Policy Research Working Paper 10217 The True Cost of War Erhan Artuc Nicolas Gomez-Parra Harun Onder Development Research Group & Macroeconomics, Trade and Investment Global Practice October 2022 Policy Research Working Paper 10217 Abstract Measuring the economic impact of a war is a daunting task. by voting with their feet, using pre-conflict data. Then, it Common indicators like casualties, infrastructure damages, infers a lower-bound estimate for the conflict-driven wel- and gross domestic product effects provide useful bench- fare shock from partially observed post-conflict migration marks, but they fail to capture the complex welfare effects patterns. A case study of the conflict in Eastern Ukraine of wars. This paper proposes a new method to estimate between 2014 and 2019 shows a large lower-bound welfare the welfare impact of conflicts and remedy common data loss for Donetsk residents equivalent to between 7.28 and constraints in conflict-affected environments. The method 24.79 percent of life-time income depending on agents’ first estimates how agents regard spatial welfare differentials time preferences. This paper is a product of the Development Research Group, Development Economics and the Macroeconomics, Trade and Investment Global Practice. 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 eartuc@worldbank.org and honder@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 The True Cost of War∗ † ‡ Erhan Artuc, Nicolas Gomez-Parra, Harun Onder§ Originally published in the Policy Research Working Paper Series on October 2020. This version is updated on November 2022. To obtain the originally published version, please email prwp@worldbank.org. Key words: Conflict, revealed-preferences, internally displaced people JEL codes: D74, J61, I131 ∗ Previously circulated as “The Economic Impact of War”. This paper has been partly supported by the Umbrella Facility for Trade trust fund and the World Bank Knowledge for Change Program (KCP). 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 of Reconstruction and Development, the World Bank, or the Inter-American Development Bank, and their affiliated organizations or those of the Executive Directors of the World Bank, or the Inter-American Development Bank, or the countries they represent. We thank Olga Kupets for access to pre-conflict data. We are grateful to Arup Banerji, Caglar Ozden, Bob Rijkers and Daria Taglioni for their comments and support. All errors are our responsibility. † Development Economics Research Group, World Bank, Washington, DC ‡ Inter-American Development Bank, Washington, DC § Macroeconomics, Trade & Investment Global Practice, World Bank, Washington, DC 1 Introduction Wars have complex economic and social consequences. Conflict-driven deaths, physical destruction, and economic disorganization levy unambiguously heavy tolls on societies. However, accounting for the whole burden of wars on human well-being is challenging. On the one hand, it is generally difficult to measure how the intangible fallout of wars, including institutional degradation, erosion of social trust, and eruption of psycho-social trauma, can affect economies over time. On the other, even otherwise measurable factors, including common economic indicators like prices, employment, and trade, may not be recorded accurately in times of conflict, rendering a systematic assessment difficult. With these challenges, economic impact assessments of wars often reflect only a subset of the broad and persistent misery engendered by large-scale human violence. To better account for the complex welfare impacts of wars, this paper proposes an alternative approach that relies on a general and flexible migration model, where economic agents choose between regions with different characteristics. The model is estimated using gross bilateral migration patterns before the onset of the conflict. This helps first to establish how economic agents regard spatial welfare differentials by voting with their feet. Next, the estimated model is used to infer the magnitude of the conflict-driven welfare shocks from partially-observed migration patterns after the onset of the conflict. Thus, an increase in migration outflows, together with the estimated responsiveness of agents to welfare differentials across regions, yield a measure of how economic agents perceive the conflict-driven welfare shock. This approach has several desirable properties. First, it relies on the revealed preferences of economic agents to account for what matters in assessing the welfare impact of wars. A preference-based welfare concept better captures the intangible consequences of wars (e.g., cultural effects or trauma from physical and sexual violence) as it covers both pecuniary and non-pecuniary factors affecting individual well-being. Second, because the analyst abstains from specifying the composition of welfare–thus, effectively delegating the model selection problem largely to economic agents themselves, data constraints are significantly relaxed in our approach. Collecting and publishing reliable data is especially difficult during a war. As purportedly coined by the US Senator Hiram Warren Johnson in 1918, “the first casualty when war comes is truth”. While many socioeconomic indicators suffer from this problem during war, our approach relies on migration away from conflict, which is typically well-recorded by humanitarian organizations for coordinating assistance. Importantly, our method only requires data from a subset of potential migration destinations, not all of them. 2 It is possible to omit some destinations, international or local, to overcome data collection issues or other empirical problems. Overall, the ability to ease a pervasive model-selection problem and circumvent potentially prohibitive data constraints help to achieve a more complete inference about the welfare impact of wars. To show these points formally, we first set up a general discrete choice framework with random utility, where agents choose among a finite number of locations. This decision is guided by (i) location-specific fixed utility common to all agents in a given location, (ii) an individual-specific utility for each location, and (iii) bilateral mobility costs associated with migration. For generality, the framework remains agnostic about the components of location-specific fixed utility, which can include wages, local amenities and, potentially, expected future utility, among others. We do not impose any strong assumptions on whether the people are myopic or forward-looking, how much they discount the future outcomes, or how they form their expectations about the future after conflict. Then, we derive sufficient statistics equations and mark the lower-bound estimates for the welfare impact of a war. To estimate the main parameters of the model, we consider a case study: the conflict in Eastern Ukraine (i.e., the Donbas Region, covering Donetsk and Luhansk oblasts) between 2014 and 20191 . Specifically, we estimate the inter-region migration elasticity parameter using flows among Ukrainian regions prior to the conflict in 2014. This estimate is then used to back out the welfare impact of the war in Donbas as revealed by the flows of about 1.4 million internally displaced people (IDPs) after 2014. Our data includes yearly regional statistics and bilateral gross migration flows from the State Statistic Services (UkrStat) for 2008-2012 and IDP numbers from the Ukrainian Ministry of Social Policy for 2014-2019. Our empirical analysis yields a migration elasticity parameter between 0.458 and 0.682 depending on the agents’ risk aversion and time preferences. Using these estimates, and following an inversion equation similar to Hotz and Miller, 1993, we then map moving probabilities onto welfare, and compute the welfare shocks implied by the post-conflict migration outflows from Eastern Ukraine. Finally, following the literature, we compute the equivalent income losses implied by these welfare shocks. Specifically, we answer the following question: what would be the rate of income loss that makes an average individual equally worse off as the conflict? Our estimates for Donetsk oblast show a 7.28 to 24.79 percent equivalent life-time income loss, depending on time preferences, vis-a-vis the average pre-conflict income when agents are risk neutral. This range is equivalent to 27.72 to 38.06 percent income-loss for a duration of 10 years. The magnitude of welfare loss is similar for 1 Note that the data used in this paper covers a period before the war in 2022, which was not foreseen at the time of the analysis. 3 the Luhansk oblast. These estimates are driven by the (ex ante) perceived welfare shocks by economic agents, which trigger migration decisions, and they should be interpreted as the lower bounds of the welfare impact. When conflicts affect welfare in non-conflict areas significantly or boost mortality significantly,2 thereby distorting migration numbers, the actual impact can be larger. To our knowledge, this paper is the first to estimate the economic impact of war by using preference-based approaches inferred from migration patterns. There is a large and growing body of research focusing on the economic and social impact of wars.3 This includes a strand that uses synthetic control methods to estimate the GDP impact of war, starting with Abadie and Gardeazabal, 2003.4 There is also a diverse body of research that uses various methods of inference to measure conflict-driven impact on health and experienced well-being (Clark et al., 2020 and Bendavid et al., 2021), on macroeconomic indicators like GDP, investments, and fiscal flows (Edwards, 2014, and Auray and Eyquem, 2019), and on international trade flows (Glick and Taylor, 2010). Among the latter group, Korovkin and Makarin, 2019 estimate the impact of the conflict in Eastern Ukraine on Ukrainian firms’ trade with the Russian Federation between 2013 and 2016 by exploiting the spatial variation of pre-conflict Russian-speaking population as a proxy in a difference-in-difference setting. Unlike our approach, however, these studies limit attention to a subset of welfare components c et al., 2010 estimate – often pecuniary ones only. In the international trade literature, Artu¸ the welfare generated by mobility using labor flows, and Arkolakis et al., 2012 calculate income gains from trade using trade flows. Both studies use inversion equations similar to Hotz and Miller, 1993, that are related to the inversion equation in our approach. Different from them, our method is implemented using partially observed outflow data, instead of stayer data, which may be unobserved or unreliable in conflict environments. This paper proceeds as follows. Section 2 lays out the discrete-choice framework and the theory of the welfare impact analysis. Section 3 introduces the Eastern Ukraine case study. Section 4 presents the estimates for migration elasticities, welfare implications, and the equivalent income shock from the conflict. Section 5 discusses the key findings and caveats. The last section concludes. 2 Onder et al., 2019 discuss the conditions for analyzing the income equivalent of the mortality driven decreases in statistical life spans, with an application to the Syrian conflict. 3 For an excellent review of the economic causes and consequences of wars, see Blattman and Miguel, 2010, and for a comprehensive review of recent advances in understanding the long term implications of exposure to war on human behavior, with an emphasis on cooperative social behavior, see Bauer et al., 2016 4 For a detailed overview of this approach, including the conditions shaping its feasibility, see Abadie, 2021. 4 2 Model Consider an economy with K locations. In this economy, agents can choose new locations in every period based on their preferences. For an agent i located in location k , the total i,l utility associated with moving to location l can be expressed as Utl − Ctkl + zt , where Utl is the location-specific fixed utility for all agents in l, Ctkl is the cost of moving, which is equal i,l to zero for stayers, and zt is a random utility associated with location l specific to agent i. Agents are homogeneous except for the random utility shock. Individuals choose the utility maximizing location l∗ such that i,l l∗ = arg max Utl − Ctkl + zt . l We are agnostic about the components of the location-specific fixed utility Utk , which can include wages, local amenities, and, possibly, expected future utility. After a shock, like a conflict, agents can establish expectations about the future or they might be myopic. That is, they may make decisions on the basis of outcomes in a given period in a static environment or they may consider a stream of future outcomes and their present discounted values in a dynamic environment. Therefore, we do not impose a dynamic or static structure on the location-specific utility, or any strong assumptions about the risk perception of agents; they can be risk-neutral or risk-averse. Although the location-specific fixed utility is the same for everyone, agents’ experiences can vary due to individual-specific random utility shocks and the moving costs they face. To calculate the expected utility, an agent in location k must consider values from each potential decision from her optimization problem. Formally, the welfare, Wtk , is defined as: i,l Wtk ≡ Ez max Utl − Ctkl + zt , l where the expectation is taken over the random shocks before their realization. Assuming i,l that zt is drawn from a Gumbel distribution with scale parameter 1/θ, we get a closed 1 θ form solution for the expected total utility: Wtk = θ log l exp Utl − C kl , where l ∈ {1, 2, ..., K }. The expected total utility, Wtk , also gives a measure of welfare in this framework. The distributional assumption ensures tractability of the optimization problem similar to a multinomial-logit model, and the expression for the moving probability from k to l becomes: θ exp Utl − Ctkl mkl t = , (1) exp Wtk 5 where 0 < mkl t < 1. This equation will help us to map migration flows onto welfare under certain conditions even when such flows are only partially observed. We will refer to θ as the migration elasticity parameter, since it determines the mobility of agents in response to value differentials. Introducing conflict: With the onset of the conflict, location-specific utilities change and some agents respond to this shock by relocating. Imagine that conflict occurs in some of the regions with varying intensities. Let us denote variables at an arbitrary reference time before the conflict as mkl k k kl 0 , U0 , W0 and C0 . We use ∆ operator to denote change after the conflict such that ∆xt = xt − x0 for any variable xt . Thus, the flow equation (1) can be expressed as ∆ log mkl l kl k t = θ ∆Ut − θ ∆Ct − θ ∆Wt . (2) After subtracting the flow equation for movers from k to l from the flow equation for the stayers in k , and rearranging the terms, we get an expression for the fixed utility in k as 1 ∆Utk = ∆ log mkk kl t − ∆ log mt + ∆Utl − ∆Ctkl . θ This expression implies that it is possible to back out fixed utility in k using only (i) fixed utility in l net of moving costs ∆Utl − ∆Ctkl , (ii) flows to l and stayers in k ∆ log mkk kl t − ∆ log mt , and (iii) parameter θ . We would not need any other information related to other destinations. Note that this result directly follows from the properties of the discrete choice optimization problem given the distributional assumptions, as the expected maximized total utility associated with choices, conditional on choosing them, are equalized. While it is empirically desirable to use as many destinations as possible, it is feasible and practical to omit some destinations when such flows are unobserved.5 Hence, summing this expression for a subset of destinations, Φ, after multiplying with arbitrary weights yields a general expression for the utility: 1 ∆Utk = ϕl ∆Utl − ∆Ctkl + ϕl ∆ log mkk kl t − ∆ log mt , (3) l ∈Φ θ l ∈Φ where weights ϕl ≥ 0 and l ∈Φ ϕl = 1. Equation (3) comprises two economically meaningful terms: The first term on the right hand side, l ∈Φ ϕl ∆Utl − ∆Ctkl , accounts for the average loss in regions in Φ (net of the 5 This convenient feature of our method gives an advantage over backing out utility using gravity c and McLaren, 2015, when flows to some destinations are not observed. regressions, such as Artu¸ 6 changes in moving costs and weighted by ϕl ), and the second term on the right hand side, 1 θ l ∈Φ ϕl ∆ log mkk kl t − ∆ log mt , accounts for the additional loss specific to location k . These two terms, together, represent the total change in fixed utility associated with k . Following similar steps, we derive an equation for the welfare: 1 ∆Wtk = ϕl ∆Utl − ∆Ctkl + ϕl −∆ log mkl t . (4) l ∈Φ θ l ∈Φ To map the observed changes in migration patterns onto changes in location-specific utility and welfare, we need a restriction about the change in net utility, which we call the “war is not good for some destinations” assumption: Suppose there exists a subset Φ of destinations and weights ϕl ≥ 0 for l ∈ Φ such that war does not increase the average weighted utility in Φ net of moving cost, thus l∈Φ ϕl ∆ Utl − Ctkl ≤ 0, where k is a conflict location. If the utility in a non-conflict location l net of the moving cost from k to l, increases with the war, then it would be welfare improving for those who were planning to move to from k to l, i.e. war would be good for them. Our restriction implies that this should not be the case on average for a non-empty subset of destinations. Obviously, all non-conflict destinations will become more attractive relatively, i.e., because the conflict locations become less attractive in comparison. But not all of them should become more attractive in absolute terms compared to pre-war. We discuss the practical implications of this assumption in the coming sections. We can now analyze the upper bounds of changes in utility and welfare, i.e. the lower bound of a negative impact, by using information on post-conflict migration. The following propositions establish this formally. Proposition 1 The upper bound of change in location-specific fixed utility in conflict location k (i.e., the lower bound of the impact) can be expressed as 1 ∆Utk ≤ ϕl ∆ log mkk kl t − ∆ log mt , θ l∈Φ if the condition l∈Φ ϕl ∆ Utl − Ctkl ≤ 0 holds, where ϕl ≥ 0 are weights for destinations such that l∈Φ ϕl = 1. Intuitively, this proposition suggests that the upper bound of location-specific utility change can be calculated by using only the changes in flows of agents from conflict locations to a subset of destinations, mkl t , and the migration elasticity parameter θ . When expressed in logs, the upper bound of a utility change in a conflict location is proportional to the change 7 in the probability of staying minus the probability of leaving. To characterize this, we do not need the flow data between non-conflict locations or inflows to conflict locations, which may not be captured in the absence of a humanitarian situation. However the number of internally displaced people, i.e. flows out of conflict regions, is often well documented by the international organizations to program assistance. Note that the calculated utility change in Proposition 1 is an upper bound rather than a precise point estimate because it is missing the term l ∈Φ ϕl ∆Utl − ∆Ctkl from equation (3), which accounts for the average utility loss in all regions weighted by ϕl , net of the changes in moving costs. As long as this term is non-positive, Proposition 1 will hold. The next proposition extends this result to changes in welfare. Proposition 2 The upper bound of change in welfare in conflict location k (i.e., the lower bound of the welfare impact) can be expressed as 1 ∆Wtk ≤ ϕl −∆ log mkl t , θ l ∈Φ if the condition l∈Φ ϕl ∆ Utl − Ctkl ≤ 0 holds, where ϕl ≥ 0 are weights for destinations such that l∈Φ ϕl = 1. Similar to the previous proposition, Proposition 2 only needs the flow of agents out of conflict locations and the migration elasticity parameter to calculate the upper bound of welfare changes. The expression in Proposition 2 calculates the upper bounds of welfare change (lower bound of the welfare impact of the conflict) rather than the location-specific utility change, which takes the possibility of moving to other locations into account. If agents are not likely to stay in the conflict region after the conflict, then the impact on their expected welfare will be smaller. Therefore, the change in expected welfare is a better measure of the impact of conflict on people compared to the location-specific fixed utility. 3 Case Study: The Conflict in Eastern Ukraine In this section, we take our model to data using the conflict in Eastern Ukraine between 2014 and 2019. A few characteristics of this conflict make such an application feasible. First, interregional migration patterns were documented in detail before the conflict. Second, the conflict was regionally contained for the period covered in this analysis and the war in 2022 (3 years after the last data in our analysis) was not foreseen. Third, with a relatively low-intensity conflict, demographic mobility remained feasible throughout the period of 8 analysis, with many people crossing the contact line routinely for daily needs. With these characteristics, we can implement the procedure developed in the previous section. 3.1 Background Ukraine’s Eastern regions have historically been home to the country’s industrial core, including coal mining, metallurgy, and chemical industries. Before World War I, these regions produced more than three quarters of the pig iron and coal output of the Russian Empire. Under the Soviet industrialization programs, Donbas became the most heavily settled region of Ukraine, attracting people from elsewhere in Ukraine and from other parts of the union.6 After the collapse of the Soviet Union, however, the region’s industrial infrastructure saw very limited modernization. To offset the eroding competitiveness, the industry was increasingly granted subsidies in key inputs like electricity, gas, and coal. Nonetheless, with dissipating favorable external conditions after the Global Financial Crisis in 2008, which include a slowing demand for steel and increasing modern production capacity in previous export markets of Ukraine, and as the subsidy component of energy inputs shrank after disputes with Russia, the challenges faced by the region’s aging economy grew. By 2013, Donetsk and Luhansk still had relatively larger populations and economies than other regions. However, they were losing people to other parts of the country (net out-migration), partially because of social challenges (i.e., high crime rates, alcoholism, and drug abuse) and environmental problems (pollution from mining and industry). The onset of the conflict in 2014 changed the economic and demographic characteristics of these regions dramatically. Before the onset of the war in 2022, about 38 percent of the combined territories of Donetsk and Luhansk oblasts were outside the Ukrainian government’s control (less than 4 percent of total Ukrainian land), which was demarcated by a 457 kilometer line of contact. This division imposed additional obstacles to economic activity, especially the provision and transportation of industry inputs and outputs like coal and steel. Despite the hostilities, however, migration was not prohibited. Ukrainians routinely crossed the contact line for family visits, shopping, and using ATMs for withdrawing pensions. Overall, there were about 1.4 million IDPs registered by the country’s Ministry of Social Policy between 2014 and 2019, close to half of whom remained displaced within the territories of Donetsk and Luhansk oblasts, and the other moved elsewhere within Ukraine. 6 For a detailed analysis of the social and economic conditions in Donbas before and after the onset of the conflict in 2014, see World Bank, 2021. 9 3.2 Data Our analysis considers all 25 administrative regions (oblasts) in Ukraine, including Kyiv city as a special region. For the purposes of this paper, the pre-conflict period (2008-2012) is defined as the 5-year period before the Maidan protests in 2013, and the post-conflict period is between 2014 and 2019 (the last year of our data). All regional statistics for the pre-conflict period, including population and average per capita income numbers, are based on the official information gathered by the State Statistics Services of Ukraine (UkrStat). Our data also includes gross bilateral migration flows between oblasts for the same period, which were reported in UkrStat’s demographic yearbooks. Information on these flows reflects the official registrations of residence recorded by the State Migration Service of Ukraine. Descriptive statistics show that Donetsk and Luhansk regions were among the most populous oblasts in Ukraine before the onset of the conflict in 2014 (Figure 1b). With 4.5 million and 2.3 million residents, they were ranked 1st and 7th among all regions, respectively. They also had relatively high real gross regional product per capita (2nd and 9th, respectively). Compared to other regions, however, migration flows into Donbas were relatively small. Incoming migrants constituted only 0.26 percent of local population in Donetsk, and 0.28 percent in Luhansk, putting them among the least attractive regions (24th and 25th among all) for migrant arrivals (1c). Among the most important sources of migrants were neighboring regions. Donetsk, for example received 22.8 percent of all its inflows from Kharkiv, 20.7 percent from Luhansk, and 12.3 percent from Dnipropetrovsk. Migration outflows from Donbas were also small. While the nation-wide interregional migration outflow rate stood at 0.56 percent of population annually, it was only 0.3 and 0.4 percent in Donetsk and Luhansk— the 2nd and the 7th lowest rates among all regions, respectively. Before the conflict, migrants leaving Donbas largely concentrated in neighboring regions and in Kyiv city. For example, of all migrants from Donetsk and Luhansk oblasts, on average, 31 percent went to Kharkiv every year, and 18 percent to Kyiv city (Figure 2). More distant or economically less attractive regions received fewer migrants from Donbas. 10 Figure 1: Regional statistics before conflict, 2008-12 annual averages (a) Real GRP per capita (b) Population Zaporizzhia Luhansk Odesa Kharkiv Luha Kyiv sa Kyiv nsk Vin My Ode hia kol nyt zz Ch a s iv Po iv ia ri Lv erk lta po as va Za y va iv lta Ivan ark Kiro Po o-F Kh voh ran r ad kivs sk 1,814,213 2,309,930 trov k City 1,72 14,146 14,864 2,5 5,713 rope 13,0 Kyiv 57 1,6 Dnip 12 1,4 9,64 3 15,7 11 ,10 51 8 Khmeln 07 ,77 96 2,38 ,83 Lviv ,45 4 ,76 ytsky ,91 39 32 ,03 8 7 10,7 2 15 ,1 1,38 60 3,49 1 3 08 17 2,7 0,234 26 Donetsk 2 ro vsk 10,629 18,7 1,337,91 7 2,77 Dnipropet 18,994 3,359,285 Chernihiv 10,251 44,697 Cherkasy 1,297,663 4,454,673 10,24 6 ,316 7,29 Kyiv City 1,294 904 Donetsk 75 8 030 1,0 ,640 9,5 8,0 48, 18 46 myr 1,2 ,98 1 Sumy ,22 Zhyto 1,0 0,425 4 9,3 1 048 6 1,0 8,1 94 ,618 9,4 1,104,462 1,097, 39 21 Che Che 8,8 1,1 74, 9,182 90 ,77 9 9,021 9,118 rnivts rnivts 1,160 18 1,1 3 i a i 922 yts ia rp atti Vinn Za kar Z aka Kir ovo pa hra ttia d v lai lyn ko vsk Vo Te My Vo nki rno ly my n pil Fra Khe Ter Khmeln Su Chernihiv Zhytomyr Kherso Rivne Rivne no- nop rso Iva n il ytsky n (c) Migration inflows (d) Migration outflows ( of destination population) ( of source population) Zhytomyr Zhytomyr Kharkiv Rivne Vinn Polt rad iv rnih voh ytsia ava Vin Po Che y y Kiro tsk tsk nyt lta lny va lny sia Ri Su me me vn my e hiv y Kh Kh i as ern erk Sum Ch Khe rson Ch y rkas y rad voh 0.58 0.65 0.60 0.67 Che Kiro 0.56 0.63 0.61 0.68 0.5 0.6 Volyn 0.5 Mykola 0.5 2 9 0.6 0.6 5 0 2 3 iv 9 7 0.49 0.6 0.59 0.7 0.48 0.73 Kyiv 0.57 0.82 Kyiv 1.18 0.94 Mykolaiv 0.47 1.64 Ternopil 0.53 1.18 0.46 0.20 Kyiv City 0.52 0.26 Kyiv City 4 9 0.4 0.2 0.4 0.3 on 6 iv 0 Khers Khark 4 0 0.4 0.4 0.2 0.3 Zaka Zaka 2 0 0.3 0.3 0.4 0.4 8 4 0.39 0.40 0.37 0.37 0.37 0.39 rpatt rpatt 2 5 ia ia esa Vo lyn Od Do Do ne ne tsk tsk ia il zh op riz rn sk Lu po Od Te rov tsi ha Za sk es niv ns Ivano-Frankivsk pet Dnipropetrovsk kiv a k Zapori nsk er Chern Lviv Lviv ipro -Fran Ch Luha Dn zzhia ivtsi Ivano Notes: Figures show annual averages by oblast for the pre–conflict period, 2008-2012, sorted in ascending order. Donetsk and Luhansk oblasts are represented in shades of gray, while other oblasts are shown by a range of colors transitioning from purple (lowest) to red (highest). 11 Figure 2: Migration outflows and forced displacement from Donbas, annual averages (a) Pre-conflict migration outflows (2008-12) (b) IDP outflows after the conflict (2014-19) Kharkiv (5,540) Donetsk (11,227) Kyiv City (3,197) Dnipropetrovsk (1,918) Kyiv City (30,743) Luhansk (6,720) Zaporizzhia (1,139) Kyiv (1,038) Odesa (825) Poltava (636) Sumy (432) Cherkasy (394) Donetsk (84,006) Vinnytsia (349) Kharkiv (26,354) Chernihiv (349) Kherson (325) Zhytomyr (254) Khmelnytsky (250) Mykolaiv (245) Kirovohrad (231) Dnipropetrovsk (14,003) Lviv (213) Ivano-Frankivsk (116) Volyn (112) Rivne (109) Zakarpattia (101) Kyiv (11,935) Chernivtsi (95) Ternopil (79) Zaporizzhia (11,064) Luhansk (43,912) Odesa (7,261) Poltava (4,492) Kherson (2,764) Sumy (2,229) Vinnytsia (2,203) Cherkasy (2,173) Lviv (2,148) Mykolaiv (1,631) Chernihiv (1,454) Zhytomyr (1,404) Kirovohrad (1,297) Khmelnytsky (1,223) Ivano-Frankivsk (743) Zakarpattia (664) Volyn (613) Rivne (612) Chernivtsi (489) Ternopil (419) Notes: Panel 2a shows gross migration outflows from Donetsk and Luhansk oblasts between 2008 and 2012 by destination, with average annual numbers in parentheses. Panel 2b does the same by using annualized official internal displacement numbers (IDPs) between 2014 and 2019. Like in other conflict situations, Ukraine faced major data constraints after the onset of the conflict. Most importantly, a complete record of the gross bilateral migration flows among Ukrainian regions is not available in the post-conflict period. Instead, we use information on the IDPs from Donbas as an indicator of the conflict-driven out-migration. By October 2019, the total number of officially registered IDPs were reported at 1,412,589 (21 percent of the combined population of Donetsk and Luhansk oblasts before the conflict). About 55 percent of all IDPs were registered within Donbas, while the rest was registered elsewhere in Ukraine. This outflow exhibited a pattern similar to the pre-conflict migration outflows, 12 occurring at a much larger scale (Figure 2–reporting annualized values). For IDPs fleeing the conflict in Donbas, Kyiv city (153,715), Kharkiv oblast (131,769), and Dnipropetrovsk oblast (70,012) were among the top destinations. However, other oblasts received significant numbers of IDPs from Donbas too.7 4 Quantification Our theoretical results in Section 2 do not rely on the specific components of the utility function or any model parameters apart from the migration elasticity parameter θ, which needs to be quantified. However, to estimate the migration elasticity parameter, we need to define the components of agents’ utility functions and time preferences. This specification will then allow us to assess the size of the conflict-driven welfare loss in pecuniary terms. 4.1 Characterizing the utility function In this subsection, we consider a structure for the utility function to estimate the migration elasticity parameter and map changes in welfare onto changes in income. To keep the analysis as general as possible, we adopt a flexible formulation for the location-specific utility Utk to allow for different risk aversion and time discounting parameters. Consider the following form for the location-specific utility: i,l k Utk = υ wt + η k + βEt Ez max Utl+1 − Ctkl +1 + zt+1 , l 1−σ k (wt k ) −1 k where the income component of the utility is υ wt = 1−σ and wt represents wages. The parameter η k is a location-specific, non-pecuniary, and non-random utility-shifter. The intertemporal discount factor is β , and the risk aversion parameter is σ . This specification provides ample generality and covers common labor mobility models with homogeneous agents in the literature. For example, setting β = 0 makes the agents myopic and the model static. If we also set σ = 1, the flow equation becomes isomorphic to the characterization of trade flows by Eaton and Kortum, 2002, subject to a log-transformation. When we set β > 0 and assume risk-neutral agents, i.e. σ = 0, the model becomes isomorphic c et al., 2010, with a minor modification by adding the utility shifter η k . It is to Artu¸ important to note that this framework is agnostic about how workers form their expectations about the future, i.e. the structure of expectation operator Et , and the components of the 7 Crimea is excluded from the analysis as Ukrainian sources ceased reporting data for it after 2014. 13 utility shifter η k . We next turn to estimating the migration elasticity parameter using this specification. 4.2 Estimating the migration elasticity parameter and welfare effects To estimate the migration elasticities in pre-conflict Ukraine, we follow the estimation c and McLaren, 2015. Equation (1) gives the number of people strategy suggested by Artu¸ moving from region k to region l. Multiplying this expression with the number of people located in k in the previous period, we get log mkl k l kl k k t Lt−1 = θUt − θCt + log Lt−1 − θWt , (5) where Lk t−1 is the number of people located in k at t − 1. This yields an equation that can be interpreted as a Poisson Pseudo Maximum Likelihood (PPML) regression equation: kl yt = exp al k t + bt + ct log δ kl + εkl 1t , (6) kl where yt is the number of people moving from k to l, δ kl is the distance between k and l, al t is the destination fixed effect, bk t is the origin fixed effect, ct is the time varying moving cost coefficient, and εkl 1t is a sampling error. In this specification, each coefficient has a structural interpretation where al l k k k kl kl t = θUt , bt = −θWt + log Lt−1 and ct log δ = −θCt . Next, we estimate the migration elasticity parameter θ using the following regression equation: k k αt = θυ wt + γ k + εk 2t , (7) k k where αt is constructed from first stage estimates such that αt = ak k k t − β −bt+1 + log Lt . We consider various values for the time discount factor β and the risk aversion parameter σ of 1−σ k (wk ) −1 the function υ wt = t 1−σ . The coefficient θ, which is the inverse of the Gumbel scale k parameter for the random utility shock zt , can be interpreted as the migration elasticity. The coefficient γ k is a fixed effect and it can be interpreted as the location specific utility shifter mentioned in the model γ k = θη k . Finally, εk 2t is the error term. c et al., 2010 and We use two period lagged wages as instruments in (7) following Artu¸ c and McLaren, 2015. Table 1 shows the estimates for the migration elasticity parameter Artu¸ (θ) under different time discount factors (β ) and degrees of risk aversion (σ ). Other things being equal, more patience (larger β ) and a lower risk aversion (smaller σ ) both increase the estimated migration elasticity. To see this, note that where a higher risk aversion is not the culprit behind the observed migration patterns, a higher migration elasticity should be. 14 Table 1: Estimates for migration elasticity parameter (θ) β = 0.97 β = 0.90 β=0 σ=0 0.682 0.667 0.479 (0.053) (0.063) (0.197) σ=1 0.612 0.601 0.458 (0.027) (0.035) (0.162) Notes: IV regression results based on equation (7). Standard errors in parentheses. σ denotes the degree of relative risk aversion in a CRRA utility, and β is the time discount factor. While θ is estimated at 0.458 with myopic (β = 0) and relatively risk averse (σ = 1) agents, it is estimated at 0.682 with relatively patient (β = 0.97) and risk neutral (σ = 0) agents. In one of the preferred specifications, where agents are risk averse and forward looking with β = 0.90, the implied income elasticity of emigration is 0.601, which means a 10 percent increase in income reduces emigration probability by 6.01 percent. Overall, this number, and our other estimates, are comparable with the previous estimates in the literature. For example, Kennan and Walker, 2011 report approximately 0.5 for the migration elasticity in an application to the United States. We next turn to estimating the welfare impact of the conflict. As discussed in proposition 2 earlier, an upper bound of the welfare changes due to the conflict, i.e., a lower bound of the welfare impact, can be backed out by using the estimated migration elasticities and the outflow of migration after the onset of the conflict. We observe the flows from two conflict locations to 23 non-conflict locations between 2014 and 2019. For this exercise, we include all non-conflict locations as destinations and include only conflict locations as origins. We set weights proportional to flows, i.e. ϕl ∝ mkl 0 . Table 2 shows the results of this exercise for Donetsk and Luhansk oblasts separately, using the common estimated migration elasticity parameter and oblast-specific IDP outflows. Note that factors that lead to a higher migration elasticity reduce the welfare impact as agents become more sensitive to shocks that take place in a certain location and evade them by migrating to other locations. Estimates show that a higher degree of risk aversion (greater σ ) and a more forward-looking time discounting (larger β ) both increase the estimated migration elasticity as shown in Table 1 and, therefore, reduce the estimated welfare impact as shown in Table 2. These estimates provide a lower bound for the adverse impact of the conflict. To see this, 15 Table 2: The estimated welfare impact of the conflict by region σ=0 σ=1 Donetsk Luhansk Donetsk Luhansk β = 0.97 -2.91 -2.60 -3.24 -2.90 (0.15) (0.29) (0.16) (0.32) β = 0.90 -2.97 -2.66 -3.30 -2.95 (0.15) (0.29) (0.16) (0.33) β=0 -4.14 -3.70 -4.33 -3.88 (0.21) (0.41) (0.22) (0.43) Notes: σ denotes the degree of relative risk aversion in a CRRA utility, and β is the time discount factor. Standard errors, in parentheses, are calculated using bootstrapped samples of destinations included in Φ repeated 5000 times. note that the term l ϕl∈Φ ∆Utl − ∆Ctkl in equation (4) is unobserved and missing from the expression provided in Proposition 2, but expected to be negative. Intuitively, this term defines the average impact of the conflict on location-specific utility in non-conflict regions within Φ, net of the changes in moving costs. Thus, unless non-conflict regions included in Φ benefit from the conflict in average terms, the adverse impact of the conflict should be at least as large as our estimates. Next, we will convert the welfare loss calculated in Table 2 to its equivalent monetary loss, which helps to interpret the welfare loss estimates in more conventional terms. 4.3 Estimating the equivalent income shock for the welfare loss We have so far discussed the welfare shock emanating from the conflict as inferred from migration numbers. This revealed-preference based approach has several desirable properties. One of these is the feasibility of the analysis under severe data constraints commonly observed in post-conflict environments. Another one is its ability to extend the impact assessment of conflict by going beyond the shocks to income (like GDP effects) and including non-monetary dimensions of well-being as embodied in individual preferences and manifested in migration data. It is possible to consider the changes in the non-monetary aspects of well-being in monetary terms. A large literature on equivalent income as a preference-based index of 16 well-being has characterized the axiomatic underpinnings of this conversion.8 Consider a conflict-driven hypothetical income loss suffered by the agents in the conflict region k just before the onset of the conflict. The loss prevails for T periods regardless of the location choice after the onset of the conflict, and could be explained by various factors like trauma and transition costs, among others. What would be the rate of income loss that makes the agents equally worse off as the conflict? To compute this, we need to specify the duration of the loss, T , risk aversion parameter, σ , and discount factor β . k Formally, let ψt < 0 be the income loss in period t in location k , with the new income k k k k k given as wt + ψt . We set the hypothetical income loss such that υ (wt + ψt ) − υ (wt ) = l l l υ (wt + ψt ) − υ (wt ) for every k and l, therefore the mobility decisions of agents are unaffected. For simplicity, we also abstract from any time variation by fixing the changes in instantaneous k k k k k k utility over periods, υ (wt + ψt ) − υ (wt ) = υ (wt+1 + ψt+1 ) − υ (wt+1 ). With these two simplifications, the problem is to identify the specific loss that continues k for T periods for an agent with wage w0 and changes her expected present discounted value, i.e.welfare, by ∆Wtk units. The welfare loss can be expressed as T −1 ∆Wtk = k β t υ w0 k + ψ0 k − υ w0 . t=0 which, with rearrangement, yields 1−β k ψ0 = υ −1 k ∆Wtk + υ w0 k − w0 , (8) 1 − βT k k where the percent change in income is 100 × (ψ0 /w0 ) for the initial period in location k . Note that the percent change in income is calculated relative to the income at t = 0. In the absence of a prior regarding the duration of the income shock for Eastern Ukraine, we consider alternatives values, T ∈ {1, 10, ∞}. Table 3 shows the income equivalent of the welfare shocks presented earlier in Table 2. The conflict driven welfare shock in Donetsk corresponds to a 7.28 percent income loss in every period vis-a-vis the average pre-conflict income when the loss is permanent (T = ∞) and agents are risk neutral (σ = 0) and very patient (β = 0.97). With less persistent shocks (smaller T ), the same group of agents, suffer 27.72 percent income loss for 10 years. A higher future discounting (lower β ) increases the income equivalent of the welfare shock. For instance, for T = 10 and σ = 1, the equivalent 8 Starting with the seminal work of Usher (1973), the equivalent income approach has gained increasing popularity in comparing well-being across countries and years. For a more detailed discussion on this, see Becker et al., 2005, Adler and Fleurbaey, 2016, and Onder et al., 2019. 17 Table 3: Income equivalent of the welfare loss, percent T =1 T = 10 T =∞ Donetsk Luhansk Donetsk Luhansk Donetsk Luhansk σ=0 β = 0.97 -242.61 -245.41 -27.72 -28.04 -7.28 -7.36 (11.93) (27.07) (1.36) (3.09) (0.36) (0.81) β = 0.90 -247.92 -250.79 -38.06 -38.50 -24.79 -25.08 (12.20) (27.66) (1.87) (4.25) (1.22) (2.77) β=0 -345.22 -349.21 - - - - (16.98) (38.52) - - - - σ=1 β = 0.97 -96.08 -94.48 -30.93 -28.18 -9.26 -8.32 (0.62) (1.69) (1.25) (2.60) (0.43) (0.88) β = 0.90 -96.31 -94.77 -39.74 -36.43 -28.10 -25.55 (0.59) (1.63) (1.49) (3.14) (1.16) (2.40) β=0 -98.69 -97.93 - - - - (0.28) (0.86) - - - - Notes: Values in percent of the welfare in the base-year (pre-conflict) welfare. σ is the degree of relative risk aversion in a CRRA utility, β is the time discount factor, and T denotes the number of years with income losses. Standard errors, in parentheses, are calculated using bootstrapped samples of destinations included in Φ repeated 5000 times. 18 income shock is 39.74 percent in Donetsk when β = 0.90 – about a third greater than the case with β = 0.97. 5 Discussion The economic impact estimates provided in this paper differ from more conventional metrics like GDP effects in several dimensions. In addition to including non-pecuniary aspects of welfare, they also reflect the agents’ expectations about the future in a rational expectations framework (while this assumes that agents do not make systematic errors, they are not required to have perfect foresight). These differences, combined with the missing GDP impact estimates for the conflict between 2014 and 2019, make it difficult to benchmark our estimates.9 An important feature of the analysis in this paper is its ability to infer lower bound estimates of the economic impact from migration flows into only a subset of destinations. For example, when flows abroad can not be measured accurately (as in the case of the war in Eastern Ukraine), flows within the country can provide sufficient information to conduct the assessment. This feature follows from the equalization of expected maximized utility associated with choices, conditional on choosing them, regardless of whether displacement to other destinations is welfare-enhancing ex post or not. This property also allows the exclusion of some regions when moving costs might decline significantly due to humanitarian concerns (e.g., border openings for refugee outflows), potentially threatening our “war is not good for some destinations” assumption. Although having such destinations in the sample set, Φ, does not break this assumption as long as they do not tilt the sign of the average across all regions, they can be excluded from the analysis to remove any doubts. A related concern in this case could be that some IDPs who register in non-conflict areas may be using these locations as stops en route to living abroad (i.e., transit migration), potentially decreasing the effectiveness of exclusion of foreign regions. To alleviate such concerns, we reproduce our estimations by (i) excluding Ukrainian regions on the country’s western border (which could provide a natural gateway into Europe), in addition to (ii), excluding regions with airports serving more than 100,000 passengers annually plus regions in 9 Regional GDP numbers produced by the Ukrainian Statistics Office (UkrStat) excluded the non-government controlled areas (NGCAs) in Donetsk and Luhansk oblasts starting from 2014. World Bank, 2021 uses nightlight emission levels as crude proxies for economic activity in the Donbas region. Estimates show a 7.2 percent reduction in Luhansk GCA, 42.9 percent reduction in Luhansk NGCA, 20.2 reduction in Donetsk GCA, and 28.1 percent reduction in Donetsk NGCA. 19 the western border, (iii) focusing only on the neighbors of Luhansk and Donetsk oblasts, (iv) focusing on the 10 largest regions in Ukraine, and (v) focusing on the 10 smallest regions. The results (presented in the appendix) are similar to our original estimates both qualitatively and quantitatively. Robustness tests suggest that a sharp decline in the utility in Donbas was the main factor causing a sudden but balanced increase in mobility to non-conflict regions within Ukraine, rather than destination specific positive shocks or an unbalanced decline in moving costs. Finally, like in other conflict affected environments, our data is not rich enough to explicitly consider ex ante heterogeneity in migrant profiles. Nevertheless, while we are unable to assess if there has been a shift in such profile after the conflict, we observe a strong similarity in the composition of migration destinations before and after the conflict (Figure 2). This is consistent with the idea that a pervasive shock to utility in Donbas is the primary driver of the post-conflict mobility, and it suggests no major discontinuity in the decision making processes regarding mobility. 6 Conclusion This paper proposes an alternative approach to assessing the welfare impact of wars. First, it shows that such welfare impact can be inferred from changes in migration outflows from the conflict-struck area by using a discrete-choice framework. Next, it applies this framework to the conflict in Eastern Ukraine between 2014 and 2019–long before the war in 2022. Empirical estimates yield migration elasticity parameters ranging between 0.458 and 0.682 depending on agents’ risk aversion and time preferences. Using these estimates, and the post-conflict migration patterns, the paper then estimates the implied welfare shocks and their income equivalent. Accordingly, the pre-conflict residents of Donetsk oblast are estimated to suffer at least 7.28 to 24.79 percent equivalent life-time income loss when agents are risk neutral. This income loss range is equivalent to 27.72 to 38.06 percent income-loss for a duration of 10 years. The estimates for the Luhansk oblast are of similar magnitude. The approach developed in this paper helps to avoid a pervasive model-selection problem, i.e., characterizing accurately the structure of individual welfare, including both pecuniary and non-pecuniary aspects. It also helps to circumvent potentially prohibitive data constraints, enabling a more complete inference about the welfare impact of wars. However, it is important to note that both the welfare impact estimates and the corresponding income equivalents discussed in this paper present lower bound estimates. The actual impact can be significantly greater. 20 References Abadie, A. (2021). Using synthetic controls: Feasibility, data requirements, and methodological aspects. Journal of Economic Literature, 59 (2), 391–425. Abadie, A., & Gardeazabal, J. (2003). The economic costs of conflict: A case study of the basque country. American Economic Review, 93 (1), 113–132. Adler, M. D., & Fleurbaey, M. (2016). The oxford handbook of well-being and public policy. Oxford University Press. Arkolakis, C., Costinot, A., & Rodr´ ıguez-Clare, A. (2012). New trade models, same old gains? American Economic Review, 102 (1), 94–130. Artu¸c, E., Chaudhuri, S., & McLaren, J. (2010). Trade shocks and labor adjustment: A structural empirical approach. American economic review, 100 (3), 1008–45. Artu¸c, E., & McLaren, J. (2015). Trade policy and wage inequality: A structural analysis with occupational and sectoral mobility. Journal of International Economics, 97 (2), 278–294. Auray, S., & Eyquem, A. (2019). Episodes of war and peace in an estimated open economy model. Journal of Economic Dynamics and Control, 105, 203–249. Bauer, M., Blattman, C., Chytilov´ a, J., Henrich, J., Miguel, E., & Mitts, T. (2016). Can war foster cooperation? Journal of Economic Perspectives, 30 (3), 249–74. Becker, G. S., Philipson, T. J., & Soares, R. R. (2005). The quantity and quality of life and the evolution of world inequality. American Economic Review, 95 (1), 277–291. Bendavid, E., Boerma, T., Akseer, N., Langer, A., Malembaka, E. B., Okiro, E. A., Wise, P. H., Heft-Neal, S., Black, R. E., Bhutta, Z. A., et al. (2021). The effects of armed conflict on the health of women and children. The Lancet, 397 (10273), 522–532. Blattman, C., & Miguel, E. (2010). Civil war. Journal of Economic literature, 48 (1), 3–57. Clark, A. E., Doyle, O., & Stancanelli, E. (2020). The impact of terrorism on individual well-being: Evidence from the boston marathon bombing. The Economic Journal, 130 (631), 2065–2104. Eaton, J., & Kortum, S. (2002). Technology, geography, and trade. Econometrica, 70 (5), 1741–1779. Edwards, R. D. (2014). Us war costs: Two parts temporary, one part permanent. Journal of public economics, 113, 54–66. Glick, R., & Taylor, A. M. (2010). Collateral damage: Trade disruption and the economic impact of war. The Review of Economics and Statistics, 92 (1), 102–127. Hotz, V. J., & Miller, R. A. (1993). Conditional choice probabilities and the estimation of dynamic models. The Review of Economic Studies, 60 (3), 497–529. Kennan, J., & Walker, J. R. (2011). The effect of expected income on individual migration decisions. Econometrica, 79 (1), 211–251. Korovkin, V., & Makarin, A. (2019). Trading with the enemy: The impact of conflict on trade in non-conflict areas (tech. rep.). Onder, H., Pestieau, P., & Ponthiere, G. (2019). Equivalent income versus equivalent lifetime: Does the metric matter? World Bank. (2021). The economics of winning hearts and minds: Programming recovery in Eastern Ukraine (tech. rep.). 21 Appendix: Proofs A Proof of Proposition 1 Proof. Using (2) we can show that 1 ∆Utl = ∆Wtk + ∆ log mkl kl t + ∆ Ct , θ and 1 ∆Utk = ∆Wtk + ∆ log mkk t , θ then subtracting U l from U k we find 1 ∆Utk − ∆Utl = ∆ log mkk kl t − ∆ log mt − ∆Ctkl , θ after adding the expression above across locations we find 1 ∆Utk = ϕl ∆ log mkk kl t − ∆ log mt + ϕl ∆Utl − ∆Ctkl , θ l∈Φ l Note that l∈Φ ϕl ∆Utl − ∆Ctkl ≤ 0 by assumption. Thus 1 ∆Utk ≤ ϕl ∆ log mkk kl t − ∆ log mt . θ l ∈Φ B Proof of Proposition 2 Proof. The flow equation implies that 1 ∆Wtk = ∆Utl − ∆Ctkl − ∆ log mkl t , θ then we can add this expression for destinations in Φ 1 ∆Wtk = ϕl ∆Utl − ∆Ctkl + ϕl − ∆ log mkl t , l ∈Φ l∈Φ θ Since l ∈Φ ϕl ∆Utl − ∆Ctkl ≤ 0, we can establish an inequality: 1 ∆Wtk ≤ ϕl − ∆ log mkl t . l ∈Φ θ 22 Appendix: Tables Table A1: Income equivalent of the welfare loss, percent - excluding the regions on the western border T =1 T = 10 T =∞ Donetsk Luhansk Donetsk Luhansk Donetsk Luhansk σ=0 β = 0.97 -241.18 -242.57 -27.56 -27.71 -7.24 -7.28 (13.15) (29.94) (1.50) (3.42) (0.39) (0.90) β = 0.90 -246.46 -247.89 -37.84 -38.06 -24.65 -24.79 (13.43) (30.59) (2.06) (4.70) (1.34) (3.06) β=0 -343.18 -345.17 - - - - (18.71) (42.60) - - - - σ=1 β = 0.97 -96.00 -94.29 -30.78 -27.90 -9.21 -8.23 (0.69) (1.90) (1.38) (2.88) (0.48) (0.97) β = 0.90 -96.23 -94.59 -39.56 -36.10 -27.96 -25.30 (0.66) (1.83) (1.65) (3.47) (1.28) (2.66) β=0 -98.65 -97.83 - - - - (0.31) (0.97) - - - - Notes: Values in percent of the welfare in the base-year (pre-conflict) welfare. σ is the degree of relative risk aversion in a CRRA utility, β is the time discount factor, and T denotes the number of years with income losses. Standard errors, in parentheses, are calculated using bootstrapped samples of destinations included in Φ repeated 5000 times. 23 Table A2: Income equivalent of the welfare loss, percent - excluding the regions on the western border and those with airports serving more than 100,000 passengers annually T =1 T = 10 T =∞ Donetsk Luhansk Donetsk Luhansk Donetsk Luhansk σ=0 β = 0.97 -217.98 -226.36 -24.90 -25.86 -6.54 -6.79 (6.29) (13.77) (0.72) (1.57) (0.19) (0.41) β = 0.90 -222.76 -231.32 -34.20 -35.52 -22.28 -23.13 (6.43) (14.07) (0.99) (2.16) (0.64) (1.41) β=0 -310.17 -322.09 - - - - (8.96) (19.59) - - - - σ=1 β = 0.97 -94.55 -93.09 -28.28 -26.31 -8.36 -7.70 (0.46) (1.17) (0.69) (1.37) (0.23) (0.45) β = 0.90 -94.84 -93.42 -36.56 -34.16 -25.65 -23.83 (0.44) (1.13) (0.83) (1.68) (0.64) (1.26) β=0 -97.96 -97.20 - - - - (0.23) (0.65) - - - - Notes: Values in percent of the welfare in the base-year (pre-conflict) welfare. σ is the degree of relative risk aversion in a CRRA utility, β is the time discount factor, and T denotes the number of years with income losses. Standard errors, in parentheses, are calculated using bootstrapped samples of destinations included in Φ repeated 5000 times. 24 Table A3: Income equivalent of the welfare loss, percent - using only the neighbors of Donetsk and Luhansk Oblasts T =1 T = 10 T =∞ Donetsk Luhansk Donetsk Luhansk Donetsk Luhansk σ=0 β = 0.97 -223.79 -215.78 -25.57 -24.65 -6.71 -6.47 (9.52) (70.09) (1.09) (8.01) (0.29) (2.10) β = 0.90 -228.69 -220.51 -35.11 -33.86 -22.87 -22.05 (9.73) (71.63) (1.49) (11.00) (0.97) (7.16) β=0 -318.44 -307.04 - - - - (13.55) (99.74) - - - - σ=1 β = 0.97 -94.96 -92.17 -28.92 -25.25 -8.57 -7.36 (0.59) (3.59) (1.02) (6.48) (0.35) (2.25) β = 0.90 -95.23 -92.53 -37.32 -32.86 -26.24 -22.85 (0.56) (3.46) (1.23) (7.73) (0.95) (6.02) β=0 -98.16 -96.69 - - - - (0.28) (1.80) - - - - Notes: Values in percent of the welfare in the base-year (pre-conflict) welfare. σ is the degree of relative risk aversion in a CRRA utility, β is the time discount factor, and T denotes the number of years with income losses. Standard errors, in parentheses, are calculated using bootstrapped samples of destinations included in Φ repeated 5000 times. 25 Table A4: Income equivalent of the welfare loss, percent - using only the largest 10 oblasts T =1 T = 10 T =∞ Donetsk Luhansk Donetsk Luhansk Donetsk Luhansk σ=0 β = 0.97 -247.78 -249.08 -28.31 -28.46 -7.43 -7.47 (15.23) (35.76) (1.74) (4.09) (0.46) (1.07) β = 0.90 -253.21 -254.54 -38.88 -39.08 -25.32 -25.45 (15.56) (36.54) (2.39) (5.61) (1.56) (3.65) β=0 -352.58 -354.43 - - - - (21.67) (50.88) - - - - σ=1 β = 0.97 -96.34 -94.72 -31.47 -28.54 -9.45 -8.44 (0.72) (2.03) (1.58) (3.39) (0.55) (1.15) β = 0.90 -96.56 -95.00 -40.39 -36.86 -28.60 -25.88 (0.69) (1.96) (1.88) (4.07) (1.47) (3.14) β=0 -98.80 -98.04 - - - - (0.31) (1.01) - - - - Notes: Values in percent of the welfare in the base-year (pre-conflict) welfare. σ is the degree of relative risk aversion in a CRRA utility, β is the time discount factor, and T denotes the number of years with income losses. Standard errors, in parentheses, are calculated using bootstrapped samples of destinations included in Φ repeated 5000 times. 26 Table A5: Income equivalent of the welfare loss, percent - using only the smallest 10 oblasts T =1 T = 10 T =∞ Donetsk Luhansk Donetsk Luhansk Donetsk Luhansk σ=0 β = 0.97 -218.36 -231.67 -24.95 -26.47 -6.55 -6.95 (7.43) (19.26) (0.85) (2.20) (0.22) (0.58) β = 0.90 -223.14 -236.75 -34.26 -36.35 -22.31 -23.67 (7.60) (19.68) (1.17) (3.02) (0.76) (1.97) β=0 -310.71 -329.65 - - - - (10.58) (27.41) - - - - σ=1 β = 0.97 -94.58 -93.51 -28.33 -26.84 -8.37 -7.88 (0.53) (1.48) (0.81) (1.90) (0.27) (0.63) β = 0.90 -94.86 -93.83 -36.61 -34.80 -25.69 -24.31 (0.52) (1.44) (0.98) (2.31) (0.75) (1.75) β=0 -97.98 -97.42 - - - - (0.27) (0.80) - - - - Notes: Values in percent of the welfare in the base-year (pre-conflict) welfare. σ is the degree of relative risk aversion in a CRRA utility, β is the time discount factor, and T denotes the number of years with income losses. Standard errors, in parentheses, are calculated using bootstrapped samples of destinations included in Φ repeated 5000 times. 27