-squared 0.15 0.20 0.23 0.26 0.27 0.30 Observations 12,263 12,124 9,664 7,238 4,839 2,524 Notes: T-statistics based on robust standard errors in brackets;* significant at 10%; ** significant at 5%; *** significant at 1%. TY is the first year when FAMEX support is received. The sample includes treated and control firms in the common support. The WLS regressions include treatment year fixed effects, firm age and age squared, location and sector fixed effects, a dummy variable for firm size (based on employment), a dummy variable for firms exporting 100% of their output, lagged exports, lagged number of destinations served, and lagged number of exported products. 42 Appendix F: Cost-Benefit Analysis of FAMEX Table F.1 Rates of Return on the FAMEX Program Baseline TY TY+1 TY+2 TY+3 TY+4 TY+5 (BL) Non-significant coefficients 1 2 3 4 5 6 A β Coefficient 0.511 0.723 0.571 0.272 0.043 0.200 B = exp(A)-1 Change in total export growth (treatment effect) 0.667 1.061 0.770 0.313 0.044 0.221 C Cumulative total export growth, control a/ 0.084 0.174 0.272 0.378 0.493 0.618 D = C * (1+B) Predicted cumulative total export growth, treated 0.139 0.358 0.481 0.496 0.515 0.755 E = BL * (1+C) Total exports, control b/ 2,308 2,501 2,710 2,936 3,181 3,447 3,734 F = BL * (1+D) Total exports, treated b/ c/ 2,308 2,629 3,135 3,419 3,454 3,497 4,050 G=F-E Difference in total exports due to FAMEX b/ 129 426 483 273 50 316 H Average FAMEX grant per treated b/ 21.7 21.7 21.7 21.7 21.7 21.7 I = G/H Return on public investment d/ 5.9 19.6 22.3 12.6 2.3 14.6 Notes: a/ Obtained directly from the sample as the average annual export growth over the sample period for control firms; b/ Expressed in thousand of Tunisian Dinars; c/ Given the matching procedure, pre-treatment average total exports of treated firms are assumed to be similar to those of control firms; d/ In Tunisian Dinars of additional exports per firm per Tunisian dinar of publicly-funded FAMEX grant. 43