Policy Research Working Paper 9015 Socioeconomic Resilience in Sri Lanka Natural Disaster Poverty and Wellbeing Impact Assessment Brian Walsh Stephane Hallegatte Climate Change Group September 2019 Policy Research Working Paper 9015 Abstract Traditional risk assessments use asset losses as the main systems. Such investments efficiently reduce wellbeing metric to measure the severity of a disaster. This paper losses by making exposed and vulnerable populations more proposes an expanded risk assessment based on a frame- resilient. Nationally and on average, the bottom income work that adds socioeconomic resilience and uses wellbeing quintile suffers only 7 percent of the total asset losses but losses as the main measure of disaster severity. Using an 32 percent of the total wellbeing losses. Average annual agent-based model that represents explicitly the recovery wellbeing losses due to fluvial flooding in Sri Lanka are esti- and reconstruction process at the household level, this risk mated at US$119 million per year, more than double the assessment provides new insights into disaster risks in Sri asset losses of US$78 million. Asset losses are reported to Lanka. The analysis indicates that regular flooding events be highly concentrated in Colombo district, and wellbeing can move tens of thousands of Sri Lankans into transient losses are more widely distributed throughout the coun- poverty at once, hindering the country’s recent progress try. Finally, the paper applies the socioeconomic resilience on poverty eradication and shared prosperity. As metrics framework to a cost-benefit analysis of prospective adaptive of disaster impacts, poverty incidence and well-being losses social protection systems, based on enrollment in Samurdhi, facilitate quantification of the benefits of interventions like the main social support system in Sri Lanka. rapid post-disaster support and adaptive social protection This paper is a product of the Climate Change 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://www.worldbank.org/prwp. The authors may be contacted at bwalsh1@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 SOCIOECONOMIC RESILIENCE IN SRI LANKA Natural disaster poverty & wellbeing impact assessment & Keywords: natural risks, resilience, risk assessment, welfare, Sri Lanka, social protection, cost benefit analysis JEL: D15, D30, D63, D78, Q54, R11 Contents 2 1 Introduction 5 2 Sri Lanka country risk profile 9 2.1 Risk to assets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.2 Income and consumption poverty . . . . . . . . . . . . . . . . . . . . . . 12 2.3 Risk to wellbeing at the national level . . . . . . . . . . . . . . . . . . . . 17 3 Socioeconomic resilience to disasters 18 3.1 District-level risk summary . . . . . . . . . . . . . . . . . . . . . . . . . . 21 3.2 Poorest quintile risk summary . . . . . . . . . . . . . . . . . . . . . . . . 23 4 Adaptive social protection 25 4.1 Proxy means test (PMT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 4.2 Case study: 50-year flood in Rathnapura . . . . . . . . . . . . . . . . . . 27 4.3 Post-disaster support to Samurdhi enrollees . . . . . . . . . . . . . . . . 29 4.4 Post-disaster support to PMT-qualified households . . . . . . . . . . . . 32 4.5 Nationwide cost-benefit analysis of Adaptive Social Protection . . . . . 32 5 Conclusions 33 6 Technical Appendix — Methodology and model description 37 Figure 1 Traditional DRM heuristic . . . . . . . . . . . . . . . . . . . . . . 8 Figure 2 District-level asset risk in Sri Lanka. . . . . . . . . . . . . . . . . 11 Figure 3 Impact of a disaster on per capita income and consumption . . 13 Figure 4 Time to recover 90% of assets destroyed in 100-year precipita- tion flood. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 Figure 5 Socioeconomic resilience to precipitation flooding in Sri Lanka . 19 Figure 6 District-level wellbeing risk in Sri Lanka . . . . . . . . . . . . . . 22 Figure 7 Asset and wellbeing losses in Rathnapura flood . . . . . . . . . 27 Figure 8 Asset and wellbeing losses, before and after Samurdhi scaleup . 28 List of Tables 3 Figure 9 Cost-benefit analysis of Samurdhi scale-up . . . . . . . . . . . . 29 Figure 10 Cost-benefit analysis of scale-up versus scale-out . . . . . . . . . 30 Figure 11 Cost-benefit analysis of post-disaster support scale-out using PMT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Figure 12 Avoided wellbeing losses for various ASP programs . . . . . . . 33 Figure 13 Household asset vulnerability (vh ) . . . . . . . . . . . . . . . . . 47 Figure 14 Household asset losses and reconstruction . . . . . . . . . . . . . 48 Figure 15 Household consumption losses during reconstruction . . . . . . 52 Figure 16 Household reconstruction time (τh ) . . . . . . . . . . . . . . . . . 56 Table 1 Asset risk from precipitation flooding, by district . . . . . . . . . 10 Table 2 Risk to assets, socioeconomic resilience, and risk to wellbeing . 21 Table 3 Risk to assets, socioeconomic resilience, and risk to wellbeing for the poorest quintile . . . . . . . . . . . . . . . . . . . . . . . . 23 Table 4 Per capita risk to assets and wellbeing for the poorest quintile . 24 Table 5 Total household expenditures, by district, as reported by the 2016 HIES. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 Table 6 Asset vulnerability categories . . . . . . . . . . . . . . . . . . . . 46 Traditional risk assessments use asset losses as the main metric to measure the sever- ity of a disaster. This paper proposes an expanded risk assessment based on a frame- work that adds socioeconomic resilience and uses wellbeing losses as the main mea- sure of disaster severity. Using an agent-based model that represents explicitly the recovery and reconstruction process at the household level, this risk assessment pro- vides new insights into disaster risks in Sri Lanka. The analysis indicates that regular List of Tables 4 flooding events can move tens of thousands of Sri Lankans into transient poverty at once, hindering the country’s recent progress on poverty eradication and shared prosperity. As metrics of disaster impacts, poverty incidence and well-being losses fa- cilitate quantification of the benefits of interventions like rapid post-disaster support and adaptive social protection systems. Such investments efficiently reduce wellbe- ing losses by making exposed and vulnerable populations more resilient. Nationally and on average, the bottom income quintile suffers only 7 percent of the total as- set losses but 32 percent of the total wellbeing losses. Average annual wellbeing losses due to fluvial flooding in Sri Lanka are estimated at US$119 million per year, more than double the asset losses of US$78 million. Asset losses are reported to be highly concentrated in Colombo district, and wellbeing losses are more widely distributed throughout the country. Finally, the paper applies the socioeconomic re- silience framework to a cost-benefit analysis of prospective adaptive social protection systems, based on enrollment in Samurdhi, the main social support system in Sri Lanka. 5 Sri Lanka has enjoyed rapid progress on poverty reduction and shared prosperity since the cessation of the decades-long conflict in 2009. In the last 10 years, GDP per capita has doubled from US$2,000 to US$4,000 (current US), and the national poverty incidence has been halved, from 9% to 4%, as measured by the national household income & expenditures survey. Despite this progress, the country scored modestly on several indicators related to disaster risk management in the 2015 Country Policy and Institutional Assessment, including quality of public administration; fiscal policy and macroeconomic management; social protection; and transparency, accountability, and corruption in the public sector. Given this context, effective disaster risk management (DRM) strategies are essen- tial to the continuation of peace and prosperity in Sri Lanka. Parts of the country are regularly affected by destructive floods, particularly during regional monsoon seasons. In 2017, for example, high winds, severe flooding, and landslides struck the southwestern districts of Sri Lanka–most directly, Galle, Kalutara, Matara, and Rathnapura–causing approximately 50-60 billion LKR (333-400 million USD) in dam- ages, as estimated in the Ministry of Finance (2017) Annual Report. According to the World Bank post disaster needs assessment, the government disbursed US$6.6 million in emergency relief, and total recovery needs were later estimated at nearly US$800 million. Although the 2017 floods were exceptional, 2016 also saw unusually destructive flooding, and such events could become more frequent in the future as a result of climate change. Asset losses are routinely used to connote disasters’ scope, but they provide only a partial measure of the total cost of any event. Other loss dimensions include lost wages and other income, insurance liabilities, days of school and doctor visits skipped, and meals foregone. For example: imagine that two identical, neighboring households experience the same property damage from a flood. If only one of the households is insured, that one will be compensated for its losses, while the other one might relocate, pull its children out of school, or reduce its consumption below the poverty line in order to recover. There are many different types of disaster costs, 6 and their magnitude, distribution, and urgency depends on the socioeconomic status of affected households. Asset losses measure a dimension of disaster costs that accrues mostly to the wealthy. By definition, wealthy individuals have the most assets to lose. Therefore, asset losses often define–and accurately summarize–their experience of a shock. Con- versely, the poor definitionally have few assets to lose, so damage to their collective property is a small fraction of aggregate asset losses. However, the poor also lack the resources and instruments to smooth income shocks while maintaining their con- sumption, or to recover their asset stock. Therefore, their experience of shocks is better described by other loss metrics. After most shocks, the poor are more likely than the wealthy to forego consumption of food, health, or education, and to take longer to recover. Disaster risk management strategies that seek to maximize impact in terms of avoided asset losses will almost inevitably favor central business districts and other areas with high concentrations of valuable assets, leaving the poor to resort to hu- manitarian assistance when disasters inevitably occur. By highlighting these areas, asset losses disincentivize investments in slums and informal settlements, even when small investments could significantly reduce disease transmission, absenteeism from work and school, transportation costs, lost wages, or other types of disaster costs. Finally, asset losses obscure the relationship between disasters and development in Sri Lanka. The long-term impacts of disasters on communities’ wellbeing and prospects depends not only on direct or immediate impacts (asset losses), but also on the duration of the recovery and reconstruction period and on the tools affected populations have to cope with and recover from the shock, such as social transfers, formal and informal post-disaster support, savings, insurance, and access to credit. Households that lack access to these tools will struggle to cope with shocks, and some could fall into chronic poverty as a result. To provide a more comprehensive measure of disaster impacts, the Unbreakable re- port introduced the concept of wellbeing losses. Wellbeing losses incorporate people’s socio-economic resilience, including (1) their ability to maintain their consumption for the duration of their recovery, (2) their ability to save or borrow to rebuild their asset 7 stock, and (3) the decreasing returns in consumption–that is, the fact that people who live on $2 per day are more affected by a $1 loss than are richer individuals. The analysis presented in this paper applies this approach to floods in Sri Lanka. It combines flood maps from a global model developed by SSBN (now Fathom) with the Sri Lanka Household Income & Expenditures Survey (HIES) to model flood im- pacts at the household level, using an agent-based model to represent the recovery dynamics [1]. The full methodology and detailed model description is provided in the technical appendix to this document, and in [2]. The analysis provides information to better target DRM interventions both spa- tially (where should we act?) and sectorally (which type of intervention is the most efficient?). Compared with previous work in Sri Lanka, the analysis innovates by measuring disaster risk with several metrics that complement asset losses (or the cost to repair or replace damaged buildings, infrastructure, productive equipment, etc.). These metrics include (1) traditional asset losses, which highlight the experience of the wealthy; (2) poverty-related measures such as poverty headcount, which high- light the experience of the poor and near-poor; (3) wellbeing losses, which provide a balanced estimate of the impact on poor and rich households; and (4) socioeconomic resilience, an indicator that measures the ability of the population to cope with and re- cover from asset losses. This broad perspective is intended to complement traditional, more spatially-detailed risk assessments in Sri Lanka. One major conclusion of this work is that alternative and new metrics of disas- ter impacts–including poverty headcount, poverty gap, and wellbeing losses–can be used to quantify the value of interventions currently outside the traditional risk- management toolbox. Asset-focused risk-management strategies focus primarily on protection infrastructure, such as dikes, and the position and condition of assets, for instance with land-use plans and building codes. Wellbeing-focused strategies can utilize a wider set of available measures, such as financial inclusion, private and public insurance, disaster-responsive social safety nets, macro-fiscal policies, and dis- aster preparedness and contingent planning. Even if they do not reduce asset losses, these types of measures can bolster communities’ socioeconomic resilience, or their capacity to cope with and recover from asset losses when they occur, and reduce the wellbeing impact of natural disasters. 8 Figure 1: Traditional risk assessments evaluate physical hazard and asset exposure and vul- nerability to measure expected asset losses and inform DRM strategies. The Un- breakable model additionally incorporates the socio-economic resilience of the com- munities to predict wellbeing losses. Following the general spatial and sectoral analysis, we show how this analytical framework can be used to assess the benefits of cash transfers used to provide post- disaster support (PDS) to affected populations. This is an example of a resilience- building measure that helps households to smooth consumption, but does not affect asset losses. Tools like cash transfers can help disaster-affected households to recover, but they have not been traditionally included in DRM analysis because their benefits are difficult to measure. Here, we look at the potential of the Samurdhi program to act as an “adaptive social protection” scheme able to provide emergency support to households after floods. We find that the program can be a useful tool for responding to disasters and helping Sri Lankans to recover when they are affected–even assuming no changes to the enrollment of the program—and propose options to improve its effectiveness. 9 . Risk to assets Typically, risk assessments incorporate information on the hazards (the natural occur- rence of destructive events such as severe winds, surges, floods, and earthquakes); exposure (the value of natural and built assets that might face a destructive event); and vulnerability (the expected consequences to exposed assets when a destructive event occurs) of the targeted area. Together, these three dimensions describe average annual asset losses in the area of interest. Table 1 on the following page displays expected annual asset losses to precipitation flooding, by district. These values, based on the SSBN model, are long-term averages of the total replacement cost of household income-generating assets damaged or destroyed by precipitation flooding each year [1].1 Figure 2 maps total asset risk, expressed at left in US$ and at right as a percentage of district-level aggregated household expen- ditures (AHE). According to this model, asset risk in Sri Lanka is highly concentrated in greater Colombo (over 50% of the total; left panel), but other districts throughout the country also face significant losses, when expressed in percentage of local AHE.2 Significant differences across districts are explained by variations in the hazard, exposure, and vulnerability of each part of the country. Most notably, asset losses are heavily concentrated in greater Colombo and surrounding districts. This is to be expected, as it reflects the high concentration of valuable assets in and around the capital (elevated exposure). Differences in hazards are secondary but still important: 1 The total repair and replacement value of assets damaged or destroyed by floods in each district is a direct output of the SSBN/Fathom model [1], but this analysis is particularly interested in disaster impacts on the assets, consumption, and welfare of households. One issue faced here is the difference between the Gross Domestic Income (GDI) derived from national accounts and the aggregate house- hold expenditures (AHE) calculated from household surveys (in this case, HIES). As is well known, the latter tend to report lower expenditures than national accounts [3]. In this analysis, we work based on the AHE. Therefore, the expected losses in Table 1 have been reduced by a scale factor, equal to total household consumption over nominal regional productivity (cf. Tab. 5), and express average annual damage or destruction to assets used by Sri Lankan households to generate income, by dis- trict. Asset losses in Jaffna, Kilinochchi, and Matara are omitted due to their exclusion from available SSBN/Fathom flood maps. 2 SSBN is a global model subject to limited availability and variable quality data. Therefore, its precision and accuracy can be low in some places, and we would seek ideally to replace these inputs with high- resolution local models. 10 Risk to assets from precipitation flooding District mLKR mUS$ Colombo 6,040 40.2 Rathnapura 1,230 8.19 Kalutara 1,120 7.43 Gampaha 1,040 6.93 Puttalam 498 3.32 Kurunegala 363 2.42 Polonnaruwa 315 2.10 Hambantota 233 1.55 Trincomalee 189 1.26 Kegalle 151 1.00 Galle 101 0.67 Batticaloa 89 0.59 Mannar 79 0.53 Anuradhapura 71 0.47 Kandy 64 0.43 Ampara 46 0.30 Badulla 44 0.29 Matale 26 0.17 Vavuniya 7 0.04 Moneragala 6 0.04 Mullaitivu 4 0.03 Nuwara Eliya 0 <0.01 Jaffna – – Kilinochchi – – Matara – – Total 11,700 78.02 Table 1: Expected annual asset losses (tabulated in millions of LKR and thousands of US$) from exposure to precipitation floods in Sri Lanka. (US$1 = LKR 150). Asset loss exceedance curves were not available from the Fathom model for Jaffna, Kilinochchi, or Matara. 11 while severe flooding can strike any part of the country, the meteorological hazard is greatest in the southwestern and northeastern quadrants of the country, due to their respective monsoon climates. Figure 2: Total risk to assets (expected annual losses) from exposure to precipitation floods in Sri Lanka. At left, annual expected asset losses are expressed in US dollars (US$1 = LKR150). At right, losses are shown in dollars per capita and per year. The simple comparison in Figure 2 illustrates an essential point: different metrics can lead to different disaster risk “hotspots", or investment priorities. At a minimum, this suggests that asset risk does not give a complete picture of disaster impacts in Sri Lanka. Both maps in Figure 2 describe mostly what happens to the wealthiest regions–and wealthiest people in these regions–because they represent aggregate as- set loss, and wealthy regions and people have the most to lose. 12 Clearly, a more disaggregated approach will be required. To do so, the rest of this analysis goes deeper in the distributional analysis, moving from the regional scale to the household level. In the next section, we merge regional asset loss data with household-level socioeconomic characteristics to examine disaster impacts on household income and consumption. This novel approach will allow us to develop estimates of the number of individuals pushed into transient poverty each year by disasters. Subsequently, we study how exposure to flooding affects households’ con- sumption and wellbeing and use these insights to assess new policy tools for manag- ing the risk of the poor and vulnerable. . Income and consumption poverty The disaggregation of asset losses to the household level allows us to move from expected asset losses to the secondary effects, including income and consumption losses, that must be expected when households are affected by shocks. From this perspective, we begin to see disaster losses not just in terms of economic costs, but as an obstacle to poverty reduction and other development imperatives. The technical details of asset loss disaggregation to the household level are dis- cussed in the technical appendix to this report. In this section, we present the main insights generated by the union of the SSBN model results with the household in- come & expenditures survey (HIES), focusing on how individual households’ socioe- conomic characteristics can mitigate or magnify the impact of disasters. Income poverty When disasters damage or destroy the assets on which individuals rely for their liveli- hood — including not only their own shop or field, but also somebody else’s factory — affected households face income losses. Some may receive extraordinary public as- sistance or additional remittances, which supplement their regular income while they rebuild, while others will be left to their own devices. Net income losses incorporate these socioeconomic characteristics, which influence households’ recovery pathways. 13 50-year precipitation flood in Rathnapura 120 Subsistence line Increase of 7,500 (0.6% of regional pop.) in income subsistence 100 Poverty line Increase of 11,300 (1.0% of regional pop.) in income poverty 80 Population (,000) 60 Pre-disaster income (FIES data) Post-disaster income (modeled) 40 20 0 0 250 500 750 1000 1250 1500 1750 2000 Income (USD per person, per year) 120 Subsistence line Increase of 0 (0.0% of regional pop.) in consumption subsistence 100 Poverty line Increase of 11,000 (0.9% of regional pop.) in consumption poverty 80 Population (,000) 60 Pre-disaster consumption (FIES data) Post-disaster consumption (modeled) 40 20 0 0 250 500 750 1000 1250 1500 1750 2000 Consumption (USD per person, per year) Figure 3: Per capita income (top) and consumption (bottom) in Rathnapura district, before and immediately after 50-year precipitation flooding, in US$. Income losses take into account the lost productivity of disaster-affected assets, while consumption losses additionally include reconstruction costs, post-disaster support, and savings. Therefore, income losses describe better than asset losses the real impacts of disasters on Sri Lankans’ wellbeing and prospects. To return to a historical example: damages from the 2017 floods totaled roughly US$400 million in the 5 most directly-affected districts (Galle, Hambantota, Kalutara, Matara, and Rathnapura). In Rathnapura, total damages were estimated at US$51.6 million, more than 6 times greater than long-term average annual losses in the dis- trict (US$8.2 million, cf. Table 1). These losses correspond roughly with the 50-year 14 precipitation flood, meaning that there is a 2% chance of flooding damage in excess of this value in any given year. The top panel of Figure 3 on the previous page illustrates the expected impact of a 2017-like flood event on individual incomes in Rathnapura district. The black outline indicates the regional income distribution as reported in HIES, while the red histogram illustrates the expected income distribution immediately following a 50- year flood in the region. This distribution shows a mode around US$450 per person, per year, with 10% of the district living below the poverty line. The large bar on the right shows the number of people with incomes higher than US$2,000 per year; the poverty and subsistence lines are indicated by the dotted lines around US$300 and US$350 per year, respectively. Income losses resulting from this 2017-like event are expected to push some 11,300 individuals into income poverty in Rathnapura, and 5,700 below the subsistence income level, meaning they cannot meet basic needs. At the national level and for the average year, we estimate that over 5,600 Sri Lankans face income poverty or subsistence for some amount of time every year due to precipitation flooding. However, large events can impoverish broad segments of the population when they occur. For example, the expected destruction from nation- wide 50-year flooding could move nearly 34,000 individuals nationwide into transient income poverty or subsistence. Of these, the model indicates that just over a thou- sand will still be in poverty 10 years later. This suggests that disasters may have a long-term impact on the income and consumption of some Sri Lankan households. Consumption poverty Income losses represent a complementary perspective to asset losses, and generate new insights into the costs of disasters in Sri Lanka. However, as a metric of disaster impacts, income losses may yet underestimate the link between poverty and disas- ters. First, they do not account for the reconstruction costs that households incur in rebuilding their assets after a disaster. Second, while wealthy households may have access to considerable savings, credit, remittances, and insurance to finance this recovery, these resources are often available only informally, if at all, to poor house- holds [4, 5]. This is important, because access to these tools can mean the difference 15 between a speedy recovery and a long-term poverty trap [6, 7, 8]. These costs and resources all impact households’ ability to maintain consumption after a shock, or their consumption losses. Over the long term average, and inclusive of all districts for which flooding data are available, nearly 6, 000 Sri Lankans face transient consumption poverty due to precipitation flooding every year. This is equivalent to 0.5% of national poverty in- cidence, with significant variation across districts. In some districts, the number of individuals pushed into subsistence represents a significant fraction of the HIES inci- dence of subsistence. This result is most notable in greater Colombo, reaching 3.5% of the chronic poverty incidence there. These results suggest that precipitation flood- ing events may be major drivers of extreme poverty in certain districts of Sri Lanka. In these areas, disaster risk management strategies can be a highly efficient tool to decrease poverty and subsistence incidence, either by reducing asset losses, or by enhancing households’ abilities to recover from disasters.3 Socioeconomic characteristics and disaster recovery dynamics Asset losses (and the foregoing income poverty analysis) describe individual house- holds’ status in the instant after a hazard occurred, and provide useful insights into how governments and other first responders should target humanitarian relief. How- ever, another important question is whether disaster-affected households become mired in poverty or achieve a speedy recovery. In contrast to asset losses, income and consumption losses can inform policy makers on how to manage the duration of disaster impacts. Expanding our analysis of the 2017-like floods, Figure 4 on the following page maps the expected time to recover 90% of the assets destroyed when a 50-year precipitation flood strikes each district. The map indicates that coastal districts in the northern half of the country take the longest to recover. When disasters strike communities in these districts, many of which are national minorities, recovery is expected to proceed 3 Consumption losses are net of household savings, which can be used to smooth consumption after a shock, or to finance each household’s recovery from the initial event. In the absence of reliable data on household savings, we assume that each household has "precautionary" savings–i.e., liquid savings available for consumption– equivalent to one month’s worth of wages. 16 Figure 4: Time for households in each district to recover 90% of assets damaged or destroyed in 50-year precipitation flood. slowly. Disaster strategies seeking to minimize time to recover from disasters should focus on providing post-disaster support to these areas. Notably, this indicator of the ability to recover is found to be largely independent of the hazard exposure and risk level. It represents the capacity of each district to cope with losses, regardless of the likelihood of these losses. 17 . Risk to wellbeing at the national level The foregoing has shown that, even if disaster-affected households suffer identical asset losses, their consumption losses and recovery time will vary according to their socioeconomic status and the resources available to them (for instance because insur- ance, savings, remittances, and public support help some households to smooth the consumption shocks). Going further, $1 in consumption losses can have very different consequences for individual households, depending on their income. In particular, while rich households will be able to spend down their savings or cut on luxury consumption, poorer households will often have to cut on basic needs and essential consumption like food, health, or education, threatening their health, human capital, and long-term prospects. Disaster risk strategies and budgets should account for these differences, and well- being losses do precisely this. While $1 in asset or consumption losses affects a poor individual more than a rich one, wellbeing losses are defined such that a $1 wellbeing loss affects a rich and a poor individual equally. Wellbeing losses are calculated from consumption losses using a classical “welfare function.” This operation translates into wellbeing the value of a household’s consumption at each point in its unique re- covery, with decreasing returns to represent the fact that increasing consumption by $1 increases more the wellbeing of a poor individual (compared with a rich person).4 The difference in the wellbeing generated by $1 of consumption is a simple proxy for the continuum from survival consumption (the very first units of consumption that have the largest impact on wellbeing) to luxury consumption (which enhances wellbeing less and less). Wellbeing losses integrate each household’s consumption losses over the duration of its recovery, and give more weight to the consumption losses experienced by poor people than to the losses experienced by richer people. In this way, wellbeing losses account for socioeconomic differences among households, and correct for the pro- wealthy bias inherent in asset losses without relying on binary thresholds like the 4 Here, the unit of analysis is the household. It would be useful to do it at the individual level, to uncover intra-household distributional effects, linked to gender or age. In the absence of within- household data, we make the strong assumption that pre-disaster consumption and disaster losses are distributed equally per capita within each household. 18 poverty line. As a metric, they capture more fully the costs of disasters and the benefits of prospective DRM investments than do asset losses. Therefore, wellbeing- informed strategies are not merely more equitable, but more cost-effective than asset- informed strategies. As discussed in Section 2.1 on page 9, we estimate the replacement value of house- hold income-generating assets damaged or destroyed by precipitation flooding in Sri Lanka at US$78 million (cf. Table 2 on page 21), equivalent to 0.35% of AHE. Well- being losses are much higher, at US$122 million (0.53% of AHE) per year. In other words, the real impact of flooding in Sri Lanka is equivalent to a decrease in national consumption by US$122 million.5 The analysis has moved from asset to income, consumption, and wellbeing losses, in- corporating additional relevant household-level socioeconomic characteristics at each stage. To summarize these developments, we return to the traditional risk-assessment framework (cf. 1 on page 8). The three traditional components (i.e., hazard, exposure, and vulnerability) predict asset losses, but not the full impacts on people’s wellbeing. In response to this theoretical and practical shortcoming, we have effectively added a fourth component, called socioeconomic resilience, which we now define as the capacity of affected populations to cope with and recover from these losses and quantify as the ratio of expected asset losses to wellbeing losses. At the national level, precipitation flooding causes Sri Lankan households to have US$78 million per year in asset losses and US$119 million per year in wellbeing losses. In other words, the average annual impact of flooding on wellbeing in Sri Lanka is 53% greater than is represented by asset losses. Socioeconomic resilience is the ratio of these two metrics, and, therefore, the socioeconomic resilience of Sri Lanka to 78 disasters is 66% ( 119 = 0.66). As a matter of macro-fiscal policy, the government 5 More precisely, the impact on wellbeing is equivalent to a US$122 million decrease in consumption that would be optimally shared across households in the country and across time. Or it is equivalent to a US$122 million decrease in consumption that would be equally shared across households in the country, if all households were identical (i.e. if all inequality had disappeared). 19 Figure 5: Socioeconomic resilience to precipitation flooding in Sri Lanka. of Sri Lanka can use this scale factor to estimate the costs of floods to household consumption, taking into account the relative wealth of disaster-affected communities. For example, wellbeing losses can be used as inputs to disaster risk management budgeting, prioritization, and investment decisions.6 6 This estimate is significantly lower than that from the Unbreakable report, and cannot be directly com- pared, considering the difference between the models. The difference is explained by the improved consideration of distributional impacts (using the full survey instead of only two categories of house- holds) and the explicit representation of the reconstruction pathway. The difference confirms the need to model the reconstruction pathway in a dynamic manner and to include the impact of short-term consumption drops. 20 Although this result already deepens our understanding of the distributional im- pacts of flooding in Sri Lanka, the national average of socioeconomic resilience be- lies significant regional, sub-regional, and socioeconomic variation in the capacity of households to cope with and recover from disasters. Increasing spatial resolution just one step, Figure 5 on the preceding page maps socioeconomic resilience to dis- asters at the regional level. Among all regions, greater Colombo enjoys the highest resilience (105%). This is due not just to its overall wealth and poverty rate of just 1%, but also to its high degree of financial inclusion and social protection coverage (30% higher than the national average). As a result of these resilience-building char- acteristics, wellbeing losses to disasters in Colombo district are slightly lower than asset losses. Of course, among the 2.3 million residents of the capital, there are sig- nificant disparities in disaster impacts, and this analysis should not be construed as neglecting or minimizing the challenges facing the urban poor. However, the "aver- age” resident of Colombo is better prepared to cope with and rebuild from disasters than the "average" Sri Lankan outside the capital. At the other end of development in Sri Lanka, socioeconomic resilience to disasters is less than or equal to 30% in Trincomalee, Batticaloa, Batticaloa, and Mullaitivu. This indicates that the average resident of these districts struggles to cope with and recover from shocks when they occur. This struggle results in a lower likelihood of recovery in the long term, and is due to a complex of factors (e.g., poverty incidence, diversity of income sources, financial inclusion, and social protection enrollment), the net effects of which, wellbeing losses are designed to measure. On average, residents of these districts suffer wellbeing losses three times greater than their asset losses when floods occur. Because of these dynamics, wellbeing losses are less concentrated in greater Colombo than are asset losses: while flooding in greater Colombo causes on average 52% of nationwide asset losses, it causes 32% of nationwide wellbeing losses. This suggests that well-designed DRM interventions have a large potential to help vulnerable communities (which are almost always fragile in other dimensions, as well) cope when shocks inevitably occur, and particularly outside the capital. 21 Socioeconomic Asset risk resilience Well-being risk District [,000 US$/year] [%] [,000 US$/year] Colombo 40,235 105 38,260 Rathnapura 8,194 32 25,487 Kalutara 7,434 62 12,050 Gampaha 6,931 73 9,449 Puttalam 3,321 46 7,252 Trincomalee 1,257 29 4,315 Kurunegala 2,418 56 4,315 Polonnaruwa 2,099 49 4,312 Hambantota 1,552 61 2,540 Kegalle 1,006 41 2,442 Batticaloa 592 26 2,240 Mannar 528 38 1,396 Galle 673 53 1,280 Badulla 293 31 948 Kandy 430 49 880 Anuradhapura 473 55 860 Ampara 304 42 722 Matale 170 44 384 Moneragala 40 30 136 Mullaitivu 27 26 104 Vavuniya 44 54 82 Nuwara Eliya 1 34 4 Matara – 40 – Jaffna – 31 – Kilinochchi – 18 – Total 78,021 65 119,462 Table 2: Risk to assets, risk to wellbeing, and socioeconomic resilience for the entire popula- tion of each district of Sri Lanka. Note: though our hazard dataset does not report asset loss exceedance curves for Matara, Jaffna, or Kilinochchi, still we are able to as- sess their respective levels of socioeconomic resilience using household survey data, and the fact that resilience is the ratio of asset to wellbeing losses. . District-level risk summary The annual risk to assets, risk to wellbeing, and socioeconomic resilience for each district of Sri Lanka are summarized in Table 2. According to the SSBN model, annual asset losses to precipitation flooding are highly concentrated in Colombo district, totaling over US$40 million per year (1.0% of AHE). Rathnapura district suffers the second highest losses, valued at US$8.2 million per year (0.9% of AHE). 22 However, in Rathnapura, wellbeing losses are driven less by asset losses from pre- cipitation flooding, than by the low socioeconomic resilience of the district. With a district-level poverty rate over 10%, Rathnapura is ill-equipped to deal with shocks. Among all households in Rathnapura, 47% report income from social transfers or re- mittances. This value is close to the national average (48%), but far lower than other rural and relatively poor provinces (for example, 77% of households in Mullaitivu report social income). Further, the total value of social transfers and remittances is on average 35% lower than the national average, and 66% lower than in Colombo district. Figure 6: Risk to wellbeing (expected annual losses) from exposure to precipitation floods in Sri Lanka. Losses are shown in thousands of dollars (left) and in dollars per capita, per year. 23 Annual Socioeconomic Annual asset risk resilience wellbeing risk District [,000 US$] [% of total] [%] [,000 US$] [% total] Colombo 2,891 7 25 11,574 31 Rathnapura 701 9 8 8,587 34 Kalutara 451 6 12 3,627 31 Gampaha 519 7 19 2,751 30 Puttalam 279 8 13 2,211 31 Trincomalee 97 8 6 1,568 37 Kurunegala 176 7 12 1,430 34 Polonnaruwa 162 8 13 1,231 29 Kegalle 81 8 10 846 35 Batticaloa 50 8 7 767 35 Hambantota 118 8 16 722 29 Galle 50 7 13 396 31 Mannar 48 9 16 304 22 Kandy 29 7 10 303 35 Badulla 22 8 7 303 32 Anuradhapura 36 8 14 258 31 Ampara 28 9 14 200 28 Matale 13 7 10 128 34 Moneragala 4 10 8 49 36 Mullaitivu 2 8 7 32 31 Vavuniya 3 8 15 23 29 Nuwara Eliya 0 9 9 1 34 Matara – – 9 – – Jaffna – – 8 – – Kilinochchi – – 5 – – Total 5,761 7 15 37,313 32 Table 3: Risk to assets, socioeconomic resilience, and risk to wellbeing for the poorest 20% (by expenditures) of the population of each district. . Poorest quintile risk summary Because our results are based on information at the household level, we are able to assess asset and wellbeing losses not just at the district level, but also for different income groups.7 Table 3 lists annual asset and wellbeing losses for the poorest quin- tile in each district. Across all districts, the asset losses of the poorest 20% are US$5.7 million, or just 7% of total asset losses. On the other hand, the wellbeing losses of the 7 It is also possible to look at different subgroups in the country or at the district scale (e.g., per oc- cupation, head of household gender, household size, ethnic background or religion, social transfer enrollees), within the limits of the representativeness of the household survey. 24 Per capita losses Q1 population Asset Well-being District [thousands] [US$ per cap., per year] national result 4,367 1.42 9.47 Rathnapura 236 2.96 37.07 Colombo 473 6.08 24.80 Trincomalee 82 1.17 19.18 Mannar 20 2.32 14.89 Kalutara 259 1.74 14.44 Polonnaruwa 90 1.83 14.19 Puttalam 165 1.68 14.14 Batticaloa 114 0.45 7.25 Gampaha 479 1.10 6.02 Hambantota 133 0.89 5.61 Kegalle 185 0.44 4.73 Kurunegala 350 0.51 4.26 Galle 231 0.22 1.81 Badulla 181 0.12 1.68 Mullaitivu 18 0.11 1.67 Anuradhapura 178 0.20 1.47 Ampara 141 0.19 1.39 Matale 111 0.11 1.19 Kandy 303 0.10 1.03 Vavuniya 37 0.09 0.65 Moneragala 100 0.04 0.51 Nuwara Eliya 154 0.00 0.01 Kilinochchi 24 – – Jaffna 124 – – Matara 179 – – Table 4: Annual per capita asset and wellbeing losses for the poorest 20% (by expenditures) of the population of each district. poorest quintile give a better sense of their experience of disasters: valued at US$37.3 million per year, the wellbeing losses of the poorest quintile are 32% of total wellbeing losses. It means that on average, and in wellbeing terms, the poorest quintile suffers from losses that are twice as large as the average individual loss in the country. To this point in the discussion of results, it is likely that the district-level averages underrepresent the urban poor, who face higher costs of living in search of economic opportunity, and who may not only not receive support from their home communities, but are expected to send remittances home. Indeed, when we limit our analysis to just the poorest quintile in each district, the range of socioeconomic resilience 25 values narrows significantly. For example, recall that the entire population of greater Colombo enjoys on average a resilience of 105% (cf. Table 2). However, the poorest 20% of residents of the capital have a resilience of only 25% (cf. Table 3). On average, the poorest residents of Colombo district lose just over US$6 (LKR 900) per person, per year to precipitation flooding (7% of total asset losses), but this equates to US$25 (LKR 3,800) per person, per year in wellbeing losses, or 30% of total wellbeing losses (cf. Table 4 on the previous page). In other words, the aggregate wealth of the capital district does not imply that its poorest residents are well-protected from or resilient to asset losses. To the contrary: the low resilience of the poor to asset losses, combined with the large contribution of disasters to poverty incidence in Colombo, suggests that interventions to reduce asset losses and build resilience of the poor could be very effective for reducing disaster exposure and poverty incidence in Colombo and other urban areas. In addition to these diagnostics, the extension of the traditional hazard-exposure- vulnerability framework to include socioeconomic resilience has the benefit of ex- panding the disaster risk management "toolbox." Traditional risk-management strate- gies seek to mitigate hazards as well as the exposure and vulnerability of assets. Other tools (e.g., post-disaster support, financial inclusion, private and public insur- ance, and sovereign contingent credit) are not easily included in risk assessments because they do not affect asset losses. In a wellbeing-informed framework, however, the benefits of these interventions become obvious and quantifiable: to the degree that they allow households to maintain a healthy degree of consumption while they rebuild, these interventions increase socioeconomic resilience and reduce wellbeing losses to disasters. Here, we use avoided wellbeing losses to assess the benefits of post-disaster sup- port in general, and in terms of various reforms to Samurdhi, the most significant social protection system in Sri Lanka. The Samurdhi program was established by 26 the Government of Sri Lanka in 1995, with the main goal of reducing poverty in the country. . Proxy means test (PMT) A proxy means test (PMT) is an index of observable and verifiable household charac- teristics, such as the quality of the household’s dwelling, ownership of durable goods, demographic structure, and the education and employment status of adult household members [9]. PMT scores provide a proxy for household incomes, and are commonly used to determine or optimize eligibility for social safety net programs in situations where verifiable income data are not available. Since the PMT essentially ranks households by welfare level, the same formula can be used for targeting a range of different welfare programs, adopting different eligibility cutoff scores for each program and possibly adding other criteria as well. Having a common targeting approach for various programs, based on a single PMT score, ensures consistency of targeting across programs, minimizes overlaps, and provides a basis for future harmonization of programs. Samurdhi program benefits are currently distributed on the basis of self-reported income. However, the Sri Lankan government is in the process of updating this system, and one component of this reform is to identify beneficiaries using a standard vector of household attributes in place of self-reported incomes. This shift is expected to significantly improve the targeting of Samurdhi benefits to the poor in Sri Lanka. In this analysis, we construct four illustrative disaster-responsive adaptive social protection (ASP) systems. Like Samurdhi, these ASP systems distribute cash transfers to designated households on a one-time or monthly basis. Enrollment and disburse- ments for each ASP alternative are based disaster-affected status, existing Samurdhi enrollment, and PMT scores, as described in the following sections. In some cases, we use a PMT score of 887 as a cutoff for benefits. This value corresponds to the 25th percentile of the “true" per capita consumption distribution in Sri Lanka, as measured by the detailed HIES survey [10]. 27 . Case study: 50-year flood in Rathnapura When the 50-year precipitation flood event strikes Rathnapura, it is expected to cause US$43 million in asset losses and US$131 million in wellbeing losses. In Figure 7, cumulative asset and wellbeing losses are plotted against PMT scores (i.e. the y-axis shows total asset and wellbeing losses for all households with PMT less than the value on the x-axis). The annotations inside the figure describe the asset and well- being losses for each income quintile, independently. Notably, wellbeing losses are 8.4 times greater than asset losses for the poorest quintile in Rathnapura (socioeco- 4 .7 nomic resilience = 38.0 = 12%), while wellbeing losses are less than asset losses for the 16.1 wealthiest quintile (socioeconomic resilience = 15.0 = 107%). This indicates that the wealthiest residents of Rathnapura are as resilient to shocks as the average resident of Colombo (cf. Table 2). 50-year precipitation flood in Rathnapura Total wellbeing losses = $130.8 million Wealthiest quintile Total asset losses = $42.9 million $15.0 mil. $16.1 mil. Fourth $20.1 mil. 120 $8.9 mil. Third $25.3 mil. 100 $7.2 mil. Cumulative losses [mil. US$] Second 80 $32.3 mil. $6.1 mil. 60 Poorest quintile $38.0 mil. 40 $4.7 mil. 20 0 850 900 950 1000 1050 1100 Household income [PMT] Figure 7: Expected household asset and wellbeing losses in a 50-year precipitation flooding event in Rathnapura. The y-axis shows cumulative expected losses for all affected households with PMT less than the value shown on the x-axis. 28 In Figure 8, the two leftmost sets of bars present again expected asset and wellbeing losses, by quintile, for this event. The figure shows that while the richest households are expected to lose more assets, the poorest households are expected to suffer the greatest wellbeing losses. The figure also shows how small, uniform, lump sum post- disaster support delivered with perfect targeting to all affected households (“uniform payout") could theoretically reduce wellbeing losses, especially for the poorest house- holds. The first quintile sees its wellbeing losses reduced by around a third, while impact is small on the richest quintile.8 In the next section, we will look in much greater detail at optimizing ASP response to a particular event–here, the 50-year pre- cipitation flood in Rathnapura. 1400 Poorest quintile Second Third Disaster losses (US$ per affected person) 1200 Fourth Wealthiest quintile 1000 800 600 400 200 0 Asset loss Wellbeing loss Net cash benefit of Wellbeing loss uniform payout with uniform payout Figure 8: Cost-benefit analysis of uniform payout to all affected individuals in response to a 50-year precipitation flooding event in Rathnapura. 8 It is important to note that even if the amount of post-disaster support is equal to asset losses, it does not fully cancel wellbeing losses: indeed, post-disaster support maintains consumption, but consumption losses are larger than asset losses. This result is consistent with intuition: even if people are immediately given in cash the cost of rebuilding their houses and replacing their assets, they would still experience wellbeing losses during the reconstruction period, since assets and houses cannot be replaced instantaneously. 29 benefit = 2.57 1 month-equivalent top-up to existing cost Samurdhi enrollees in Rathnapura 3.0 Benefit after 50-year flood 3.0 2.5 Cost of Samurdhi scaleup [mil. USD] 2.6 2.18 2.0 1.5 Cost 2.69 1.2 1.33 1.0 0.84 0.49 0.48 3.12 0.5 0.15 0.0 Scaleup to Scaleup to Scaleup to Perfect all Samurdhi all affected & all affected & targeting within beneficiaries 66% of unaffected 33% of unaffected Samurdhi Inclusion error Figure 9: Cost-benefit analysis of Samurdhi scaleup in response to a 50-year precipitation flooding event in Rathnapura, as a function of the inclusion error on post-disaster support. . Post-disaster support to Samurdhi enrollees In practice, it is much easier to leverage existing social protection systems to deliver post disaster support, taking advantage of established beneficiary rolls and delivery mechanisms, than to implement new systems. In order to model a more realistic post-disaster response than the “uniform payout" presented above, we simulate a scaleup of Samurdhi benefits to households already enrolled in the program. In this simulation, households enrolled in Samurdhi before the disaster receive a one-time top up, valued at one month of their normal receipts. This system takes advantage of the established assets and procedures of Samurdhi, on the assumption that the program is already well-targeted for delivering aid to the poor. The cost of the program is 30 distributed to all Sri Lankan households via a flat tax on income, and we assume that there is no transaction cost to run the program. 1 month-equivalent Samurdhi benefit = 4.17 payment to Rathnapura Benefit: 0.82M cost 0.8 after 50-year flood Perfect targeting Cost, benefit of post-disaster support [mil. USD] 0.6 benefit = 3.12 cost Benefit: 0.48M 0.4 0.2 Cost: 0.2M Cost: 0.15M 0.0 Scaleup to affected Scaleout to affected Samurdhi enrollees using PMT (3.6% of district (4.6% of district are beneficiaries) are beneficiaries) Figure 10: Cost-benefit analysis of scale-up using Samurdhi enrollment, compared to scaleout using PMT (upper threshold = 887) in response to a 50-year precipitation flooding event in Rathnapura. Perfect targeting is assumed in both cases. In response to the 50-year flood in Rathnapura, the Samurdhi scaleup disburses US$1.2 million to US$150,000 in post-disaster support, depending on how accurately disaster-affected households can be identified among all Samurdhi enrollees in the district (cf. Figure 9). At the high end of this cost range, a payout to every house- hold in the district that is enrolled in Samurdhi (i.e., without any targeting of affected households) reduces wellbeing losses by US$3.2 million, delivering a benefit-to-cost ratio (BCR) of 2.67 on the investment in post-disaster support.9 At the other end of the estimated range, assuming it would be possible to deliver aid strictly to af- 9 Note that the benefits calculated here include the wellbeing benefits of non-affected people who receive post-disaster support; the assessment is not limited to affected populations. 31 fected households without any inclusion or exclusion error, we calculate US$500,000 in avoided wellbeing losses. This idealized scenario achieves a BCR ratio of 3.22.10 In Figure 10, we compare the costs and benefits of scaling Samurdhi benefits up (top up to existing beneficiaries), versus out (disbursements based on PMT scores). For the “out" scenario, poor households are ranked according to their PMT score, and all households with PMT less than 887 (including those already enrolled in Samurdhi) receive the average per capita, monthly Samurdhi payout for their district[10]. We assume perfect targeting for both “up" and “out." Our analysis indicates that, while the scaleout is more expensive at US$1.5 million, due to the high PMT threshold for support, it achieves a better BCR than the more straightforward scaleup of benefits to households already enrolled in Samurdhi. This result suggests that targeting of Samurdhi benefits can be improved in Rathnapura district. 50-year precipitation flood in Rathnapura Samurdhi to affected for 1 month (perfect targeting) Poorest Second Third Fourth Wealthiest 20.0 quintile quintile Marginal impact at threshold [US$ per next enrollee] 17.5 15.0 12.5 10.0 7.5 5.0 PDS cost 2.5 0.0 850 900 950 1000 1050 1100 1150 Upper PMT threshold for post-disaster support Figure 11: This plot shows the marginal cost and benefit of increasing the PMT threshold for PDS, across the range of PMT scores in Rathnapura. 10 Note that the flat tax payment mechanism used here effects a net transfer from unaffected districts into Rathnapura; district-level risk pooling, progressive taxation, social insurance-like systems, and more complex alternatives can also be modeled. 32 . Post-disaster support to PMT-qualified households Finally, Figure 11 presents the marginal cost and benefit of post-disaster support to flood-affected households in Rathnapura. For each PMT value on the x-axis, the plot shows the expected value (blue) and cost (green) of raising the PMT threshold to add an additional household. Because the payout is uniform, the cost is constant across all PMT values. However, the expected benefit falls from US$20 per enrollee with the lowest PMT scores, to negative BCR for enrollees with the highest PMT scores. The lines cross around 950, indicating a net benefit to the provision of post disaster support to all households with PMT scores below this threshold. In this way, the PMT can be used to optimize the coverage and value of post-disaster support to affected communities using Samurdhi as a delivery mechanism. . Nationwide cost-benefit analysis of Adaptive Social Protection In Figure 12, we plot the expected benefits of four different ASP programs, by flood- ing magnitude (return period), for all districts in Sri Lanka. In bold at right, the criteria for beneficiary selection are detailed. In order of increasing cost, we measure the benefits of delivering a one-time payout, equal to a month of Samurdhi, to the following groups: all existing Samurdhi enrollees who are affected by flooding; all af- fected individuals with PMT 887 (irrespective of Samurdhi enrollment); all affected individuals; and all Samurdhi enrollees, without targeting. Assuming any flood with at least a 10-year return period triggers a payout from these programs, their annual cost ranges from US$48 thousand to US$1.22 million, depending on the number of beneficiaries in each program. Figure 12 also details the average, expected benefit-cost ratio (BCR) for each of these programs. Even assuming perfect targeting, the program that delivers lump sum payouts to all affected individuals has a relatively high cost and low return on investment (BCR = 1.2) because this program does not impose a means test on recip- ients. The top-up to all existing Samurdhi enrollees delivers an BCR of 2.0, followed by perfect targeting of affected individuals who are already enrolled in Samurdhi 33 Expected benefit of ASP (1 month Samurdhi top-up) by return period & beneficiary group Cost of ASP All Samurdhi enrollees after 50-year flood (no targeting) $14.3M Annual cost = 1.43M 25 Program BCR = 2.0 Avoided wellbeing losses [mil. US$] 10 All affected 5 (perfect targeting) Annual cost = 306K Program BCR = 1.2 $3.5M Affected and PMT 887 (perfect targeting) Annual cost = 60K Program BCR = 3.5 $676K Samurdhi enrollee and affected (perfect targeting) Annual cost = 48K Program BCR = 2.4 $532K 1 10 25 50 100 500 1000 Return period [years] Figure 12: Expected benefit of four ASP responses to precipitation flooding, by disaster mag- nitude (return period) and inclusive of all districts in Sri Lanka. Benefit is the expected reduction to wellbeing losses with ASP, compared to a scenario with no post-disaster support. Note: both axes are logarithmic. (BCR = 2.4). Finally, a new social protection system targeted at poor (PMT 887), flood-affected individuals delivers the greatest return (BCR = 3.5) at a cost of US$60K per year. These cost and benefit estimates are based on simple coverage criteria, applied uniformly to all districts. In practice, criteria for post-disaster support may vary by province and disaster magnitude. This framework can model all such complexities, up to the granularity and representativeness of the household survey and hazard maps used as inputs. This paper has presented the results of a risk assessment based on an expanded framework, which includes in the analysis the ability of affected households to cope with and recover from disaster asset losses and uses “wellbeing losses” as its main measure of disaster severity. This framework adds to the three usual components 34 of a risk assessment — hazard, exposure, and vulnerability — a fourth component, socioeconomic resilience. Like the traditional components of risk management, so- cioeconomic resilience can be measured at any degree of spatial resolution, from the household to national averages. Using a new agent-based model that represents explicitly the recovery and recon- struction process at the household level, this risk assessment provides more insight into flood impacts in Sri Lanka than a traditional risk assessment. In particular, it shows how the regions and communities identified as priorities for risk-management interventions differ depending on which risk metric is used. While a simple cost- benefit analysis based on asset losses would drive risk reduction investments toward the richest regions and areas, a focus on poverty or wellbeing rebalances the analysis and leads to a set of priorities to integrate DRM into the larger development agenda. In parallel, measuring disaster impacts through poverty and wellbeing implications helps quantify the benefits of interventions that may not reduce asset losses, but do reduce their wellbeing consequences by making the population more resilient. These interventions include financial inclusion, social protection, and more generally the provision of post-disaster support to affected households. The model and data used in this analysis have many limitations. For instance, the 2016 HIES is only statistically representative at the district level and does not offer the geolocalization of households. Further, the SSBN model is itself a complex and imperfect exercise. Finally, the model does not represent all coping mechanisms avail- able to households, such as international remittances or temporary migrations. How- ever, even with these limits and simplifications, the introduction of socioeconomic resilience and household characteristics into risk assessments provides actionable in- sights and appears as a promising research agenda. 35 The authors wish to recognize the work of many colleagues who contributed to this report, including Thomas Walker, Jinqiang Chen, Adrien Vogt-Schilb, Mook Banga- lore, the World Bank country office in Colombo, and Ivan Vuarambon and Laksiri Nanayakkara (World Food Program). Asset risk is a product of the SSBN/Fathom model, not the authors. All errors, interpretations, and conclusions are the responsi- bility of the authors. [1] Christopher C. Sampson, Andrew M. Smith, Paul B. Bates, Jeffrey Neal, Lorenzo Alfieri, and Jim E. Freer. A high resolution global flood hazard model. Water Resources Research, 51, 08 2015. [2] Brian Walsh and Stephane Hallegatte. Socio-economic resilience in the philip- pines: Managing disaster risk to protect the well-being of individuals and com- munities. Policy Research Working Papers, 2018. [3] Angus Deaton. The analysis of household surveys: a microeconometric approach to development policy. World Bank Publications, 1997. [4] Robert M Townsend. Consumption insurance: An evaluation of risk-bearing systems in low-income economies. Journal of Economic Perspectives, 9(3):83–102, 1995. [5] Jonathan Morduch. Income smoothing and consumption smoothing. Journal of Economic Perspectives, 9(3):103–114, 1995. [6] Michael R. Carter and Christopher B. Barrett. The economics of poverty traps and persistent poverty: An asset-based approach. Journal of Development Studies, 42(2):178–199, 2006. 36 [7] Michael R Carter, Peter D Little, Tewodaj Mogues, and Workneh Negatu. Poverty traps and natural disasters in Ethiopia and Honduras. World Development, 35(5):835–856, 2007. [8] Stephane Hallegatte, Adrien Vogt-Schilb, Mook Bangalore, and Julie Rozenberg. Unbreakable: Building the Resilience of the Poor in the Face of Natural Disasters. Cli- mate Change and Development. Washington, D.C. : World Bank Group., 2016. [9] Margaret E Grosh, Carlo Del Ninno, Emil Tesliuc, and Azedine Ouerghi. For protection and promotion: The design and implementation of effective safety nets. The World Bank, 2008. [10] Ashwini Sebastian, Shivapragasam Shivakumaran, Ani Rudra Silwal, David Newhouse, Thomas Walker, and Nobuo Yoshida. A Proxy Means Test for Sri Lanka. Policy Research Working Papers, No. 8605. World Bank, Washington, DC., Oct 2018. [11] L. Le De, J.C. Gaillard, and W. Friesen. Remittances and disaster: a review. International Journal of Disaster Risk Reduction, 4:34–43, Jun. 2013. [12] Dean Yang and HwaJung Choi. Are Remittances Insurance? Evidence from Rainfall Shocks in the Philippines. The World Bank Economic Review, 21(2):219– 248, Jan. 2007. [13] Alvina Erman, Elliot Motte, Radhika Goyal, Akosua Asare, Shinya Takamatsu, Xiaomeng Chen, Silvia Malgioglio, Alexander Skinner, Nobuo Yoshida, and Stephane Hallegatte. The road to recovery: the role of poverty in the exposure, vul- nerability and resilience to floods in Accra. Policy Research working papers, No. WPS 8469. World Bank, Washington, DC., 2018. [14] Stéphane Hallegatte, Jun Rentschler, and Brian Walsh. Building Back Better: Achieving Resilience through Stronger, Faster, and More Inclusive Post-Disaster Re- construction. World Bank Group, 2018. [15] Stephane Hallegatte and Adrien Camille Vogt-Schilb. Are Losses from Natural Disasters More than Just Asset Losses? The Role of Capital Aggregation, Sector Interac- — 37 tions, and Investment Behaviors. World Bank Policy Research Working Papers, No. 7885. Washington, D.C.: World Bank Group, 2016. [16] Stefan Dercon and Catherine Porter. Live aid revisited: Long-term impacts of the 1984 Ethiopian famine on children. Journal of the European Economic Association, 12(4):927–948, Aug. 2014. [17] Alain de Janvry, Frederico Finan, Elisabeth Sadoulet, and Renos Vakis. Can con- ditional cash transfer programs serve as safety nets in keeping children at school and from working when exposed to shocks? Journal of Development Economics, 79(2):349–373, 2006. [18] Germán Daniel Caruso. The Legacy of Natural Disasters: The Intergenerational Impacts of 100 Years of Disasters in Latin America. Journal of Development Eco- nomics, 127(March):209–233, 2017. [19] Solomon M Hsiang and Amir S Jina. The causal effect of environmental catas- trophe on long-run economic growth: Evidence from 6,700 cyclones. Technical report, National Bureau of Economic Research, 2014. [20] Marc Fleurbaey and Peter J. Hammond. Interpersonally Comparable Utility. In Handbook of Utility Theory, pages 1179–1285. Springer US, Boston, MA, 2004. — This section explains the methodology and describes the model used to translate asset losses into wellbeing losses. The code of the model is freely available, and the reader is invited to refer to the code for the implementation of the principles and equations presented in this section.11 The household survey data cannot be made available directly, as they need to be requested from the Department of Census & Statistics of Sri Lanka. 11 https://github.com/walshb1/hh_resilience_model — 38 In all applications, the model assumes a closed national economy. In terms of dis- aster risks, this means that 100% of household income is derived from assets located inside the country, and that post-disaster reconstruction costs can be distributed to non-affected taxpayers throughout the country, but not outside its borders.12 Pre-disaster situation Population & weighting The pre-disaster situation in the country is represented by the households described in the Household Income and Expenditure Survey (HIES). We use a per capita weight- ing (ωh ), such that summing over all households in the survey or in an administrative unit (Nh ) returns the total population (P): Nh P= ωh (1) h=0 One essential characteristic of each household is its income (ih ). As defined by the HIES, ih combines primary income and receipts from all other sources, including the imputed rental value of owner-occupied dwelling units, pensions and support, and the value of in-kind gifts and services received free of charge. The data recorded in the survey are assumed to capture the household’s permanent income, which is smoothed over fluctuations in income and occasional or one-off expenditures.13 It is important to note that the value of housing services provided by owner-occupied dwellings is included in the income data, so that the loss of a house has an impact on income (even though it would not affect actual monetary income).14 12 This is a serious limit in a country where international remittances have reached more than 8 percent of the gross national income in 2017, and where remittances have been shown to support post-disaster recovery [11, 12]. 13 In some countries, it may be necessary or preferable to infer household income from expenditures, whether because incomes are not reported, because consumption is more stable over time, or because the official poverty statistics are calculated from consumption rather than income. 14 Similarly, the services provided by other assets (e.g., air conditioners, refrigerators) could be added as an additional income that can be threatened by natural disasters. — 39 Social transfers, taxation, and remittances The enrollment and value of social transfers (isp h ) are listed in HIES, and the total cost (Csp ) of these programs to the government is given by a simple sum:15 Nh Csp = ωh isp h (2) h=0 All incomes reported in HIES are assumed to be reported net of the income tax that finances general spending of the government (for infrastructure and other services) and of an additional flat income tax that finances social programs (rate = δtax sp ). The rate δtax sp can be estimated with the following equation: 16 Nh Nh Nh Csp ωh isp h = ωh ih = ωh ih δtax sp (3) ωh ih h=0 h=0 h=0 Note that, since ih includes isp h , income from social programs is treated as taxable in the model. A similar approach is used to derive the tax rate (δtax pub. ) to fund post- disaster reconstruction of public assets. Remittances and transfers among households play a very important role for peo- ple’s income. Some of these transfers are within a family or a community, while others are international. The HIES provides estimates of the amount received, but it is of course impossible to represent the bilateral flows of resources among households. For this reason, remittances are modeled like an additional social protection scheme: the transfers received from friends and family are added to the social transfers, and it is assumed that these transfers comes from a single fund, in which all households contribute proportionally to their income (like a flat tax). Under these assumptions, remittances can be aggregated with social protection and redistribution systems. This is of course a simplification, especially in that it does not account from international remittances, which have been shown to play a role after disaster [12]. 15 Administrative costs are not included in program cost estimates. When household data do not in- clude the transfers, then transfers from social programs can be modeled on the basis of the actual disbursement rules that qualify households for participation in each program (eg, PMT score, house- hold number of dependents or senior citizens, employment status, etc.). 16 Although we assume a flat tax, the model is capable of handling more complicated tax regimes, in- cluding progressive taxation. — 40 Income, capital, & consumption Household income is equal to the sum of the social transfers and domestic and inter- national remittances (as discussed above) and the value generated by a household’s effective capital stock (keff h ), times the average productivity of capital in the country (Πk ), reduced by the flat tax at the rate δtax sp . 17 ih = isp tax eff h + (1 − δsp ) · kh · Πk (4) In practice, the household’s effective capital stock (keff h ) is estimated based on the income and transfers reported in the HIES, and the tax level δtax sp that would balance the budget: Income from assets ih − isp keff h · Πk = h tax (5) 1 − δsp Gross of taxes All household income not from transfers is assumed to be generated by household effective assets, including some assets not owned by the household (like roads and prv factories). Some of the assets represented by keff h are private (kh ), such as equipment used in family business or livestock; some assets are public (kpub h ), such as road and the power grid (and possibly the environment and natural capital); and some assets are owned by other households (koth h ), but still used to generate income by the household, such as factories. The productivity and vulnerability of these assets to various hazards can vary, so it is useful to disambiguate among them as much as the data allow. By construction, our model distinguishes among asset classes, as allowed by data. In addition, these distinctions are important to understand the consumption and wellbeing losses that follow a disaster, since the liability for reconstruction varies: households rebuild their own assets (unless they carried private insurance); the national, regional, or provincial 17 Since general spending of the government is not explicitly represented in the income ih , the effective capital stock estimated here keff h is net of the resources used to finance this general spending through taxes. — 41 taxpayers rebuild public assets; and other privately-held assets are reconstructed by private business owners or corporations. prv pub For each household, we distribute keff h to private (kh ) and public (kh ) using the fraction of private to total asset losses in the district, assuming that (1) the ratio of the different capital categories are similar for all households, and (2) the vulnerability of each household’s public assets is given by the vulnerability of its private assets. See the technical appendix of [2] for a discussion of data in the Philippines. Lacking these data in the case of Sri Lanka, we make the necessary assumption that all income- generating assets are the personal property of each household. This may overestimate the reconstruction costs paid by households to rebuild disaster-affected assets. Precautionary savings Precautionary savings play a key role in managing disasters, but there is no esti- mate of these savings in the HIES. Also, although the HIES provides information on both income and consumption, the difference between income and consumption is highly variable and negative for many households (aggregate consumption greater than reported income), making it an uncertain indicator of savings at the household level. Instead, we calculate the average gap (income less consumption) by district and decile. We then assume that each household maintains one year’s surplus as precau- tionary savings: separate from their productive assets, and available to be spent on recovery or consumption smoothing. Household versus nominal GDP Based on this definition of each household’s income and assets, we note that the total capital stock (K) of any country is given by a sum over the effective capital of all households, and the portion of national GDP from household consumption (hhGDP) as reported in the HIES is given by the product of K and the average productivity of capital (Πk ). Nh K= ωh keff h (6) h=0 — 42 Table 5 lists the aggregated household expenditures (AHE) for each district in Sri Lanka. AHE National accounts District (bil. LKR/year) Colombo 609.8 – Gampaha 460.3 – Kalutara 233.9 – Kandy 218.3 – Matale 69.0 – Nuwara Eliya 88.3 – Galle 172.8 – Matara 116.1 – Hambantota 109.4 – Jaffna 66.8 – Mannar 13.5 – Vavuniya 25.4 – Mullaitivu 9 .0 – Kilinochchi 9 .8 – Batticaloa 49.8 – Ampara 81.2 – Trincomalee 45.3 – Kurunegala 288.8 – Puttalam 135.4 – Anuradhapura 129.1 – Polonnaruwa 61.3 – Badulla 102.8 – Moneragala 52.1 – Rathnapura 135.2 – Kegalle 118.1 – Total 3,401.6 11,400.0 Table 5: Aggregated household expenditures (AHE) at district level. At left, we sum total expenditures as reported in the 2016 HIES; at right, we present the nominal GDP of Sri Lanka. Values are expressed in millions of LKR per year. In the following sections, we will trace the impacts of disasters on household assets and wellbeing through the following steps: 1. Disasters result in losses to households’ effective capital stock (∆keff h ). 2. The diminished asset base generates less income (∆ih ). — 43 3. Reduced income contributes to a decrease in household consumption (∆ch ), but households affected by a disaster must further reduce their consumption to finance the repair or replacement of lost and damaged assets. 4. Household consumption losses are used to calculate wellbeing losses (∆wh ). One of the limits of the study is that we treat every event as independent, assuming that disasters affect the population as described by the 2016 HIES, and that two dis- asters never happen simultaneously (or close enough to have compounding effects). Asset losses The model starts from exceedance curves, produced by AIR Worldwide, which pro- vide the probable maximum (asset) loss (PML) for several types of natural disasters (earthquakes, tsunamis, tropical cyclones, storm surges, and fluvial and pluvial flood- ing), each administrative unit in the country, and various frequencies or return peri- ods. We make the simplification that a disaster affect only one district at a time, so that total losses in the affected district are equal to national-level losses. For each district, the input data details the total value of assets lost due to hazards as well as the frequency of each type of disaster over a range of magnitudes. Magni- tudes are expressed in terms of total asset losses (L). For example, the curves specify "An earthquake that causes at least $X million in damages in Y district is, on average, expected to occur once every Z years." When distributed at the household level, the losses L in the affected district can be expressed as follows: Nh L = Φa · K = ωh fah keff h vh (7) h=0 Setting aside for the moment the probability of a disaster’s occurrence (the "Haz- ard" component of Fig. 1 on page 8), Eq. 7 expresses total losses (L) in terms of total exposed assets (K) and the fraction of assets lost when a disaster occurs (Φa ). In the rightmost expression, losses are expressed as the product of each household’s prob- — 44 ability of being affected (fah , the "Exposure" component of Fig. 1 on page 8), total household assets (keff h ), and asset vulnerability (vh , cf. Sec. 6 on the next page). We make one important simplifying assumption, imposed by the data that are available: we assume that households are either affected or not affected; and, if they are affected, they lose a share vh of their effective capital keff h , with vh a function of household characteristics, independent of the local magnitude of the event. In case of a flood, one household is flooded or not, and if it is flooded, the fraction of capital lost will depend on the type of housing and other characteristics of the households and on a random process — but the losses do not depend on the local water depth or velocity. Similarly, an earthquake will affect a subset of the population who will experience building damages that depends on luck and the type of building — but the model does not take into account the ground motion at the location of the household. While this is of course a crude approximation, it is made necessary by the uncertainty on the exact localization of households in the HIES. With this approach, a bigger disaster is a disaster that affects more people, not a disaster that affects people more. Household exposure On the right side of Eq. 7 on the preceding page, fah is an expression of household exposure to each disaster. If we had perfect knowledge of each household’s exposure – for example, super high-resolution flood maps overlaid with the coordinates of every household (including those represented only implicitly in HIES) – we could assign a value of 0 or 1 to fah for each household and each event. Lacking this information, we interpret household exposure as the probability for any given household to be affected by the disaster when it occurs, and we assume that this probability is determined by household localization (at the highest resolution available) and characteristics. If the localization of a household in the HIES is known through the district, for instance, then the likelihood of one household to be flooded can be estimated by the fraction of the area of the district that is within the flood zone. Or, if population density maps are available and reliable, by the fraction of the population of the district living within the flood zone. This assumes that there is no relationship between income and exposure within a district. If poor people are found — 45 to be systematically more likely to be flooded, it is possible to introduce a "poverty bias" in the form of a higher probability of being affected for households with lower income. We do not have strong evidence that it is the case in Sri Lanka, and reviews suggest that such a bias is far from universal (see a review in [8], [13]). We therefore assume that the odds of being impacted by a given disaster are the same for all households in each administrative district. As a result, we can move fah out of the sum and drop the "h" subscript to indicate it is no longer household-specific (but it remains district-specific):18 Nh L = Φa · K = fa ωh keff h vh (8) h=0 This assumption also allows us to make a critical conceptual shift: if exposure is constant for all households in a given area, then we can reinterpret exposure (fa ) as the fraction of each household affected by a given disaster. After each disaster, of course, every household will be in exactly one of only two possible states: either it suffered direct impacts, or it escaped the disaster. On average, however, we can adopt a probabilistic approach by bifurcating each household in the HIES into two instances: affected and non-affected. We introduce this split in such a way that the total weight of each household (as well as asset losses at the household and provincial levels) remains unchanged:    ωh a = fa · ωh affected households ωh = ωha + ωhna (9)   ωh na = (1 − fa ) · ωh non-affected households Asset vulnerabilities The model assigns to each household a vulnerability (vh ), which describes the frac- tion of assets lost when a household is affected by a disaster. Again, this fraction does not depend on the local intensity of the hazard. Vulnerabilities are based on cat- egorical, qualitative information on the construction and condition of each domicile. 18 Where higher resolution household and disaster loss data are available, it is of course possible to expand Equation 8 to the provincial or sub-district level. — 46 In the application of the model to Sri Lanka, households are grouped into just two categories: fragile and robust, with associated vulnerabilities as described in Tab. 6.19 HIES descriptor Category vh tile, concrete, asbestos, metal sheet, taka ram robust 0.14 ± 0.06 cadjan, palmyrah, straw fragile 0.70 ± 0.14 other materials fragile 0.70 ± 0.14 Table 6: Asset vulnerability categories The right-most column in Tab. 6 indicates the fraction of assets lost when a house- hold is affected by a disaster. Each category includes a smearing factor (±20% for the moderate and fragile categories, ±40% for robust dwellings). This randomness recog- nizes a degree of irreducible uncertainty, including the fact that actual losses depend not only on whether a household is affected, but also on many situational factors (e.g., for floods, water depth and velocity; for earthquakes, local soil conditions) and some random factors. For each household, a value is chosen at random within the indicated range, allowing for variation as plotted in Fig. 13 on the next page. This method of assigning asset vulnerabilities involves a critical, simplifying as- sumption: the condition of each dwelling is assumed to be a direct proxy for the vulnerability of all assets that generate income for the household. This vulnerability factor is applied not just to household (private) assets, but also the assets that a house- hold does not own, but from which it derives income (e.g., roads, utilities, factories, agriculture, and other infrastructure). In other words, the model assumes that the vulnerability of assets not owned by a household but which it still uses to generate income is well-described by the condition of their private assets—for example, that the roads used by people who live in makeshift dwellings are not paved, and equally vulnerable to being destroyed as is their home. This assumption avoids significant increases in data requirements—indeed, global data on the vulnerability of infrastruc- ture are not available—and is necessary to avoid overly-complex representations of economic interactions between each household and assets held in common. 19 The effect of this too-simplistic vulnerability assumption is probably to underestimate the asset losses of the poor. — 47 4 10000 10 8000 3 10 HIES entries HIES entries 6000 2 10 4000 2000 1 10 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Asset vulnerability (vh) Asset vulnerability (vh) (a) Household asset vulnerability (vh ) (b) Same as (a), with y-axis plotted on a log scale Figure 13: Household asset vulnerability (vh ), constructed from qualitative domicile descrip- tions in the 2016 HIES from Sri Lanka. Vulnerability modifiers: early warning systems and access to finance When available, we incorporate data on the presence of early warning systems in affected districts. This reflects the assumption that early warning systems allow ex- posed households to move, reinforce, or otherwise protect their most fragile or valu- able assets, thus reducing their vulnerability to disaster. Using the same assumption as in the Unbreakable report [8], we assume that households who receive a warning are able to reduce their vulnerability by 20%, relative to identical households without access to early warning systems, by moving valuable items (from important papers to car or motorbikes) and implementing other mitigating measures (e.g., boarding windows, sandbagging doors). Summary of asset losses and calibration Returning to Eq. 7 on page 43, total losses in each district are defined as the losses suf- fered by each household, times that household’s likelihood of being directly affected by a disaster when it occurs. As mentioned in the previous section, losses include all assets that produce an income for the household—even those that are not owned by — 48 Disaster Household capital occurs k0 prv(t = ) = 0.05× k prv kh h 0 eff Effective household capital kh Private prv asset k0 losses eff(t) = k prve t/ h kh 0 + k0 othe t/ oth + k0 pube t/ pub Other asset k0 oth losses Public pub asset k0 losses -1 t0 1 2 3 4 Time after disaster [years] Figure 14: When disasters occur, affected household reconstruct its assets at the optimal rate while staying out of subsistence, as described in Sec. 6 on page 54. Here, we illustrate the reconstruction process for a household that suffers ∆keff 0 = ∆kprv 0 + pub ∆k0 in losses at time t=to , and reconstructs with period τh = 2.1 years. the household. If the vulnerability of all asset types is assumed to be linked to the vulnerability of the household’s private assets: Nh L= ωh · fa · ∆keff h (10) h=0 where: prv pub ∆keff h = v h · ( kh + kh + koth h ) (11) The calibration of fa and vh depends on the data that is available on hazards. Here, we start from results from the AIR catastrophe model, which provides an estimate of L in each district and each possible event (different hazards, different return periods). The value of vh is based on the damage function from 6 on the previous page, and is — 49 independent of the intensity of the hazards. Then, the number of affected people (or, equivalently, the probability for households to be affected) fa is calibrated such that the estimated losses are consistent with the AIR estimate.20 Income Losses To represent the longitudinal impacts of a disaster, it is not sufficient to consider the initial aggregate and distributional asset losses: one needs to explore the impacts on income and consumption, not only assets. Further, one needs to consider the dynamics of these impacts, not only the initial shock. The same asset losses do not cause the same effects if reconstruction and full recovery can be completed in a few months, compared with a case where various constraints make the recovery span years. To investigate this issue, asset losses are assumed from this point to be time- dependent: ∆keff eff h → ∆kh (t). When it is possible, the time variable is omitted below for simplicity and readability. Initial asset losses decrease throughout the reconstruction and recovery process as houses, infrastructure (i.e., roads and electric lines), and natural assets are repaired, replaced, and regrown. However, we assume that the income these assets had gen- erated for each household is diminished unless and until they are repaired. This includes the value households derive from their domicile; their appliances, vehicles, and livestock; and the infrastructure they use to commute to work or market. In this way, asset losses translate to income losses. Further, the reconstruction process is not free: households and governments have to invest in the reconstruction, at the expense of consumption (for households) or budget reallocation and increased taxes (for government). The objective of the model is to represent these processes, in order to estimate their longitudinal impacts on consumption, wellbeing, and poverty. It is important to note that the model assumes that households and governments aim at returning to the pre-disaster situation. It is well known that reconstruction can be used to “build back better” (for instance with more resilient building and in- 20 In exceptional cases where fa exceeds an upper threshold of 0.95, or 95% of households affected, the exposure is capped at 95% and the vulnerability vh is increased to match the asset losses from the AIR model. — 50 frastructure, but also with more efficient and productive assets); see [14]. And the reconstruction process is sometimes transformative for an economy [15]. However, measuring the impact of a disaster as the cost of returning to the pre-disaster sit- uation is non-ambiguous and objective, making shocks comparable even when the reconstruction leads to a different end point. sp ∆ih = (1 − δtax eff sp ) · Πk · ∆kh + ∆ih (12) Post-disaster income losses are described by Eq. 12. The first term specifies direct losses, while the last two terms incorporate secondary and indirect impacts on house- hold income, beyond the income losses resulting directly from a disaster. Each of these links between asset and income losses will be treated separately in this section. Direct income losses The definition of keff h includes all assets used by a household to generate income, including the value of owner-occupied dwellings that are generating a “virtual” in- come in the form of housing services. The first term in Eq. 12 represents the post-tax reduction in income due to the loss of assets, assuming that the income loss is simply proportional to the asset loss. sp ∆ih = (1 − δtax eff sp ) · Πk · ∆kh +∆ih (13) Direct losses For instance, in the absence of social transfers, a pastoralist losing 3 of 10 goats would see her income reduced by 30 percent. A factory worker working in a factory losing half of its machinery would experience a 50 percent loss in income. [15] pro- vides the theoretical basis for this linear relationship, in spite of decreasing returns on capital, based on imperfect substitutability of assets. Direct income losses are par- tially offset by the social transfers tax (δtax sp ), assuming that households’ tax burden is directly proportional to their asset base. — 51 Social transfers The second term in Eq. 12 on the preceding page, ∆isp h , represents the change in social transfers due to the decrease in tax revenue. As discussed above, these transfers in- clude also remittances, which are modelled as an additional social protection scheme. Households’ asset losses directly reduce their income, and as a result the tax they pay and the financial transfers they do to other households.21 Based on Eq. 3, it is easy to verify that the reduction in transfers is proportional to national asset losses, and these losses are fully diversified at the national level: L(t) sp ∆isp h ( t) = · ih (14) K We include explicit time dependence in Eq. 14 to indicate that social transfers re- cover to pre-disaster levels throughout the recovery and reconstruction process. Total asset losses L(t) is inclusive of all asset classes, irrespective of ownership (private, public, and other). These assets are rebuilt independently, and at different rates. Therefore, social transfers tend to recover even for households that are unable to recover their direct income losses through private asset reconstruction. Importantly, the fact that the assets used by households to generate an income have different ownerships introduces interactions across households, with each household benefiting from a rapid recovery of the others. For instance, poor people benefit from a more rapid recovery of asset-rich households, if it allows them to re-open shops and factories earlier, thereby protecting jobs and increasing income of workers. Similarly, a rapid recovery of taxpayers helps governments restore social transfers and rebuild public assets. In Eq. 15, we update Eq. 12 on the previous page to reflect the structure of ∆isp h . In Eq. 16, we show that the aggregate loss in income is equal to the average productivity of capital multiplied by the total asset losses. Note that the final term, ∆iPDS h , is 21 This is equivalent to assuming that the government budget is always balanced and that inter-household transfers respond instantaneously to income changes. Other changes in government spending, tax rates, and remittances are represented through the third term of the equation, ∆iPDS h , see below. — 52 Disaster c0 occurs Income Area = lost productivity of losses destroyed assets Household consumption ch Area = total value of destroyed assets Reconstruction costs Area = total value of savings + PDS Savings + PDS expenditure Household consumption -1 t0 1 2 3 4 Time after disaster [years] Figure 15: After being affected by a disaster, each household reconstructs its assets at the op- timal rate, while avoiding falling below the subsistence line, as described in Sec. 6 on page 54. Here, we illustrate consumption losses through the reconstruction process for a household that suffers ∆keff 0 = ∆kprv 0 + ∆kpub 0 in losses at time t=to , and reconstructs with period τh = 2.1 years. omitted in Eq. 16, since we have not yet discussed its funding mechanism, but costs and revenue sum to zero for all PDS systems independently. L sp ∆ih (t) = (1 − δtax eff sp ) · Πk · ∆kh + · i + ∆iPDS h (15) K h Nh Nh L sp ωh ∆ih = ωh (1 − δtax eff sp ) · Πk ∆kh + i = Πk L (16) K h h=0 h=0 — 53 Consumption losses After their capital has been diminished by a disaster, households are able to generate less income and, therefore, can sustain a lower rate of consumption.22 Ideally, this decrease in income and consumption is not permanent, as households usually repair the damages to their dwelling, replace lost assets such as fridges and furniture, and rebuild their asset base (for instance re-growing their livestock). Because assets do not rebuild themselves, affected households will also have to forego an additional portion of their income (∆creco h ) to fund their recovery and re- construction.23 Total consumption losses, then, are equal to income losses plus recon- struction costs, less savings and post-disaster support (together represented by Sh ), as indicated by Eq. 17): ∆ch = ∆ih + ∆creco h − Sh (17) Total reconstruction costs are equal to the reduction in consumption needed to re- build their asset stock, plus the increase in taxes needed for the government to rebuild public assets such as roads and water infrastructure. The contribution of reconstruc- tion costs to consumption losses at each moment depends on the ownership of the damaged assets, and on reconstruction rate. These two dimensions will be discussed next. Consumption losses due to reconstruction costs vary by asset type (i.e., private, public, or other): 1. Affected households pay directly and entirely the replacement of the lost assets that they owned (∆kprv ). 2. All households pay indirectly and proportionally to their income for the replace- ment of lost public assets through an extraordinary tax (∆kpub ) 3. Households do not pay for the replacement of the assets they use to generate an income but do not own (such as the factory where they work; ∆koth ). 22 In addition to the loss of monetary income, this includes the loss in virtual income if the housing services provided by their home or their asset (fridge, fans, air conditioning systems) are also lost. 23 Even though natural capital "rebuilds itself,“ people may have to reduce consumption to allow for accelerated growth. — 54 Private asset reconstruction In the event of a disaster, affected households lose productive assets, which directly reduces their income. Household-level consumption losses do not end there, how- ever, as the destroyed assets do not rebuild themselves. Rather, affected households will have to increase their savings rate–that is, avoid consuming some fraction of their post-disaster income–to recover these assets. Assuming each household pur- sues an exponential asset reconstruction pathway, we calculate a reconstruction rate for each household that maximizes its wellbeing over the 10 years following the dis- aster while avoiding bringing consumption below the subsistence level (if possible). If the households cannot avoid having consumption below the subsistence line (for instance because consumption is below the subsistence level even without repairing and replacing lost assets), then we assume that reconstruction takes place at the pace possible with a saving rate equal to the average saving rate of people living at or below subsistence level in Sri Lanka (according to the HIES). To model each household’s recovery, we assume that disaster-affected households rebuild their lost assets exponentially over some number of years (τh ) after the shock, where τh specifies the number of years each household takes to recover 95% of initial asset losses. τh is related to reconstruction rate λh as follows: 1 τh = ln · λ− h 1 (18) 0.05 Given these assumptions for the response of each affected household to a disaster, the asset losses at time t after a disaster (occurring at time to ) are given by: −λh ·t ∆keff eff h → ∆kh (t) = ∆kh · e (19) In order to rebuild at this rate, the reconstruction costs to household consumption are given by: d ∆creco h ( t) = − ∆kh (t) = λh · ∆kh (t) (20) dt — 55 In the above equation, we have introduced a negative sign in order to keep this con- tribution to consumption losses positive, in accordance with our convention. To calculate for each household a reconstruction rate that maximizes its wellbeing, we plug Eq. 17 on page 53 into the canonical definition of wellbeing: 1 1−η W= × ch − ∆ch (t) · e−ρt dt (21) 1−η Expanding these terms (and omitting social transfers and taxes for simplicity), we arrive at the following equation, where λh is the optimal reconstruction rate for each household: keff 10 1−η W= h × Π − (Π + λh ) · ve−λh t · e−ρt dt (22) 1−η t=0 This integral cannot be solved analytically, but we know that each household will ∂W maximize its wellbeing if it chooses a reconstruction rate (λh ) such that ∂λ = 0: 10 −η ∂W =0= Π − (Π + λh ) · ve−λt t(Π + λh ) − 1 · e−t(ρ+λ) dt (23) ∂λ t=0 We use this expression to determine the value of λh numerically. We note again that the optimum depends only on productivity of capital (Π), asset vulnerability (v), and future discount rate (ρ), while dependence on initial assets and absolute losses has dropped out of the expression. Fig. 16 on the following page displays the full distribution of reconstruction times τh , for two 25-year flooding events, affecting alternately the greater Colombo and Puttalam districts in Sri Lanka. In Colombo, affected households take 43 months to recover on average, while the average recovery time stretches 5 additional months in Puttalam (even though the average per capita asset loss is 30% lower than in the capital). This result is consistent with the idea of a poverty trap (see, e.g, [6]) and with the observation that poor households sometimes need a long time to get back to their pre-disaster situation [16, 17, 18]. It may also provide a theoretical explanation for the long-term consequences on incomes and growth that have been identified in the literature [19]. — 56 Event: 25-year precipitation flood 35 Region : Colombo Mean per cap asset loss ( k eff): US$607 Mean recovery time ( ): 43 months Fraction of affected households in district [%] 30 Region : Puttalam Mean per cap asset loss ( k eff): US$428 25 Mean recovery time ( ): 48 months 20 15 10 5 0 0 2 4 6 8 10 12 14 Optimal recovery time ( h) [years] Figure 16: Optimal reconstruction time distribution and mean for disaster-affected house- holds in Colombo (red histogram) and Puttalam (blue) districts after a 25-year flooding event. Reconstruction time is calculated for each household after each disaster, and parameterizes the disaster recovery pathway that minimizes wellbe- ing losses individually for each household. In addition, households must maintain consumption above a certain level to meet their essential needs. To reflect this, we use the following heuristic: if a household cannot afford to reconstruct at the optimal rate without falling into subsistence (i.e. if ih − ∆ih − λh ∆keff h < isub ), then the household reduces its consumption to t=t0 the subsistence line less the district-level savings rate for households in subsistence (Rsub sav ), and uses the balance of its post-disaster income to reconstruct. Its consump- tion remains at this level until its reconstruction rate reaches the optimum. This leads 1 to an initial reconstruction rate equal to λh = · ih − ∆ih − isub + Rsub sav . ∆kpriv h t=t0 — 57 Public asset reconstruction When disasters occur, we assume that the government borrows externally to finance the cost of public asset reconstruction, in order to speed recovery and minimize the financial burden on affected households. Eventually, the government recovers these costs through a tax, but only when recovery is complete. Through this mechanism, all households throughout the country share the cost of public asset reconstruction in the affected area. sp ωh vh kpub ih · δtax pub = Πk keff tax h (1 − δsp ) + ih · h (24) K pre−disaster income fractional public losses Note that Eq. 24 represents a distribution of the costs of rebuilding public assets to both affected and non-affected households. Like the tax to fund social protection programs, it should be understood as a universal tax with a flat rate (δtax pub ) that is by construction proportional to pre-disaster household income. Because this is conceived as a one-time tax to fund reconstruction, public asset re- construction costs are not spread across the duration of the reconstruction (in contrast to income lost due to the destruction of public assets, which does last for years after the disaster). Therefore, all of the time-dependence has been eliminated in Eq. 24, which indicates our assumption that the government does not collect the special tax at any point during recovery, but rather covers the cost of public asset reconstruction for the duration of reconstruction and collects taxes to fund this process many years later, after full recovery. Savings and post-disaster support The final term in Eq. 17 on page 53, Sh , represents households’ precautionary sav- ings, increased by post-disaster support, potentially including cash transfers to af- fected households, increases in social protection transfers, help through informal mechanisms at the community level, potential increases in remittances, and other exceptional cash transfers to households following a disaster. When available, these resources help households to smooth their consumption over time, or decrease con- sumption losses. — 58 In Section 6 on the preceding page, we discussed one example of a post-disaster support system in which all Sri Lankans affected by a disaster receive a uniform pay- out, which is equal to 80% of the asset losses suffered by the poorest quintile. In this and other systems, the value of this PDS in included in Sh , along with any savings the household may have had before the disaster. In more realistic applications, these ben- efits can also accrue to non-affected households: for example, due to targeting errors in post-disaster support systems. Like the cost of public asset reconstruction, the cost of all exceptional post-disaster transfers is distributed among all households, includ- ing those in unaffected districts, long after reconstruction is complete (cf. Sec. 6 on the previous page). Similarly, the rebuilding of the savings of affected households is assumed to take place far in the future, when reconstruction is complete and affected households’ incomes are back to their pre-disaster levels. Optimal consumption of savings and post-disaster support As illustrated by the gray shaded region in Fig. 15 on page 52, each household uses its savings, plus any post-disaster support it receives to smooth its consumption. More specifically, each household spends its liquid assets to establish a floor, or offset the deepest part of its consumption losses. The floor each household is able to afford is a function of the value of its income losses, savings and post-disaster support, and reconstruction rate, as well as of the average productivity of capital. Having assumed an exponential recovery with rate λh and a total value of savings Stot h , we can determine the level of this floor (γ) and the time at which the household’s savings ˆ ) by solving the following coupled equations: are exhausted (t keff h vh ˆ Stot ˆ h + γt = Π + λh 1 − e−λh t (25) λh ˆ γ = keff h Π − v h ( Π + λh ) e −λh t (26) — 59 As with the reconstruction rate, this optimization cannot be completed in a closed form without resorting to series expansions, so we combine Eqs. 25 and 26 into Eq. 27, and numerically find the value of γ that satisfies this equation: 0 = keff tot h vh Π + λh 1 − β + γln β − λh Sh (27) where β = γ · (keff h vh (Π + λh )) −1 Importantly, we assume here that the provision of PDS and savings do not accel- erate reconstruction pace, since the utilization of these resources is determined only after the rate of reconstruction is determined. This is a simplification that allows the two questions (rate of reconstruction and utilization of savings and PDS) to be solved sequentially, making it easier to solve the model. Wellbeing losses A $10 reduction in consumption affecting a rich household does not impact welfare or threaten health and wellbeing as much as the same loss would affect a poor house- hold. Welfare economics theory quantifies this difference by evaluating the utility (w) derived from a given level of consumption. Here, we use a simple constant relative risk aversion (CRRA) utility function: c1−η w= (28) 1−η The value of η, representing the elasticity of the marginal utility of consumption, is important to the modeling of the wellbeing losses; it represents both the risk aversion and the aversion to inequality in a society and is linked to preferences and values. It describes how $1 in consumption loss affects differently poor and non-poor people. Implicitly, it sets distributional weights, i.e. the weight attributed to poor people vs. the rest of the population in the aggregation of costs and benefits in an economic analysis [20]. In this study, we use a standard value of 1.5. Higher values give more importance to poor people, lead to higher estimates of wellbeing losses, and make it — 60 relatively more important to use policy instruments targeted towards poor people to reduce wellbeing risks. In order to determine the wellbeing losses that accumulate to each disaster-affected household during the reconstruction period (defined as the time to return consump- tion is back to its pre-disaster level), we calculate wellbeing as the future-discounted time integral over 10 years after a disaster. c1 −η ∞ 1−η ho ∆ch (t) λh t ∆Wh = 1− e − 1 e−ρt dt (29) 1−η 0 coh Note that the integral evaluates to 0 when ∆ch = 0. For all other values of 0 < ∆ ch < co h , Eq. 29 has to be evaluated numerically. To balance the need for precision with computational limitations, Eq. 29 is evaluated within the model with tmax = 10 years and dt = 1 week. tmax reconstruction c1 ho −η ∆ch (t) −λh t 1−η ∆Wh = dt × 1− e − 1 e−ρt (30) 1−η coh t=0 We have assumed that the reconstitution of household savings and the taxes that fund public asset reconstruction and post-disaster support are widely distributed and far in the future, so that they reduce consumption but only after reconstruction is complete. Therefore, we assume that the wellbeing impact of using savings and PDS-related taxes can be estimated using the marginal utility of consumption of each household: long−term ∂W ∆Wh = ∆c = c− h η × ih · δtax tot pub + ∆Sh (31) ∂c Total wellbeing losses one household are equal to the sum of the loss along the reconstruction path and the long-term losses: reconstruction long−term ∆Wh = ∆Wh + ∆Wh (32) — 61 Then, the total wellbeing losses are calculated by summing over all households, using the number of individuals in the household as weight: Nh ∆W = ωh ∆Wh (33) h=0 Finally, we translate ∆W from an expression of utility back into an “equivalent consumption loss” (∆Ceq ) by determining the value of the consumption loss that an imaginary individual earning the national mean income would have to suffer in order to experience wellbeing losses equivalent to each "real" individual’s losses. This final step allows us to express wellbeing losses, like asset losses, in currency unit and as a percentage of national or district-level GDP. We derive ∆Ceq as follows: ∆W ∆Ceq = (34) W where: ∂W ∂ c1−η W = = = c−η avg. (35) ∂c ∂c 1 − η cavg. cavg. The result ∆Ceq is the metric we use to measure the wellbeing impact of a disaster (or of risk) on the population. It is a measure — expressed in domestic currency — of the wellbeing loss due to a disaster. If a disaster causes = P1 in wellbeing losses, it means that its wellbeing impact is equivalent to a = P1 decrease in the consumption of the average Sri Lankan (i.e. an hypothetical individual with a consumption level equal to the average consumption in Sri Lanka).