Policy Research Working Paper 11284 Heat, Informality, and Misallocation Firm Adaptation in the Short and Long Run Jonah Rexer Siddharth Sharma South Asia Region Office of the Chief Economist January 2026 Policy Research Working Paper 11284 Abstract How do climate shocks shape resource allocation across informal firms, generating allocative efficiency losses of up firms? Rising temperatures might worsen allocative effi- to 4.3 percent. In long difference estimates spanning several ciency if large, productive firms face constraints in adapting. decades, the relationship reverses: large firms adapt and This paper assesses this question in India, an economy char- absorb labor, while small firms contract. This adaptation acterized by informality, misallocation, and extreme heat. offsets nearly 60 percent of the short-run labor demand The paper uses census data on 42 million non-farm estab- shock. These results highlight a general mechanism of lishments from 1990 to 2013 linked to granular climate climate adjustment: in the short run, shocks exacerbate histories to estimate the impact of heat on the firm size misallocation by pushing labor into low-productivity firms, distribution. A 1 degree Celsius temperature shock reduces but in the long run, adaptation by larger firms restores firm size by 11.6 percent, with losses concentrated among efficiency. large, formal firms. Displaced workers reallocate to smaller, This paper is a product of the Office of the Chief Economist, South Asia Region. 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 jrexer@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 Heat, Informality, and Misallocation: Firm Adaptation in the Short and Long Run* Jonah M. Rexer† Siddharth Sharma ‡ World Bank World Bank Authorized for distribution by Franziska Ohnsorge, Chief Economist, South Asia Region, World Bank Group JEL: Q54, J23, J46, L11, D61 Keywords: climate change adaptation, allocative efficiency, labor markets, informality * We are grateful for excellent research assistance from Yurui Hu. We thank Adrien Bilal, Ben Collier, Jonathan Colmer, Raisa Sherif, Andrea Caggese and seminar participants at the 2024 Preparing for a Changing Climate Conference at Stanford, NEUDC 2024, and the Barcelona School of Economics Summer Forum 2025 for helpful comments. We thank Paul Novosad and Teevrat Garg for facilitating access to data. 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. † Office of the Chief Economist, South Asia Region, the World Bank: jrexer@worldbank.org ‡ Office of the Chief Economist, South Asia Region, the World Bank ssharma1@worldbank.org 1 Introduction There is growing evidence that climate change harms firm productivity (Heal and Park, 2016). Much of this evidence comes from the agriculture sector, where rising temperatures, changing weather patterns, and rainfall variability have reduced agricultural productivity around the world (Burke and Emerick, 2016; Nath, 2020). Studies increasingly show that climate change also harms firms in the non-farm sector, where rising temperatures reduce labor productivity (Adhvaryu et al., 2020; Somanathan et al., 2021) and increase absenteeism (Somanathan et al., 2021). In addition to the firm-level productivity effect, climate change also has a reallocative effect. Rising temperatures tend to shift resources away from agriculture, the most climate- sensitive sector and toward higher productivity non-farm sectors, triggering structural change (Colmer, 2021; Nath, 2020). They also spur migration to less-exposed, more productive urban locations (Baez et al., 2017). In low-and middle-income countries, where numerous small, informal firms typically coexist with less numerous but more productive formal firms, wide dispersion in productivity suggests that climate-induced labor reallocation between small and large firms—even within narrowly defined sectors—could have important implications for aggregate productivity (Hsieh and Klenow, 2009; Hsieh and Olken, 2014; Restuccia and Rogerson, 2017; Hopenhayn, 2014). Reallocative effects could help offset the first-order pro- ductivity loss from climate change if larger firms adapt more effectively (Goicochea and Lang, 2023; Xie, 2024a). At the same time, reallocation could also magnify productivity losses, at least in the short-run, if the returns to adaptation by larger firms take time to materialize. Workers laid off from formal firms following a heat shock might transition into the informal sector or self-employment, lowering aggregate productivity. We examine this underexplored question in the context of India, which is among the world’s most climate change-exposed countries. Under an SSP2-4.5 warming scenario, the number of days with heat index exceeding 35 degrees Celsius in India is projected to reach a median of 80 by 2040-59 and 95 by 2060-79 (WorldBank, 2024). Like many other emerging economies, India also has a skewed firm-size distribution. Together with a small number of large firms that operate close to the global technology frontier are a multitude of informal firms with fewer than 10 employees. The latter account for 99 percent of all firms and more 2 than 75 percent of total employment in India’s non-farm sector (Bussolo and Sharma, 2022); in comparison, the modal firm size in the U.S. manufacturing sector is 45 employees (Hsieh and Klenow, 2014). Hence, vulnerability in India’s non-farm sector is both biophysical (that is, exposure to heat) and structural (that is, a preponderance of informal micro-enterprises). India is also estimated to have substantial firm-level misallocation; by one estimate, costing the economy 40-60% of total productivity (Hsieh and Klenow, 2009). Yet, existing research has focused almost exclusively on the direct productivity effects of climate change on Indian firms. We estimate the impact of heat shocks on firms using census data on the universe of 42 million establishments—both formal and informal—from 1990-2013, combined with spa- tially granular, long-term climate data. Notably, the establishment census covers all non-farm medium, small and micro-enterprises; in other words, the "MSME" sector, which is gener- ally under-represented in firm-level surveys. Of particular interest is heat-induced realloca- tion across the firm size distribution, including between informal and formal firms. Climate change in India has produced location specific warming trends with substantial spatial het- erogeneity across the country. We leverage both cross-sectional and temporal variation in temperature to identify the impact of heat shocks on the moments of the firm size distribu- tion. For identification, we rely on the spatial granularity of our data, which allows for a variety of fixed effects to control for cross-sectional differences across locations, as well as location-by-year trends at administrative levels above our unit of analysis. We first estimate the relationship between average firm size and local temperature shocks, the latter measured as annual deviations (anomalies) relative to long-run annual temperature averages. We find that a 1°C anomaly in average annual maximum temperature reduces firm size by 0.26 employees, or 11.6% on average. Our preferred specification includes village and district-by-year fixed effects, but the result is robust to other combinations of fixed effects. This relationship between temperature and firm size holds within both manufacturing and services, and survives the inclusion of highly disaggregated sector fixed effects, suggesting that it does not just reflect broad reallocation across sectors that vary in average firm size. Even within narrowly defined sectors, firms exposed to heat shocks shed labor on average. Next, to identify distributional effects, we estimate the conditional distribution of firm size using fixed effect quantile regressions. The results show that negative firm size effects of heat 3 shocks are concentrated entirely among large, formal firms. Firms below approximately the 40th percentile in firm size do not experience any contraction with a 1°C increase in tempera- ture, and actually grow slightly in some specifications. In contrast, firms above this threshold shrink, with effects increasing along the firm size distribution. Quantile regression estimates for our preferred specification reveal that a 1°C heat shock induces an increase in firm size of 10% at the 5th percentile, while firms at the 90th percentile contract by nearly 20%. These quantile effects hold, and if anything strengthen, after conditioning on disaggregated sectoral fixed effects. As such, despite the fact that we do observe some evidence of sectoral labor reallocation in response to heat shocks, such reallocations do not seem to drive effects along the firm size distribution. A potential explanation for these results is that heat shocks cause large, formal firms to shrink and shed labor, which is absorbed by small, informal firms. We explore this mechanism by estimating the impact of heat shocks on sub-district level employment shares of different firm size categories—single-person microenterprises, informal firms with 1 to 7 employees, and formal firms with 8 or more employees.1 We find that heat shocks cause a decline in the sub-district employment share of larger, formal firms and an increase in the share of smaller, informal firms, equivalent to a 2 p.p. reallocation in labor from the formal to informal sectors for each 1°C temperature increase. The results are consistent with the informal sector serving an absorptive role for labor displaced by negative shocks, a common theme in the trade lit- erature (Ponczek and Ulyssea, 2021; Ulyssea, 2020), but one that is less well-explored in the context of climate change. Given the large labor productivity differentials between formal and informal firms, our estimates imply an allocative efficiency loss of 1.3-4.3% of aggregate productivity due only to reallocation effects for each degree Celsius of temperature increase. While growing informality may serve as a transitional form of adjustment to climate shocks in environments with labor market frictions, the long-run adjustment dynamics also depend on other factors, such as differences in the adaptive capabilities of informal and for- mal firms. Since our main estimates use transitory weather shocks for identification, we there- fore estimate only short-run effects. To examine the long-run impacts of temperatures on the firm size distribution, we estimate long-difference regressions in a panel of census villages.2 1 Thislocal labor market approach is consistent prior research on the local labor markets effects of shocks in India, such as Jayachandran (2006). 2 Throughout, we refer to the smallest identifiable spatial units in the firm census as “villages,” although in 4 We regress the change in the village average firm size on the change in temperature over lags of 8, 15, and 23 years, corresponding to the available lags between census rounds. To estimate the differential effects across the firm size distribution, we interact the long-run temperature change with a flexible function of the average firm size in the initial year. Based on the long-difference regressions, a 1°C increase in long-run temperature is esti- mated to reduce average firm size by between 1 and 7.5 percent. These effect sizes are notably smaller than the estimated short-run impacts of temperature on firm size. This gap between short- and long-run impacts is consistent with prior literature, which interprets it as evidence of adaptation reducing the adverse impacts of climate shocks over time (Burke and Emer- ick, 2016; Bento et al., 2023). Under this interpretation, our estimates suggest that adaptation offsets nearly 60% of the short-run negative labor demand shock from heat. The long-difference results also show that the impact of heat on the firm size distribution reverses over time. As the lag lengthens, the temperature versus firm size curve flattens, and eventually inverts completely. For the 8-year lag, a 1°C increase in long-run temperature has a very small negative impact on firm size at the 5th percentile, but reduces it by roughly 15% at the 95th; for the 23-year time horizon, effects are quantitatively similar at the bottom of the distribution, but heat has no significant effect at the 95th percentile. This suggests that larger firms have superior adaptive capacity in the long run. Our core argument is that a heat-induced decline in labor productivity reduces labor de- mand. In the short run, displaced labor reallocates to small, unproductive, informal firms due to limited labor reallocation within the formal sector, reducing aggregate productivity. This allocative effect dissipates in the long run, as larger firms adapt to heat. Put otherwise, heat shocks worsen misallocation in the short run by shifting labor into low-productivity informal firms, but long-run adaptation by larger firms offsets much of this effect — implying high transitional adjustment costs. However, we also consider several alternative explanations for our results. One possibility is that our short-run results are explained by sample composition or measurement artefacts. For example, large firms may contract while small firms exit, leading to observed contractions among large firms and no change in size for small firms. However, we find no evidence of a reduction in the number of small firms in response to temperature shocks that would suggest urban areas these may represent entire towns, or neighborhoods in larger urban agglomerations. 5 substantial net exit in this segment. We also test for correlated measurement error induced by the urban heat island effect, which might lead to us to understate heat shocks in the urban areas where large firms concentrate. However, our average and distributional results hold for both urban and rural locations. We argue that our results likely operate through declining labor productivity and labor demand. Several possible supply-driven explanations may also explain firm contraction—for example, migration of working-age men in response to heat, or the expansion of India’s large government workfare schemes in heat-affected areas, either of which could draw labor out of local firms. However, we find no evidence that heat shocks increase migration in our sample, nor do we find evidence that impacts are larger in areas with greater coverage by government schemes. An alternative demand-side explanation that does not involve labor productivity argues that firms face negative demand shocks in product markets, which might be larger for formal firms. Using sectoral indices of tradeability as a proxy for local demand exposure, we find no evidence that impacts are smaller for firms producing more tradable products. A final explanation for our short-run results is that larger firms respond to heat shocks by switching to more capital-intensive production technologies. If this is true, then our short- run results do not necessarily map onto allocative efficiency, and may just reflect differences in adaptation methods.3 We cannot examine this mechanism directly because our main dataset has no information on output, capital investment, or input choices. However, we present de- scriptive evidence from a recent survey of South Asian firms which suggests that although larger firms are more likely to invest in building upgrades and cooling technology for protec- tion against heat shocks, these capital upgrades do not substitute for labor. We also test the robustness of the results to a variety of specification choices. We consider several definitions of the temperature shock, including using varying lag periods, binned specifications, controlling for precipitation, and testing for nonlinear heat effects. Through- out, we also test robustness of the main results to exponential models that account for the skewed, discrete nature of the firm size outcome variable. We also estimate OLS and quan- tile models that include sector fixed effects to account for differential heat exposure and firm 3 In the firm-level misallocation literature, the extent of misallocation and its aggregate productivity implica- tions are inferred from observed production patterns. This inference is sensitive to assumptions about returns to scale and technological similarity between firms in the same sector (see, for example, Haltiwanger et al. 2018; Carrillo et al. 2024; Gollin and Udry 2021). 6 size distributions across sectors. The core results remain robust to these measurement and specification choices. Our paper contributes to the growing literature on climate change impact and adapta- tion in firms. This literature has largely focused on the direct productivity impacts of climate shocks on firms, and identified a range of mechanisms deployed by firms to mitigate the im- pact of environmental shocks, including the adoption of technologies such as air conditioning (Somanathan et al., 2021), better management (Adhvaryu et al., 2022), and changes to the sup- plier base (Balboni et al., 2023).4 Some studies suggest that agricultural firms and firms that predominantly serve local markets suffer greater damage from natural disasters than manu- facturing and services firms (Nath, 2020; Gallagher et al., 2023). There is also some evidence that natural disasters have a more severe impact on smaller and weaker performing firms (Pelli et al., 2023; Xie, 2024a; Basker and Miranda, 2017). However, there is limited evidence on heterogeneous impacts, reallocation across firms as an economy-wide adjustment mecha- nism, or the dynamics of this process. While adaptation is inferred and not directly observed in our study, our results suggest that larger firms have greater long-run adaptive capacity (that is, greater managerial and financial capacity to undertake investment in cooling or other capital upgrades). Within this strand of the climate economics literature, our study builds most directly upon recent work by Xie (2024b), who examines Brazil’s labor market responses to temperature shocks. Xie (2024b) finds that a substantial fraction of workers displaced from formal manu- facturing jobs following heat shocks remain outside formal employment for extended periods (up to 36 months post-shock), with many transitioning into informal service sector work. Our findings are consistent with these results and add to them in two respects. First, Xie (2024b) focuses on worker-level heterogeneity in job transitions, while we adopt a firm-level perspec- tive, estimating the impact of heat shocks along the entire firm size distribution using a census of firms for the non-farm sector. The firm-level perspective allows us to discuss implications for allocative efficiency. Second, we estimate both short- and long-run impacts on firm size and labor reallocation, providing the first estimates of the dynamic allocative efficiency costs of climate change. 4 Recent reviews of the growing literature on climate change adaptation by households and firms include Goic- ochea and Lang (2023); Kala et al. (2023); Rexer and Sharma (2024); Carleton et al. (2024). 7 In examining the impact of temperature on the firm size distribution over varying time spans, our paper also contributes to understanding how heat shocks may have persistent effects on GDP. Several empirical studies suggest that temperature shocks have long-term ef- fects on output, both at the national level (Bilal and Känzig, 2024; Burke et al., 2015; Dell et al., 2012; Nath et al., 2024) and sub-nationally (Burke and Tanutama, 2019; Colacito et al., 2019). Our results highlight two factors that may be contributing to persistent temperature impacts: firm-level misallocation that amplifies heat’s adverse productivity effects during an adjust- ment period, and a long lag before firms are able to undertake more effective adaptation. More broadly, we contribute to the literature on misallocation across firms and how it is af- fected by economic shocks. This literature has largely focused on macroeconomic shocks, and tends to find that misallocation worsens during large-scale economic crises and protracted re- cessions (Restuccia and Rogerson, 2017). As weather shocks such as heat waves and storms become more frequent and intense, there is growing interest in understanding how they, too, might affect resource allocation across firms and industries (see, for example, Xie 2024a; Pelli et al. 2023; Pelli and Tschopp 2017). Finally, we contribute to the literature on economic shocks and informality. Transitions into and out of informality are an important margin of adjustment to trade liberalization and other changes in the economic environment, with the informal sector often absorbing workers displaced from formal jobs in the aftermath of shocks (Ponczek and Ulyssea, 2021; Ulyssea, 2020). Because informal firms tend to be less productive than formal firms, these transitions affects the efficiency of resource allocation between firms, with sizable impacts on aggregate productivity (Dix-Carneiro et al., 2021). However, these dynamics have not been studied extensively in the context of climate change. We show that similar formal-to-informal transi- tions also result from rising temperatures, consistent with evidence that inflexible regulations constrain re-allocative adjustment to weather shocks within the formal sector (Adhvaryu et al., 2013; Chaurey, 2015; Colmer, 2021). The next section describes our data and presents descriptive statistics of temperature trends and the firm size distribution in India. Section 3 discusses our empirical strategy. Sec- tion 4 presents our results on the short-run impacts of heat shocks, and section 5 quantifies their reallocation-driven effects on aggregate labor productivity. Next, section 6 presents our estimates of the long-run impacts of heat. Section 7 discusses robustness checks and alterna- 8 tive explanations, and section 8 concludes. 2 Data and measurement 2.1 Data The primary outcome variable of interest is firm size. This is taken from establishment- level data come from Indian Economic Census. The census collects data on the universe of Indian economic establishments—both formal and informal—at an 8-year frequency. The survey contains information on firm location, the quantity and gender composition of the firm’s employment, the characteristics of the firm owners (gender and caste), the economic sector (harmonized at the 3-digit National Industrial Classification of sectors), and data on access to finance and energy sources. The firm-level census data was obtained from The Socioeconomic High-resolution Rural-Urban Geographic dataset on India (SHRUG) compiled by Asher et al. (2019). SHRUG provides harmonized location keys—known as SHRIDs—to match firms to specific towns and villages over time, although specific firms cannot be tracked as a panel through the census waves. We obtain four waves of the SHRUG Economic Census: 1990, 1998, 2005, and 2013.5 Though the census questionnaire collects only a small set of firm-level variables (for e.g., there is no data on inputs and outputs), it benefits from a very large sample of over 100 million unique firm-year observations of private establishments in the non-farm sector, a sample restriction we maintain throughout the paper. The four waves allow a panel structure only at the village level.6 We merge firm-level census data to village- level temperature histories for approximately 510,000 villages across India.7 Benchmarking adjustment costs requires estimates of differentials in labor productivity between the formal and informal sectors. Since there is no survey containing detailed firm outcomes that covers both formal and informal firms in India, we follow common practice in the literature and combine data from the 2015-16 National Sample Survey of the Unincorpo- rated Enterprises (NSS), which covers informal firms, with the 2015 Annual Survey of Indus- 5 SHRUG data is publicly available here. 6 Throughout,we refer to the unique, lowest-level administrative areas in the SHRUG (the SHRID) as “vil- lages.” These units, however, vary dramatically in size, from small rural hamlets to subsections of large urban centers. 7 The merge rate varies across census years. Summary statistics on firm and village sample sizes are in Ap- pendix Table A1. Details on the data construction pipeline are in Appendix A. 9 tries (ASI), which covers the formal manufacturing sector (Nataraj, 2011; Hsieh and Klenow, 2014; Hsieh and Olken, 2014). This combined ASI-NSS dataset enables us to measure the gap in labor productivity between firms with 8 or more employees (“formal”) and smaller, infor- mal firms in a period close to the last year of our Economic Census data, 2013.8 We calculate labor productivity as value-added per worker. Temperature data are the ERA5 global reanalysis data (Hersbach et al., 2020), a standard temperature dataset used in climate science on South Asia (Kumar et al., 2023) and produced by the European Union through the Copernicus Climate Change Service. We obtain daily minimum, maximum, and average temperature readings for each of our villages from 1980- 2021. Following the climate science literature, our main independent variable of interest for analyzing short-run heat impacts measures the deviation of temperature from its long run average. We measure temperature anomalies by first averaging the daily temperature within years for each of the mean, minimum, and maximum readings. We then average these an- nual averages for the census year and the year prior to smooth out year to year variation in temperature. Finally, we subtract from this 2-year lagged average the long-run village-level average annual temperature from 1980-2021. This 2-year temperature anomaly calculation expresses all deviations relative to the long run mean, accounting for spatial differences in long-run temperature levels. 2.2 Descriptive statistics We plot the primary identifying variation of the study in Figure A3. The left-hand panel shows maximum temperature anomalies for the 1,743,430 village-years in our sample. This variation in unexpected deviations from average temperature drives the estimates of short- run responses to heat, where minimal adaptation is expected. The right-hand panel shows long-difference temperature change over the study period for the 575,117 unique villages in the data. This variation drives estimates of the long-run response to temperature, which includes adaptation over a nearly 25 year period. The short-run anomalies generally do not exhibit a clear pattern, distributed randomly around a slightly negative temperature shock. Meanwhile, the long-differences reflect overall warming patterns and are distributed above 8 The NSS Sector survey was conducted at an approximate interval of 5 years and the 2015-16 round is the closest to the most recent Economic Census round. 10 zero, with substantial clustering within a 0-0.5 degrees Celsius increase over the period. Figure 1: Village-level temperature: anomalies and long-differences .125 .125 .1 .1 Fraction Fraction .075 .075 .05 .05 .025 .025 0 0 -1.5 -1 -.5 0 .5 1 -1.5 -1 -.5 0 .5 1 Max temperature anomaly (°C) Temperature change (°C), 1990-2013 Note: Figure shows distribution of village-year temperature anomalies (left) and village-level long- difference temperature changes from 1990-2013, both measured in degrees celsius. Anomalies are calculated as the deviation between the average annual village-level maximum temperature in the two years before the census year and the long-run average annual maximum temperature from 1980-2021. Sample is all villages (SHRIDs) for the last four Economic Censuses of India (1990, 1998, 2005, and 2013). Appendix B investigates warming trends across India in greater detail. Following Kumar et al. (2023), we estimate subdistrict-level linear temperature trends from 1980-2021 for each sub-season (plus annually) and for each district. We then collect and plot these estimates by region and season. The results reveal localized warming trends across the country that are uneven, seasonal, and spatially heterogeneous. The largest warming trends, reaching up to 0.7 degrees every decade, are concentrated in India’s northern regions, in colder seasons, and observed more commonly for daily minimum and average, rather than maximum, tempera- tures. Appendix Table A1 also provides summary statistics on the firm size distribution in the Indian economic census. The table contains 3 panels (A-C), each of which gives census round- wise statistics of the firms size distribution for all firms, services firms, and manufacturing firms, respectively. As is well-established, average firm size in India is small, ranging from 2.05-3.06 employees per firm across the various subsamples and census rounds. Manufactur- ing firms tend to be roughly 0.4-1 employee larger than services firms. In Panel A, the firm size distribution is also quite stable across years. In each round, the median firm is a single- 11 person enterprise, indicating informal self-employment. The firm size distribution is also highly skewed. Only at the 75th, and particularly the 95th percentile do firms grow to sizes consistent with organized economic activity. The final column of the table gives the count (in millions) of surveyed firms in the census which can be linked to temperature histories. 3 Empirical strategy Our aim is to estimate the impact of heat shocks on both the conditional expectation and conditional quantiles of the firm size distribution. Using a repeated cross section of firm-level data, the baseline linear specification for firm i in village v at census year t is: yivt = α + β Avt + δv + δdt + ϵivt (1) Where yivt is firm size, Avt is the 2-year village-level maximum temperature anomaly, δv is a village fixed effect, and δdt is an administrative unit-by-time effect, where the level of the administrative unit is higher than the village and may vary depending on specification. Typi- cally, we use district fixed effects, given the important role that districts play in the implemen- tation of many national and state-level policies. Unit-by-time trends are important given the substantial regional heterogeneity in temperature trends, as well as potential location-specific changes in measurement and sample composition in the economic census.9 We test the robustness of the results to various specifications of fixed effects and interacted trends, as well as different measurement approaches for Avt . Our main results contain five core specifications: (i) unconditional, (ii) with district fixed effects, (iii) with district-by-year fixed effects, (iv) with village (SHRID) fixed effects, and (v) with village and district-by-year fixed effects. Note that with village fixed effects, a specification including the temperature anomaly on the right-hand side is equivalent to one in which only the temperature level is included, since the long-run average is subsumed by the village fixed effect. Standard errors are clustered at the village level in all specifications. We estimate both the conditional expectation E[yit | Ait ] and the conditional quantiles Qτ (y| A) for τ from to 5 to 95 in intervals of 5. Since the nonlinearity of quantile fixed effects models 9 Potentially confounding composition effects may be particularly important given that the share of firms that can be matched to precise locations, and therefore temperature histories, varies substantially over census rounds. 12 presents an incidental parameters problem, we use the location-scale quantile fixed effects es- timator of Machado and Santos Silva (2019). Our identifying assumption is that, conditional on the fixed effects, the residual temperature variation across villages and over census rounds is orthogonal to unobservables that determine firm size. Our results would be confounded if spatially differentiated trends in economic activity – or sample composition and measure- ment – are correlated with year-to-year changes in temperature. These correlations would have to occur at a level below the district, given the inclusion of district-by-year fixed effects in our preferred specification. 4 Short-run results 4.1 Linear estimation The estimated results of the baseline linear model are in Table 1. The estimates here re- veal the negative impact of temperature shocks on average firm size. Quantitatively, in our preferred specification in column (5), a 1°C increase in the annual average maximum tem- perature reduces firm size by 0.26 employees, equivalent to 11.6% since the average firm size in our sample is approximately 2.3 employees. This coefficient is similar in magnitude and significant at 1% for specifications with district fixed effects (2), village fixed effects (4), and village fixed effects with district-by-year trends (5). However, it falls to just 0.05 with district- by-year effects alone in (3). Still, across all specifications, the estimated coefficient is always negative. In Section 7, we discuss the robustness of these results to alternative specifications of the heat shock and the estimation model. The results are similar across specifications that use mean and minimum temperature anomalies, extend the lag period for measuring anomalies, measure heat shocks by the number of hot days, control for precipitation anomalies, and employ an exponential model with a fixed effects Poisson regression estimator to account for outliers given the skewness of the firm size distribution. 13 Table 1: Heat shocks and firm size Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Max temp anomaly -0.108 -0.169*** -0.056 -0.161*** -0.263** (0.081) (0.027) (0.093) (0.026) (0.085) Observations 110584807 110584805 110584805 110570800 110570800 District FE No Yes No No No District × Year FE No No Yes No Yes SHRID FE No No No Yes Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sec- tor firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Outcome is the total number of employees of the firm. Anomalies are calculated as the difference between the average annual maximum temperature of t and t − 1 and the long-run average temperature from 1980-2021. Fixed effect specification given in table footer. *** p < 0.01, ** p < 0.05, * p < 0.1. 4.2 Quantile effects One possible explanation for small average estimates might be heterogeneity across the firm size distribution. If firms’ sensitivity to heat is size dependent, then even a zero aggregate effect might mask substantial reallocation across firms. Figure 2 plots quantile regression coefficients and 95% confidence intervals for our main fixed effects models. Each coefficient gives the marginal effect of a 1°C increase in the maximum temperature anomaly on quantile τ of firm size, conditional on fixed effects. The quantile estimates reveal that the negative effects in Table 1 are concentrated in large firms. Depending on specification, firms below the 40th percentile see either no change or small gains in average firm size. In contrast, firms above 40th percentile see substantial losses. At the 95th percentile, firms lose an average of roughly 1 worker for every 1°C increase in tem- perature. The shape of the quantile curve is robust to fixed effects, though the steepest slope, as well as the positive effects for small firms, obtains with the most exacting specification of village fixed effects and district-by-year trends. To aid in the interpretation of magnitudes, Appendix Figure A4 shows the estimated quantile effects normalized by the quantile values to give percentage changes. In the pre- ferred specification, firms at 5th percentile gain just over 10% of firm size following a 1°C shock. At the same time, firms at the 90th percentile shed approximately 20% of their labor, going from 5 employees on average to 4, an economically large contraction. However, the 14 Figure 2: Quantile regressions of firm size on temperature anomalies Effect on firm size 1 Effect on firm size 1 .5 .5 0 0 -.5 -.5 -1 -1 -1.5 -1.5 -2 -2 5 25 45 65 85 5 25 45 65 85 Quantile Quantile District FE District-Year FE Effect on firm size 1 Effect on firm size 1 .5 .5 0 0 -.5 -.5 -1 -1 -1.5 -1.5 -2 -2 5 25 45 65 85 5 25 45 65 85 Quantile Quantile SHRID FE SHRID and District-Year FE Estimate 95% CI Note: Figure shows estimates and 95% confidence intervals of β τ from quantile regressions of firm size on temperature anomalies and geographic fixed effects. Standard errors clustered at the village level. Anomalies are calculated at the village level as the deviation between the average annual temperature in the two years before the census year and the long-run average annual temperature from 1980-2021. Sample is the unit level data of the last four Economic Censuses of India (1990, 1998, 2005, and 2013). normalized quantile curves are no longer monotonic, with a slightly smaller effect at the 95th than 90th percentiles. Still, these results confirms that the small and sometimes insignificant average effects of Table 1 are masking meaningful distributional heterogeneity. Even in the specifications with the weakest average results, firms at the 9th percentile contract by just under 10%. Since large firms lose and small firms gain labor from climate shocks, this plau- sibly suggests reallocation along the firm size distribution. Such reallocation could imply worsening misallocation, given the well-established relationship between firm size and TFP 15 (Van Biesebroeck, 2005). 4.3 Sector-specific heterogeneity One plausible explanation for these distributional effects is reallocation across sectors dif- fering in average size. For example, if certain industrial processes are highly heat-sensitive, then labor that was released from industrial sectors with large firms might enter self-employment and micro-enterprises across a range of sectors. In contrast, if large firms within any sector face greater regulatory constraints in adapting to shocks, for example, one could imagine that even within sectors, labor shedding primarily occurs among larger firms. Existing literature, for example Colmer (2021), has primarily studied reallocation from agriculture to industry and services, and has not considered reallocation within non-agriculture, nor has it consid- ered the degree of sectoral granularity that our data contains. We estimate reallocation effects across 90 non-farm subsectors in India’s National Indus- trial Classification (NICS) divisions. To obtain heat-induced changes in sector shares, we estimate regressions of the sector labor market share on heat shocks at the village-year level, controlling for village and census round fixed effects. We run this model separately for each of the 90 NICS sectors in the data to obtain a coefficient for each. Figure 3, Panel A plots the sector-specific coefficients and 95% confidence intervals, or- dered by magnitude. The results show substantial heterogeneity in sector sensitivity to heat. The most negatively affected sectors lose roughly 2 p.p. of their labor market share, which is compensated for by gains among the winners of equal magnitude. In the distribution of sector-level t-statistics, 33 sectors gain and 32 lose with effects significant at 5%. The remain- ing 25 sectors do not see significant changes in their labor market share. The average sector loss for statistically significant losers is 0.21 p.p, which is quite meaningful given the very low level of sectoral disaggregation. Heat shocks produce substantial labor reallocation across India’s non-farm sub-sectors. In Panel B of Figure 3, we test whether sectors that gain (lose) market share contain firms that are systematically smaller (larger), which would suggest that cross-sectoral reallocation drives the quantile effects of Figure 2. We find no evidence of a relationship between average firm size and the sectoral reallocation coefficient. This suggests that, despite the fact that heat shocks generate some sectoral reallocation, this structural change does not systematically shift 16 Figure 3: Impact of heat shocks on sector labor market shares (a) Distribution of sector reallocation effects (b) Correlation with sector-average firm size 2 Average reallocation effect (p.p.) 1 0 -1 -2 -2 -1 0 1 2 0 5 10 15 Impact of 1°C anomaly on sector labor market share Firm size (sector average) Note: Panel A shows sector-level coefficient estimates and 95% confidence intervals of the impact of heat anomalies on sector employment shares at the village level. All models include village and year fixed effects. Standard errors clustered at the village level. Anomalies are calculated at the village level as the deviation between the average annual temperature in the two years before the census year and the long-run average annual temperature from 1980-2021. Panel B shows the correlation between sector-specifc reallocation coefficients plotted in Panel A and sector-level average firm size in number of employees. Sample is all villages in India for which Economic Census data is available. labor toward sectors with smaller firms on average. Therefore, structural change is unlikely to explain the quantile result, which is more likely to be the result of within-sector effects. We consider three additional tests for sector effects. Table A2 estimates the main firm-level linear regression, splitting the sample into large macro-sectors of services and manufacturing. Effect sizes are substantially larger in the manufacturing sector: a 1 degree increase in max- imum temperature reduces firm size by 23.5% in manufacturing and 9.7% in services. It is therefore possible that the quantile effects are driven in part by larger negative responses among larger manufacturing firms relative to smaller services firms. However, Table A3 goes a step further by including fixed effects for 90 3-digit subsectors in the main model. In our pre- ferred specification (column 5), the estimated average firm size effects hardly change, whether we allow for fixed sector heterogeneity (Panel A) or unrestricted sector-specific time trends (Panel B). Finally, Appendix Figure A5 plots quantile curves conditional on sector fixed ef- fects. The results are nearly identical to those of Figure 2. In sum, it is indeed highly likely that sectors are not equally exposed to, or equally able to adapt to, increasing heat. However, these sectoral heterogeneities do not appear to be driving effects across the firm size distribu- tion. Instead, it appears that large firms contract employment most regardless of the sector in 17 which they operate. 4.4 Formal-informal reallocation Our quantile regressions have two possible interpretations. First, labor may be reallocat- ing from large formal to small informal firms. This would imply an aggregate expansion in the labor share of the informal segment of the distribution. On the other hand, the quantile effects may not reflect any reallocation at all. Instead, it could be the case that all segments of the firm size distribution are shedding labor, but for large firms this shows up as a drop in firm size, whereas for small firms, particularly single-person firms, it shows up as exit, without any change in average firm size. For this reason, we turn to aggregate employment outcomes to distinguish these explanations. To test for reallocation across sectors, we aggregate total employment shares to the sub- district level by firm size category. We use subdistrict aggregates because there are likely to be substantial cross-village spillovers, given that labor, product, and factor markets are highly integrated within relatively narrow local areas (Jayachandran, 2006). There are 5883 unique subdistricts in our sample. We calculate the temperature shock as the average maxi- mum temperature anomaly in the subdistrict.10 For each subdistrict, we calculate the share of total employment in firms of the following size groups: (i ) 1 employee, (ii ) 2-7 employ- ees, and (iii ) 8 or more employees. The first two categories represent the self-employed and the household or informal sector, respectively. We placed the second cutoff at eight employ- ees because it is the threshold at which the Economic Census classifies firms as "directory" business establishments (as opposed to household business establishments).11 We note that this cutoff yields a relatively inclusive definition of the "formal" sector in the Indian context, where a threshold of 10 employees is more commonly employed given that manufacturing establishments above this size must register under the Factories Act (1948) (Kanbur, 2017; Amirapu and Gechter, 2020).12 We investigate employment shares instead of levels in order 10 Specifically, we take the weighted average of village-level anomalies, where the weights are village geo- graphic areas. 11 Directory establishments must fill in an additional module of the Census questionnaire which includes ques- tions on their tax and business registration status and associated identification numbers. This module was used to generate a "Business Registry" from the Sixth Economic Census (CSO, 2013). 12 More broadly, there are multiple business registration requirements for firms that vary by industry and state, and have different size thresholds. 18 to focus exclusively on reallocations across segments, accounting for the fact that aggregate employment may contract for all firm size segments. For subdistrict s at time t in firm size group g, we estimate the regression: g g g g est = α g + β g Ast + δt + δs + ust (2) g Where est is the employment share of firm size group g, Ast is the subdistrict average maxi- g g mum temperature anomaly, and δt , δs are time and unit fixed effects, respectively. Standard errors are clustered at the subdistrict level. Table 2: Heat shocks and employment shares, subdistrict level Dependent variable Share of employment (1) (2) (3) (4) (5) (6) Panel A: Firm size = 1 Max temp anomaly -0.004 0.020*** 0.004 0.020*** 0.006** -0.005 (0.003) (0.002) (0.003) (0.002) (0.003) (0.015) Panel B: 1 < Firm size < 8 Max temp anomaly 0.016*** 0.005* 0.016*** 0.005** 0.016*** 0.013 (0.003) (0.002) (0.003) (0.002) (0.003) (0.016) Panel C: Firm size ≥ 8 Max temp anomaly -0.012*** -0.024*** -0.020*** -0.025*** -0.021*** -0.008 (0.003) (0.002) (0.003) (0.002) (0.003) (0.014) District FE No Yes Yes No No No Year FE No No Yes No Yes No Subdistrict FE No No No Yes Yes Yes District × year FE No No No No No Yes Observations 22335 22335 22335 22335 22335 22335 R2 0.001 0.271 0.275 0.588 0.592 0.728 Note: Standard errors in parentheses clustered at the subdistrict level. Sample is all private sector firms surveyed aggregated to the sub-district level in four Economic Census rounds (1990, 1998, 2005, and 2013). Outcome is the subdistrict-level share of employment in each firm size category. Anomalies are calculated as the difference between the average annual maximum temperature of t and t − 1 and the long-run average temperature from 1980-2021, and then averaged the subdistrict-level weighted by area. Fixed effect specification given in table footer. *** p < 0.01, ** p < 0.05, * p < 0.1. Table 2 2 shows the results for a range of specifications of fixed effects. Across the speci- 19 fications, the evidence is suggestive of reallocation between large, formal firms, and smaller informal ones. In our preferred two-way fixed effects specification of column (5), we estimate that a 1 degree increase in the temperature anomaly reduces the share of employment in large firms by 2.1 p.p. This reduction is offset by an increase in the subdistrict employment shares of single-person microenterprises (0.6 p.p.) and small firms below eight employees (1.6 p.p.). All estimates are significant at 5%. Across specifications, the estimates are qualitatively consistent. For mid-sized and large firms, the estimates are always positive and negative, respectively, and significant in 5 out of 6 specifications. For single-person firms, the estimates are somewhat less stable, ranging from precise zero estimates to a 2 p.p. reduction in employment shares. The most exacting specification in column (6) adds district-by-year time trends to the two-way fixed effects of (5). In this specification, the estimates lose significance, though they are directionally similar. Since there are only 9 subdistrict observations per district-year on average, there is limited variation to exploit, and this specification likely imposes too-strict requirements on the data. Overall, the estimates provide suggestive evidence of a reallocation in employment from large to small firms across aggregated labor markets. 5 Aggregate productivity impact of heat shock reallocation In this section, we employ a standard macroeconomic accounting framework to quantify the aggregate productivity implications of the heat-induced reallocation of workers from for- mal to informal firms in Table 2. The analysis closely follows the approach used by McCaig and Pavcnik (2018) to estimate the aggregate labor productivity gain from an export-induced reallocation of workers across formal and informal firms in Viet Nam.13 5.1 Model In a frictionless, perfectly competitive benchmark economy, the marginal revenue product of labor (MRPL) should be the same across sectors in equilibrium. Labor market frictions or distortions may lead to departures from this condition, generating scope for changing aggre- 13 McCaig and Pavcnik (2018) build their approach on papers such as Gollin et al. (2014). Note that in their case, the trade shock causes a reallocation from lower-productivity informal firms to higher-productivity formal firms, leading to an increase in allocative efficiency. In our paper, the reallocation occurs in the opposite direction. 20 gate labor productivity by shifting workers across sectors that differ in MRPL. For example, as in our paper, if MRPL is lower in the informal sector, then reallocating workers from formal to informal firms would reduce aggregate labor productivity. The aggregate labor productivity impact of labor reallocation between formal and infor- mal firms depends on their MRPL gap. This gap can be estimated from the observed gap in their average productivity of labor. Suppose output in sector s ∈ {in f ormal , f ormal } is given α s 1− α s by a Cobb-Douglas production function of the form Ys = As Ks Ls . If markets are com- petitive, each sector will employ labor until its marginal revenue product equals the wage rate: ws = MRPLs = (1 − αs ) ARPLs (3) where ARPLs is the average revenue product of labor in sector s. Assuming that the output elasticity of labor, α, is the same in each sector, this implies that win f ormal MRPLin f ormal ARPLin f ormal = = (4) w f ormal MRPL f ormal ARPL f ormal In other words, the ratio of the marginal revenue product of labor equals the ratio of the average revenue product of labor. Given this ARPL ratio, we can estimate the percentage change in aggregate labor productivity due to reallocation by applying the following account- ing formula: ∆s( ARPLratio − 1) ARPL f ormal ∆ PL = (5) sin f ormal ARPLin f ormal + (1 − sin f ormal ) ARPL f ormal Here, ∆ PL is the change aggregate ARPL, while ∆s is the percentage of total employment reallocated from the formal to the informal sector and ARPLratio is the ratio of ARPL in the informal and formal sectors. 5.2 Estimation of productivity gaps Our estimate of ∆s, the percentage of employment reallocated to informal firms, is based on our subdistrict-level regressions of employment shares in various firm size categories on temperature anomalies. According to the estimated coefficient in our preferred two-way fixed effects specification (column 5 of Table 2), a 1 degree Celsius temperature anomaly reduces the share of employment in large firms by 2.1 percentage points. This share of total employment 21 is reallocated to the informal sector, which we have defined as firms employing fewer than 8 individuals, including the self-employed. We estimate ARPLratio using the combined NSS-ASI dataset by regressing firm-level value- added per worker on an indicator variable for formality, defined as firm size above 8 employ- ees. The productivity regression results are presented in Appendix Table A11. One concern is that part of the gap in observed labor productivity could be due to a higher average human capital level in formal firms; this component should be excluded when assessing the change in aggregate productivity resulting from worker reallocation. Ideally, we would control for worker level attributes such as years of education, as in McCaig and Pavcnik (2018). However, since ASI and NSS do not collect worker-level data, we estimate gaps in productivity condi- tional on state and 5-digit industry fixed effects. Estimation details are available in Appendix D. The first column of Table 3 presents our estimates of the ratio of labor productivity in informal and formal firms. The first row presents our baseline or "raw’" ARPLratio , which shows that labor productivity in informal firms is just 38% that of formal firms. The second and third rows present the ARPLratio adjusted for 5-digit industry fixed effects and state fixed effects. As expected, using within-industry variation industry reduces the estimated labor productivity gap between formal and informal firms. The state fixed effects do not make a substantial additional difference.14 The last two rows present two additional estimates of ARPLratio for robustness. de Mel et al. (2009) estimate that micro-enterprises in low- and middle-income countries may under- report revenue by about 30 percent. As this would tend to bias the productivity gap upwards, we present an estimate where reported revenue of informal firms has been adjusted for 30 percent under-reporting. Lastly, we present the ARPL gap using 10 workers as the formality threshold. Our baseline 8-worker cutoff—which follows the Economic Census’s categoriza- tion scheme—is a relatively inclusive definition of formal firms in the Indian context (Kanbur, 2017). By including borderline informal firms in the formal segment, the 8-worker cutoff may lead to an underestimation of the latter’s productivity advantage. Indeed, the ARPLratio based on the 10-worker cutoff is nearly half of that based on the baseline 8-worker cutoff.15 14 However, these adjustments will be overly conservative, since a substantial share of the cross-sector variance in labor productivity will be due to firm rather than worker effects. 15 This may explain why our raw productivity differentials are low compared to the range of estimates re- 22 5.3 Aggregate productivity losses Column (2) of Table 3 presents the main estimates of the decrease in aggregate labor pro- ductivity obtained by applying the accounting formula in equation (5) to the ARPL ratios in column (1).16 The estimated decrease in aggregate labor productivity due to reallocation caused by a 1 degree increase in the temperature anomaly ranges from 4.3% in the 10-worker cutoff specification to 1.3% in the 8-worker model with industry and state fixed effects. Even the bottom end of this range represents a sizable negative productivity impact, considering that the temperature shock is only 1 degree Celsius. Estimates from a moderate-high climate scenario (SSP3-7.0) forecast that by 2075, India’s temperature could rise by 2.17 degrees Cel- sius relative to the final year of our sample period. This implies an aggregate productivity loss of 2.8-9.3% due only to reallocation effects. Table 3: Labor reallocation and change in aggregate labor productivity Method ARPLratio ∆ PL (1) (2) Raw 0.38 -2.47 Industry FE 0.55 -1.44 Industry + State FE 0.58 -1.33 Underreporting-adjusted 0.55 -1.46 10-worker cutoff 0.20 -4.32 Note: Estimates based on National Sample Survey of Unincorporated Non-agricultural Enterprises (2015-16) and the Annual Survey of Indus- tries (2015). The ARPL ratio is the ratio of value added per worker in informal and formal establishments. The "raw" or baseline ARPL ratio is estimated by regressing value added per worker on a dummy for infor- mality of the establishment. The "industry FE" and "industry + state FE" ARPL ratios are based on including 5-digit industry and state fixed ef- fects in the baseline regression. The "underreporting-adjusted" estimate is based on correcting value added per worker in informal establishments for an assumed under-reporting by 30%. The "10-worker cutoff" estimate is based on defining the formal sector using a minimum size of 10 work- ers, as compared to a baseline cutoff of 8 workers. We acknowledge that our estimates are subject to significant methodological caveats. A key limitation of our approach, as discussed, is that we cannot control adequately for differ- ported in McCaig and Pavcnik (2018) and previous papers such as Nataraj (2011) and La Porta and Shleifer (2014). La Porta and Shleifer (2014) report that the median ARPLratio across a sample of low-and middle-income countries is 0.15. Their estimate for India, 0.18, is close to our 10-worker cutoff estimate. 16 Appendix Table A12 presents detailed calculations. 23 ences in worker skills across formal and informal firms due to a lack of worker-level data in the NSS and ASI surveys. However, our estimates with sector fixed effects are only sensi- tive to differences in skills within narrow 5-digit industries. Another limitation is that since the ASI dataset only covers manufacturing firms, we are extrapolating the ARPL gap from manufacturing to services firms. Other caveats, which also apply to McCaig and Pavcnik (2018) and other papers sharing this approach, concern the validity of using the ratio of the average revenue product of labor to measure the ratio of marginal labor productivity in informal and formal firms. Differences in the output elasticity of labor between formal and informal firms would drive a wedge between the ARPL and MRPL ratios (equation 3). Furthermore, revenue-based ratios may be biased measures of physical productivity ratios if there are systematic differences in price markups by firm size due to market power. While data limitations prevent further robustness checks to address these concerns, it is worth reiterating that our estimates of the productivity gap between formal and informal firms are already at the lower, more conservative end of the range reported in the literature. This may be due to our conservative size cutoff for defining formality. More broadly, our binary treatment of heat-induced reallocation—formal versus informal—might be understat- ing its productivity impact by ignoring reallocation within the formal sector. As the quantile regressions suggest, heat shocks cause a reallocation from larger to smaller firms even within the formal sector, with the largest contractions seen for the largest firms. The two-sector pro- ductivity accounting framework, while desirable because of its tractability and grounding in prior literature, does not account for the productivity impacts of this reallocation. 6 Long-run results The empirical result that smaller firms are less affected by heat shocks – and even absorb labor from larger firms – may appear counterintuitive. After all, larger formal firms in India have greater access to capital and management expertise, both of which are strongly corre- lated with climate adaptation investments (Lang et al., eds, 2025). Evidence from a global sample of firms also show that small and medium enterprises tend to be less resilient to tem- perature shocks (Berg et al., 2025). 24 However, we argue that in the short-run, the informal sector can serve as a safety net for labor released from larger, formal firms affected by shocks. Large firms shed labor because they face adjustment costs to investing in adaptations that mitigate the impact of heat on labor productivity. For instance, new machinery takes time to install, while more heat-resilient production processes must be designed and implemented. It is therefore plausible that in the long-run, larger firms eventually adapt, the labor market adjusts, and labor is re-absorbed into larger firms. In this case, the expansion of the informal sector is a transition dynamic that represents an economy-wide adjustment cost in adapting to heat. In this section we use long difference specifications to compare the impact of heat on the firm size distribution in the short and long run. 6.1 Long difference specification We estimate the long run average effects of using the following specification for village v ∆yvkt = λk + β k ∆ Tvkt + υvkt (6) Where the outcome y, is the average firm size in village v and the independent variable is average annual maximum temperature T , averaged over the year of the census and the year prior. The long difference is defined as ∆yvkt = yvt − yv,t−k for a given lag k. The longest lag in our data is 23 years, from the initial round of the economic census in 1990 to the final round in 2013. However, we also calculate differences for all intermediate lag periods, the 8-year lag (1990-1998, 1998-2005, and 2005-2013) and the 15 year lag (1990-2005, 1998-2013). We then stack each t, k cross-section into a single dataset, estimate the model separately for each k and cluster standard errors at the village level. By differencing, the specification removes village fixed effects; by including the constant , it removes time trends within a long-difference period. The specification is therefore quite similar to our earlier specification (1), except for two key differences: i ) we exploit longer- term shocks rather than two-year lagged transitory anomalies for identification, and ii ) the specification is estimated at the village rather than the firm level, given the absence of panel data at the firm-level. To estimate the differential effects of long-run temperature change across the firm size 25 distribution, a quantile regression is no longer appropriate. This is because equation (6) is specified in changes, so the quantile regression will estimate the conditional quantiles of the distribution of firm size changes ∆y rather than y, as desired. To solve this problem, we esti- mate the following specification: ∆yvkt = αk + ∆ Tvkt × θk (yv,t−k ) + ε vkt (7) Where the long difference change in temperature ∆ Tvkt is interacted with the flexible func- tion θk of average firm size in the initial period of the difference yv,t−k . The function θk eval- uated at a given initial firm size yv,t−k therefore gives the impact of a 1°C long-run increase in temperature at a given point in the firm size distribution. We estimate the θ functions us- ing a flexible, kernel-weighted local linear regression (Hainmueller et al., 2019). Because the firm size distribution is highly skewed, we also use a variable kernel estimation procedure in which the bandwidth varies along y and is inversely proportional to the density at a given point in the distribution (Terrell and Scott, 1992).17 6.2 Long-run average effects The estimates of equation 6 are presented in Table 4. Columns (1)-(2) provide estimates for k = 23, columns (3)-(4) for k = 15, and columns (5)-(6) for k = 8. Across all specifications, a 1°C increase in the long-run temperature change is associated with a 0.02 to 0.13-employee reduction in firm size. All estimates are significant at the 5 or 1% levels. The inclusion of state fixed effects – which is equivalent to a state-by-time trend given the long differencing – increases the relationship from 0.016 to 0.069 for k = 15, but weakens it from 0.13 to 0.073 for k = 8. Overall, it appears that effects are slightly larger for the longest difference, but remain robustly negative and significant across all specifications. In the state fixed effects specifications, the k = 23 estimates are statistically significantly larger than the other two differences, though the magnitudes are moderate, while k = 15 and k = 8 coefficients are not statistically different from each other. The results are robust to using log differences in Appendix Table A4. In these models, a 1°C increase in long-run temperature is estimated to 17 Specifically, h we set h(y) = p( y) , where p(y) is the density function of initial firm size and h is a constant equal to the desired bandwidth at the modal point of the distribution ( p(y) = 1). 26 reduce average firm size by between 1 and 7.5%. Figure A6 visualizes the long difference correlations with binned scatterplots. Table 4: Heat shocks and firm size: long difference estimates Dependent variable Change in firm size (employment) Difference 23-year 15-year 8-year (1) (2) (3) (4) (5) (6) Temperature change (C) -0.102*** -0.110*** -0.016** -0.069*** -0.130*** -0.073*** (0.008) (0.012) (0.007) (0.009) (0.003) (0.007) State × Period FE No Yes No Yes No Yes Observations 331998 331998 697146 697146 1097327 1097327 R2 0.001 0.010 0.000 0.022 0.003 0.023 Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013), collapsed to the village level. Outcome is the change in average number of employees of private firms in a village over the long difference period t, k. Temperature shocks are calculated as the change in the 2-year lagged average annual maximum temperature over the long difference period t, k. Fixed effect specification given in table footer. *** p < 0.01, ** p < 0.05, * p < 0.1. Notably, these effects are smaller in magnitude than the estimates in Table 1, in which the coefficient in the preferred specification is 0.26. This is consistent with a large literature mea- suring the differential effects of long and short-run heat shocks on agricultural productivity and other economic outcomes. In the literature, this difference is typically interpreted as a reduced-form estimate of the magnitude of climate adaptation, since adaptation should be expected to reduce the costs of climate shocks over time (Burke and Emerick, 2016; Bento et al., 2023). Rexer and Sharma (2024) define the “adaptation ratio” as the share of the economic losses due to weather shocks that are recovered by climate adaptation. If we interpret the difference between the short-run estimate of 0.26 and the long-run estimate of 0.11 (column 2) as the effect of adaptation, then the adaptation ratio in this setting is about 0.58. 6.3 Long-run distributional effects The long-difference estimates provide evidence of firms’ long-run adaptation to heat. If large firms are better able to adapt in the long-run, they may be able to reverse the short-run distributional effects observed in Figure 2. To test this hypothesis, we estimate equation (7), allowing the impact of temperature change in the long-differences equation to vary flexibly 27 by initial average firm size. In practice, we use local kernel-weighted regressions to estimate the θk functions, with an adaptive bandwidth that varies along the distribution of initial firm size depending on the density of the data. Figure 4: Firm size impacts across the distribution: long difference kernel regressions Impact of 1C temperature shock Impact of 1C temperature shock Impact of 1C temperature shock .5 .5 .5 0 0 0 -.5 -.5 -.5 -1 -1 -1 -1.5 -1.5 -1.5 0 .5 1 1.5 2 2.5 0 .5 1 1.5 2 2.5 0 .5 1 1.5 2 2.5 Log initial firm size Log initial firm size Log initial firm size 8-year long difference, No FE 15-year long difference, No FE 23-year long difference, No FE Note: Figure shows estimates and 95% confidence intervals of the impact of a long-run one degree Celsius increase in temperature along the initial firm size distribution. Estimates are presented for each long difference lag k. Standard errors clustered at the village level. Long differences are calcu- lated at the village level as the difference between t and t − k in i ) the average annual temperature in the two years before the census year and ii ) the average firm size. θk functions are estimated using kernel-weighted local linear regression with adaptive bandwidth selection. Sample is village-level data of the last four Economic Censuses of India (1990, 1998, 2005, and 2013). The results are in Figure 4. Each panel plots the impact of a 1°C long-run temperature change on the change in average firm size, θk , estimated nonparametrically at each point along the (logged) distribution of initial firm size yv,t−k . These effect curves are plotted separately for each long-difference lag k (8, 15, and 23 years), and are estimated without fixed effects. The 95% confidence bands are derived from standard errors clustered at the SHRID level. The results provide clear evidence of a long-run reversal of the distributional consequences of heat, likely driven by large firms’ superior adaptive capacities. In the 8-year estimates, the effect curve mirrors the quantile results of Figure 2. The (villages with) larger firms tend to see larger contractions in firm size: for villages at the 95th percentile of the firm size distribution (see Table A1) the effect of a 1°C increase in temperature is to reduce employment by an av- erage of roughly 0.75 worker per firm, or 15%. For villages at the median, this effect is almost zero, though large samples yield a small and precisely estimated negative effect. While all experience somewhat negative effects, as with the quantile results, the effect curve remains relatively flat until around the 75th percentile, after which it slopes downward dramatically. 28 However, these trends moderate, and ultimately reverse, with longer lags in the differ- ences. For k − 15, plotted in the middle panel of Figure 4, the slope flattens. The impact of a 1°C long-run increase in temperature at the 95th percentile of yv,t−k is now less than 0.5, while the estimates below the median firm size are similar to the 8-year estimates. This suggests that as time passes, larger firms begin making the necessary adaptive investments, and as a consequence narrow the gap in firm size effects across the distribution. Finally, the right-hand panel shows estimates of θk for k = 23, the longest difference avail- able in our data. Remarkably, the effect curve inverts completely – now the negative impacts of heat on firm size are largest for the smallest firms. As in other lags, a 1°C increase in long- run temperature reduces firm size by a small amount at the 5th percentile of the firm size distribution, and the effect remains negative and significant until just above the median firm size. However, by the 95th percentile of the distribution, the effect of a 1°C temperature rise is to increase firm size, though this is not statistically different from zero. Taken together, the results suggest that it takes several decades for large firms to adapt and the labor market to adjust to a heat shock. This implies potentially large aggregate adjustment costs entailing long periods of heightened misallocation, as the informal sector swells to absorb labor from large firms. These costs are likely to be nontrivial, and since temperature increases continu- ously, these costs will be spread out over time, rather than just experienced in the aftermath of an individual shock. 7 Robustness and alternative explanations Throughout, out core argument is that declining labor productivity reduces labor demand, leading large firms to contract. Displaced labor reallocates to small, unproductive, informal firms, reducing aggregate productivity in the short-run. This allocative effect dissipates in the long run, as larger firms adapt to heat. However, our results could be driven by inappropriate specification choices, measurement error, firm exit, labor supply responses such as migration and government assistance, product demand effects, or capital-labor substitution. In this section, we consider each of these explanations in turn. However, we note that none of these alternative hypotheses can fully explain both the short and long-run effects observed in our data. 29 Heat shock specification choice: Our results may be sensitive to specification choices for the heat shock. We consider three robustness to specifying the heat shock. In Table A5, we find similar results when defining the temperature anomaly using average or minimum daily temperatures. In Table A6, we calculate the anomaly over lags of one and three years instead of two. We find that the effect holds only for three-year lags but not only the census- year, consistent with increased noise from year-to-year temperature fluctuation and limited response time for firms when the shock is measured in the contemporaneous year. Finally, in Table A7 we estimate a model in which temperature enters semi-parametrically as the number of days in each temperature interval. The estimates reveal that firm size remains a downward sloping function of temperature, consistent with the main results. The largest and most significant effects on firm size occur in the ≥ 40 degree bin. Nonlinear heat effects: We also test whether effects of temperature vary by initial tem- perature, since differential impacts of heat shocks between hotter versus colder places can be interpreted as evidence of long-run adaptation Barreca et al. (2015). In Table A8, we interact the temperature anomaly in our main specification with an indicator variable equaling one if the long-term average daily maximum temperature is 30C or greater. We find mixed evidence of systematic differences in firm size effects for hotter and cooler places—in the most rigorous specification (column 5), these differences shrink to zero, suggesting limited long-run adap- tation.18 Model specification choices. Table A9 tests the robustness of the main short-run effects result to model specification choices. In Panel A, we control for precipitation anomalies, con- structed identically to our temperature anomalies. The results are essentially unchanged. Next, Panel B estimates an exponential model with a fixed effects Poisson regression estima- tor, to account for outliers given the skewness of the firm size distribution in India, with a large mass at 1, as well as the discrete count nature of the outcome. Estimates from the Pois- son regression, interpreted as percent changes, are similar in relative magnitude to the main estimates, and again significant in the same three out of five specifications. Urban heat islands: Our results may also be driven by the urban heat island effect. Be- cause of the UHI, urban areas tend to be hotter, and large firms cluster in urban areas. In 18 Note that limited long-run adaptation on average does not preclude substantial long run adaptation at the tail of the firm size distribution, as shown in Section 6.3. 30 a naive specification, this would lead us to underestimate the negative effect of heat; how- ever, our use of temperature anomalies and fixed effect strategy address this concern. A more subtle issue is that a measured 1°C anomaly may feel hotter in urban areas, which may not be captured in the relatively coarse temperature data, which can underestimate tem- perature extremes. This correlated measurement error could generate the illusion of larger negative effects for large firms, when in fact these large urban firms simply experience larger unmeasured heat shocks. We test this hypothesis by splitting the sample into urban and rural SHRIDs and re-running both the fixed effects (Table A10) and quantile regressions (Figure A7) on each subsample. The main results hold in both subsamples. Entry and exit: A plausible explanation for the short-run results in Section 4.2 is that small and large firms respond identically to heat, but small firms exit the market while large firms contract but remain active. This explanation is mechanical, rather than behavioral given the large density of one-person firms in the data. To rule this out, Table A13 contains estimates of the total number of firms in each firm size bin as a function of the heat shock at the subdistrict level. Using an exponential PPML estimator to account for the highly right-skewed count outcome, we observe small and statistically insignificant effects for all firm size groups.19 Migration: It is possible that heat shocks causes migration (Mueller et al., 2014; Baez et al., 2017), which forces firms to contract due to labor shortage. In this case, contraction is driven by a labor supply rather than labor demand effect. To test this hypothesis, we use census data to estimate the relationship between annual migration rates and temperature anomalies at the district level. Figure A8 displays the results, which show no relationship between heat shocks on migration. Government workfare programs: Another labor supply explanation is that government targets welfare programs to villages most affected by heat, increasing the opportunity cost of private sector work. Evidence shows that participation in NREGA, an employment guarantee workfare scheme, increases in heat-affected areas Garg et al. (2020). We replicate this result in Table A14, with evidence of meaningful increases in NREGA activity in response to heat shocks. However, in Table A15, we interact the log of district-level NREGA participation 19 The log-linear specification in Panel A shows a small but statistically significant 4% reduction in the number of 1 person firms, though this is accompanied by an even larger 6-7% reduction in the number of large firms, suggesting exit across the board rather than differentially by size. We also note that the PPML estimator is likely to outperform the log-linear if the functional form is exponential. 31 with the temperature anomaly in our main regression specification. There is no evidence that villages with higher NREGA enrollment observe larger firm size effects, suggesting that labor supply does not drive our results. Demand-side effects: If household consumption of the products made by large firms is particularly sensitive to heat, then our quantile effects may be driven by a product rather than labor demand channel. Following Caggese et al. (2024), we test for demand effects by leveraging differences across sectors in tradeability, based on the logic that traded sectors are less sensitive to local demand shocks. Table A16 interacts the heat anomaly with a sectoral export exposure index. The results show no evidence that firm size effects are smaller for firms in tradeable sectors. Capital-labor substitution: Another explanation for our quantile results is that large firms adapt to heat shocks by shifting toward more capital-intensive technologies—an optimal re- sponse when heat disproportionately reduces labor relative to capital productivity, as doc- umented by Somanathan et al. (2021)—whereas smaller firms lack this option due to credit constraints or scale limitations. Since the Economic Census does not contain capital invest- ment, we draw on the World Bank South Asia Climate Adaptation Survey (2024–25), which asked managers in Indian firms about protective actions taken against heat (Lang et al., eds, 2025). Figure A9 shows that larger firms are more likely to invest in building upgrades and cooling technology, but also more likely to address heat concerns by hiring workers.20 Thus, while bigger firms invest more in capital upgrades, these do not substitute for labor. 8 Conclusion Firms in low and middle income countries are increasingly vulnerable to climate shocks. These countries are also characterized by highly skewed firm size distributions and corre- spondingly high levels of misallocation, where a large mass of unproductive small firms coexist with a small number of large firms operating at the international productivity fron- tier. While the negative effects of climate change on firm-level productivity are by now well- studied, the implications for allocative efficiency are not well understood, particularly beyond 20 Note, however, that the absence of firing in this data does not necessarily conflict with our overriding labor demand explanation, since the survey questions are phrased to elicit ex-ante adaptations rather than the ex-post responses to heat that we estimate. 32 reallocation between agriculture and industry. This paper represents an initial attempt to understand this fundamental problem in the context of India, a country facing severe climate risk and exhibiting a high degree of mis- allocation. Using fixed effects models, we show that a 1°C temperature shock reduces non- agricultural firm size by 11.6% on average. However, this effect is highly heterogeneous, and disproportionately affects the largest firms. While temperature shocks do induce reallocation across sectors, this does not drive the observed firm size heterogeneity. Rather, large firms contract, and small firms expand, conditional on highly disaggregated sector fixed effects. Aggregate labor market results suggest reallocation of labor from the formal to informal sec- tor as large firms shed workers in the face of rising heat. This short-run reallocation from formal to informal firms accounts for a 1.3-4.3 percent aggregate productivity loss for each degree Celsius of warming. However, in the long-run, we show evidence that larger firms adapt to rising temperatures, labor markets adjust accordingly, and ultimately the dispropor- tionate incidence of climate change on large firms reverses. The results are consistent with the role of the informal sector as a safety net in the face of negative labor demand shocks. We also provide new evidence on the transitional allocative efficiency costs of climate change when the largest, most productive firms fail to swiftly adapt. But the results suggest that more research is needed on how firms adapt and change input choices, and the frictions that cause substantial lags in adaptation. More evidence on whether size-dependent regulations with stronger compliance requirements triggered at specific firm- size thresholds, lengthy approval processes for capital upgrades or zoning restrictions make it harder for large, formal firms to adjust rapidly to heat shocks would have useful policy implication. In addition, the role of credit constraints and limited access to electricity and other infrastructure in impeding effective long-run adaptive responses among small, informal firms deserve more exploration. We view these as promising avenues for future research. 33 References Adhvaryu, Achyuta, A. V. Chari, and Siddharth Sharma, “Firing Costs and Flexibility: Evi- dence from Firms’ Employment Responses to Shocks in India,” The Review of Economics and Statistics, 07 2013, 95 (3), 725–740. , Namrata Kala, and Anant Nyshadham, “The Light and the Heat: Productivity Co- Benefits of Energy-Saving Technology,” The Review of Economics and Statistics, 10 2020, 102 (4), 779–792. , , and , “Management and Shocks to Worker Productivity,” Journal of Political Economy, 2022, 130 (1), 1–47. Amirapu, Amrit and Michael Gechter, “Labor Regulations and the Cost of Corruption: Ev- idence from the Indian Firm Size Distribution,” The Review of Economics and Statistics, 03 2020, 102 (1), 34–48. Asher, Sam, Tobias Lunt, Ryu Matsuura, and Paul Novosad, “The Socioeconomic High- resolution Rural-Urban Geographic Dataset on India (SHRUG),” International Growth Centre Working Paper, 2019. Baez, Javier, German Caruso, Valerie Mueller, and Chiyu Niu, “Heat Exposure and Youth Migration in Central America and the Caribbean,” American Economic Review, May 2017, 107 (5), 446–50. Balboni, Clare, Johannes Boehm, and Mazhar Waseem, “Firm Adaptation in Production Networks: Evidence from Extreme Weather Events in Pakistan,” Working Paper, London School of Economics 2023. Barreca, Alan, Karen Clay, Olivier Deschênes, Michael Greenstone, and Joseph S. Shapiro, “Convergence in adaptation to climate change: Evidence from high temperatures and mor- tality, 1900–2004,” American Economic Review, 2015, 105 (5), 247–251. Publisher: American Economic Association 2014 Broadway, Suite 305, Nashville, TN 37203. Basker, Emek and Javier Miranda, “Taken by storm: business financing and survival in the aftermath of Hurricane Katrina,” Journal of Economic Geography, 08 2017, 18 (6), 1285–1313. 34 Bento, Antonio, Noah Miller, Mehreen Mookerjee, and Edson Severnini, “A unifying ap- proach to measuring climate change impacts and adaptation,” Journal of Environmental Eco- nomics and Management, 2023, 121, 102843–102843. Berg, C., L. Bettarelli, D. Furceri, M. Ganslmeier, A. Grover Goswami, M. Lang, and M. Schiffbauer, “Firm-Level Climate Change Adaptation: Micro Evidence from 134 Na- tions,” Policy Research Working Paper 11081, World Bank, Washington, DC March 2025. Biesebroeck, Johannes Van, “Firm Size Matters: Growth and Productivity Growth in African Manufacturing,” Economic Development and Cultural Change, 2005, 53 (3), 545–583. Bilal, Adrien and Diego R Känzig, “The Macroeconomic Impact of Climate Change: Global vs. Local Temperature,” Working Paper 32450, National Bureau of Economic Research May 2024. Burgess, R., O. Deschenes, D. Donaldson, and M. Greenstone, “Weather, Climate Change and Death in India,” LSE Working Paper, London School of Economics and Political Sci- ence, London, UK 2017. Burke, M., S. M. Hsiang, and E. Miguel, “Global Non-Linear Effect of Temperature on Eco- nomic Production,” 2015, 527 (7577), 235–239. Burke, Marshall and Kyle Emerick, “Adaptation to Climate Change: Evidence from US Agri- culture,” American Economic Journal: Economic Policy, August 2016, 8 (3), 106–40. and Vincent Tanutama, “Climatic Constraints on Aggregate Economic Output,” Working Paper 25779, National Bureau of Economic Research April 2019. Bussolo, M. and S. Sharma, Hidden Potential: Rethinking Informality in South Asia, World Bank, Washington DC., 2022. Caggese, Andrea, Andrea Chiavari, Sampreet Goraya, and Carolina Villegas-Sanchez, “Cli- mate Change, Firms, and Aggregate Productivity,” CEPR Discussion Paper No. 19164, 2024. Carleton, Tamma, Esther Duflo, B. Kelsey Jack, and Guglielmo Zappalà, “Chapter 4 - Adap- tation to climate change,” in Lint Barrage and Solomon Hsiang, eds., Lint Barrage and 35 Solomon Hsiang, eds., Vol. 1 of Handbook of the Economics of Climate Change, North-Holland, 2024, pp. 143–248. Carrillo, Paul, Dave Donaldson, Dina Pomeranze, and Monica Singhal, “Misallocation in Firm Production: A Nonparametric Analysis Using Procurement Lotteries,” Working Pa- per, Massachusets Institute of Technology 2024. Chaurey, Ritam, “Labor regulations and contract labor use: Evidence from Indian firms,” Journal of Development Economics, 2015, 114, 224–232. Colacito, Riccardo, Bridget Hoffman, and Toan Phan, “Temperature and Growth: A Panel Analysis of the United States,” Journal of Money, Credit and Banking, 2019, 51 (2-3), 313–368. Colmer, Jonathan, “Temperature, Labor Reallocation, and Industrial Production: Evidence from India,” American Economic Journal: Applied Economics, October 2021, 13 (4), 101–24. CSO, “Sixth Economic Census of India: Instructions to Enumerators,” Technical Report, Cen- tral Statistical Organization, Ministry of Statistics and Programme Implementation, New Delhi, India 2013. de Mel, Suresh, David J. McKenzie, and Christopher Woodruff, “Measuring microenter- prise profits: Must we ask how the sausage is made?,” Journal of Development Economics, 2009, 88 (1), 19–31. Dell, Melissa, Benjamin F. Jones, and Benjamin A. Olken, “Temperature Shocks and Eco- nomic Growth: Evidence from the Last Half Century,” American Economic Journal: Macroe- conomics, July 2012, 4 (3), 66–95. Dix-Carneiro, Rafael, Pinelopi K Goldberg, Costas Meghir, and Gabriel Ulyssea, “Trade and Domestic Distortions: the Case of Informality,” Working Paper 28391, National Bureau of Economic Research January 2021. Gallagher, Justin, Daniel Hartley, and Shawn Rohlin, “Weathering an Unexpected Finan- cial Shock: The Role of Federal Disaster Assistance on Household Finance and Business Survival,” Journal of the Association of Environmental and Resource Economists, 2023, 10 (2), 525–567. 36 Garg, Teevrat, Maulik Jagnani, and Vis Taraz, “Temperature and Human Capital in India,” Journal of the Association of Environmental and Resource Economists, 2020, 7 (6), 1113–1150. Goicochea, A. and M. Lang, “Firms and Climate Change in Low- and Middle-Income Coun- tries,” World Bank Policy Research Working Papers, 2023, (10644). Gollin, Douglas and Christopher Udry, “Heterogeneity, Measurement Error, and Misalloca- tion: Evidence from African Agriculture,” Journal of Political Economy, 2021, 129 (1), 1–80. , David Lagakos, and Michael E. Waugh, “Agricultural Productivity Differences across Countries,” American Economic Review, May 2014, 104 (5), 165–70. Hainmueller, Jens, Jonathan Mummolo, and Yiqing Xu, “How Much Should We Trust Es- timates from Multiplicative Interaction Models? Simple Tools to Improve Empirical Prac- tice,” Political Analysis, 2019, 27 (2), 163–192. Haltiwanger, John, Robert Kulick, and Chad Syverson, “Misallocation Measures: The Dis- tortion That Ate the Residual,” Working Paper 24199, National Bureau of Economic Re- search January 2018. Heal, Geoffrey and Jisung Park, “Reflections—Temperature Stress and the Direct Impact of Climate Change: A Review of an Emerging Literature,” Review of Environmental Economics and Policy, 2016, 10 (2), 347–362. Hersbach, Hans, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquín Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, Adrian Simmons, Cornel Soci, Saleh Abdalla, Xavier Abellan, Gianpaolo Balsamo, Peter Bechtold, Gionata Biavati, Jean Bidlot, Massimo Bonavita, Giovanna De Chiara, Per Dahlgren, Dick Dee, Michail Diamantakis, Rossana Dragani, Johannes Flemming, Richard Forbes, Manuel Fuentes, Alan Geer, Leo Haimberger, Sean Healy, Robin J. Hogan, Elías Hólm, Marta Janisková, Sarah Keeley, Patrick Laloyaux, Philippe Lopez, Cristina Lupu, Gabor Radnoti, Patricia de Rosnay, Iryna Rozum, Freja Vamborg, Se- bastien Villaume, and Jean-Noël Thépaut, “The ERA5 global reanalysis,” Quarterly Journal of the Royal Meteorological Society, 2020, 146 (730), 1999–2049. 37 Hopenhayn, Hugo A., “Firms, Misallocation, and Aggregate Productivity: A Review,” An- nual Review of Economics, 2014, 6 (Volume 6, 2014), 735–770. Hsieh, C. and P. J. Klenow, “Misallocation and Manufacturing TFP in China and India,” The Quarterly Journal of Economics, 2009, 124, 1403–1448. Hsieh, Chang-Tai and Benjamin A. Olken, “The Missing "Missing Middle",” Journal of Eco- nomic Perspectives, September 2014, 28 (3), 89–108. and Peter J. Klenow, “The Life Cycle of Plants in India and Mexico,” The Quarterly Journal of Economics, August 2014, 129 (3), 1035–1084. Jayachandran, Seema, “Selling Labor Low: Wage Responses to Productivity Shocks in Devel- oping Countries,” Journal of Political Economy, 2006, 114 (3), 538–575. Kala, Namrata, Clare Balboni, and Shweta Bhogale, “Climate Adaptation,” VoxDevLit, 2023, June 2023. Kanbur, Ravi, “Informality: Causes, consequences and policy responses,” Review of Develop- ment Economics, 2017, 21 (4), 939–961. Kumar, Rahul, Jayanarayanan Kuttippurath, G. S. Gopikrishnan, Pankaj Kumar, and Hamza Varikoden, “Enhanced surface temperature over India during 1980–2020 and fu- ture projections: causal links of the drivers and trends,” npj Climate and Atmospheric Science, October 2023, 6 (1), 164. Lang, Megan, Jonah Rexer, Siddharth Sharma, and Maggie Triyana, eds, From Risk to Re- silience: Helping Households and Firms in South Asia Adapt, The World Bank, Washington DC, 2025. , , , and Margaret Triyana, From Risk to Resilience: Helping People and Firms Adapt in South Asia, Washington, DC: World Bank, June 2025. Machado, José A.F. and J.M.C. Santos Silva, “Quantiles via moments,” Journal of Economet- rics, 2019, 213 (1), 145–173. Annals: In Honor of Roger Koenker. McCaig, Brian and Nina Pavcnik, “Export Markets and Labor Allocation in a Low-Income Country,” American Economic Review, July 2018, 108 (7), 1899–1941. 38 Mueller, Valerie, Clark Gray, and Katrina Kosec, “Heat stress increases long-term human migration in rural Pakistan,” Nature Climate Change, March 2014, 4 (3), 182–185. Nataraj, Shanthi, “The impact of trade liberalization on productivity: Evidence from India’s formal and informal manufacturing sectors,” Journal of International Economics, 2011, 85 (2), 292–301. Nath, Ishan B, “The Food Problem and the Aggregate Productivity Consequences of Climate Change,” Working Paper 27297, National Bureau of Economic Research June 2020. Nath, Ishan B., Valerie A. Ramey, and Peter J. Klenow, “How Much Will Global Warming Cool Global Growth?,” Working Paper, Stanford 2024. Pelli, Martino and Jeanne Tschopp, “Comparative advantage, capital destruction, and hur- ricanes,” Journal of International Economics, 2017, 108, 315–337. , , Natalia Bezmaternykh, and Kodjovi M. Eklou, “In the eye of the storm: Firms and capital destruction in India,” Journal of Urban Economics, 2023, 134 (C). Ponczek, Vladimir and Gabriel Ulyssea, “Enforcement of Labour Regulation and the Labour Market Effects of Trade: Evidence from Brazil,” The Economic Journal, 06 2021, 132 (641), 361–390. Porta, Rafael La and Andrei Shleifer, “Informality and Development,” Journal of Economic Perspectives, September 2014, 28 (3), 109–26. Restuccia, Diego and Richard Rogerson, “The Causes and Costs of Misallocation,” Journal of Economic Perspectives, August 2017, 31 (3), 151–74. Rexer, Jonah and Siddharth Sharma, “Climate Change Adaptation : What Does the Evidence Say?,” World Bank Policy Research Working Papers, 2024, (10729). Somanathan, E., Rohini Somanathan, Anant Sudarshan, and Meenu Tewari, “The Impact of Temperature on Productivity and Labor Supply: Evidence from Indian Manufacturing,” Journal of Political Economy, 2021, 129 (6), 1797–1827. Terrell, George R. and David W. Scott, “Variable Kernel Density Estimation,” The Annals of Statistics, September 1992, 20 (3). Publisher: Institute of Mathematical Statistics. 39 Ulyssea, Gabriel, “Informality: Causes and Consequences for Development,” Annual Review of Economics, 2020, 12 (Volume 12, 2020), 525–546. Wooldridge, Jeffrey M., “Distribution-free estimation of some nonlinear panel data models,” Journal of Econometrics, 1999, 90 (1), 77–97. WorldBank, “Climate Change Knowledge Portal: Climate Projections,” Technical Report, The World Bank, Washington DC 2024. Xie, Victoria Wenxin, “Heterogeneous firms under regional temperature shocks: exit and reallocation, with evidence from Indonesia,” Economic Development and Cultural Change, 01 2024, 72 (2), 659–690. , “Labor market adjustment to extreme heat shocks: Evidence from Brazil,” Journal of Eco- nomic Behavior Organization, 2024, 222, 266–283. 40 ONLINE APPENDIX A Data appendix A.1 Temperature data A.1.1 Data Source The core unit of observation in our climate data is the village, or SHRID, derived from unique spatial identifiers at the lowest possible level (in both rural and urban areas) from the Socioeconomic High-resolution Rural-Urban Geographic Platform for India (SHRUG Asher et al. 2019). We compile daily weather data for 576,153 villages in India spanning 1980–2021. Villages are uniquely identified by the SHRID, a code constructed from state, district, sub- district, and village identifiers. Temperature readings are obtained obtained from the ERA5 reanalysis dataset, which provides maximum, mean, and minimum surface air temperature at a spatial resolution of 0.25° × 0.25° (Hersbach et al., 2020). In addition, daily mean precip- itation data are used [SOURCE TO BE INSERTED]. ERA5 is widely employed in historical climate reconstruction and provides consistent global coverage. Each village-level SHRID is mapped to the nearest ERA5 grid cell using village-level shapefiles from SHRUG. A.1.2 Data Processing and measurement For each daily temperature series (min, mean, and max), we construct annual aggre- gates for each SHRID. In addition, to capture the frequency distribution of temperature be- yond averages, we classified each daily temperature observation into predefined bins (< 0C, 0 − 10C, 10˘20C, 20 − 30C, 30 − 40C, > 40C). For each SHRID and year, we then calcu- lated the number of days falling into each bin. These bin-based measures provide a discrete characterization of thermal exposure, complementing the continuous averages and allowing for non-linear and semi-parametric models, following Burgess et al. (2017), among others. Our main variable of interest—the temperature anomalies—measures deviation from a long-run baseline climate. To establish reference conditions, we calculated long-term climato- logical means for each SHRID by computing the 1980–2021 averages (across years) of annual maximum, mean, and minimum temperatures. These serve as the baseline climatological 41 norms against which deviations and anomalies are measured. Villages are classified as “hot locations” if their long-term average maximum temperature exceeds 30°C. max be annual We define temperature anomalies relative to the long-term climatology. Let Tit average maximum temperature for village i at year t. Then the anomaly Ait is given by: 1 max Ait = T + Timax ,t −1 − T max i,1980-2021 2 it Note that we use the two-year average of T max for t and t − 1 in calculating the anomaly. This is done both to smooth out year-to-year variation in temperature, which is substantial (see Figure A1), but also to account for the fact that firm labor demand responses to heat shocks may occur with a lag. In Appendix C.4, we test the robustness of our results to nu- merous measurement assumptions. The final dataset is a balanced panel at the SHRID level covering 42 years (1980–2021). For each village-year, the dataset includes the max temper- ature anomalies and number of days in different temperature bins for each village in India, with the unique SHRID used to merge with other data sets. A.2 Economic census data A.2.1 Data Source We utilize establishment-level records from the Indian Economic Census (EC) for the 1990, 1998, 2005, and 2013 rounds. The Economic Census, conducted at 8-year intervals by the Ministry of Statistics and Program Implementation, collects basic data on the universe of commercial establishments (what we call “firms”) in India. Each record corresponds to an individual commercial establishment and contains information on employment size, industry classification, ownership type, and owner demographics. The raw data contains information on 21,806,546 firms in 1990, 26,772,187 firms in 1998, 35,795,897 firms in 2005 and 46,462,994 firms in 2013. We obtain the raw, firm-level census data from SHRUG, who also produce weighted aggregates at varying spatial levels. Note that firms are not uniquely identified or followed with a panel identifier throughout the EC, which we therefore treat as a repeated cross section. 42 A.2.2 Data processing and merging The raw EC files contain geographic identifiers for each establishment, including state, district, subdistrict, block, town, ward and village codes, along with the firm responses to census questions. In order to map these geographic identifiers to the unique SHRID code, SHRUG provides a set of keys for both rural and urban areas that match firms to SHRIDs on the combination of state, district, subdistrict, block, town, ward and village codes. This procedure allows us to match firm-level observations with village-level weather data. Because the EC provides separate lookup keys for rural (village-based) and urban (town- based) SHRIDs, we perform two merges for each establishment record in the EC: (i ) a merge on {state, district, subdistrict, town} with the urban key to obtain a candidate urban identifier SHRID2U , and (ii ) a merge on {state, district, subdistrict, village} with the rural key to obtain a candidate rural identifier SHRID2R . We then coalesce these candidates to a single spatial identifier according to the following rule for firm observation i:  SHRIDU if SHRIDU R  i is non-missing and SHRIDi is missing,  i      SHRIDi = SHRIDiR if SHRIDiR is non-missing and SHRIDU i is missing,      missing  if both are missing. In the rare case where both identifiers are present (e.g., along rural–urban administrative boundaries), we merge the weather data using both SHRIDs and take the average of the resulting measures. After coalescing, we drop records with missing SHRIDi and retain a harmonized set of variables—including ownership type (with a private-sector indicator), power source, caste, sex, total and female employment, financing mode, and census year, with consistent coding and value labels across census rounds. Finally, using the SHRID mapping, firm-level records are merged with the SHRID-level weather data. This linkage assigns to each firm the climate anomalies, long-term climatological averages, and temperature-bin measures corresponding to its village of location. The firm-level dataset thus provides, for each establishment and census round, both microeconomic characteristics (employment size, sector of main output, and ownership type) and local climatic conditions. 43 A.2.3 Sample selection For the purposes of our analysis, we focus exclusively on non-agricultural, private-sector firms. First, we drop all records with missing SHRIDi values, as this spatial identifier is required to merge establishments with the weather data. Second, we exclude agricultural firms using SHRIC industry codes. The SHRIC (SHRUG Industrial Classification) codes were constructed by SHRUG to harmonize India’s National Industrial Classification (NIC) codes across different NIC revisions. The SHRIC system harmonizes using 3-digit NIC08 codes to facilitate concordance with NIC 2004 and NIC 1987. There are 90 3-digit SHRIC codes. Finally, we further restrict the sample to private-sector establishments. The definition of “private” varies across census rounds. In 1990, private firms include all establishments cat- egorized as privately owned or organized under cooperative models. In 1998, the definition expands to include all non-government establishments. The 2005 and 2013 rounds follow a similar classification, where all non-government establishments are treated as private firms. This harmonization ensures consistency in our definition of the private sector across time. After merging with climate data and restricting the sample to non-agricultural establish- ments with valid SHRIC industry codes and to private-sector firms as defined for each census round, the final dataset contains 110.6 million unique establishment-year-level records across the four rounds of the Economic Census. Specifically, there are 15.1 million firms in 1990, 22.5 million firms in 1998, 31 million firms in 2005, and 42 million firms in 2013 (Table A1, Panel A). Note that given varying merge rates and sample restrictions, not all of the original 576,153 villages are matched to firms in our final sample. Final sample sizes for villages— which represents the number of clusters for our main identifying variation—are in Table A1, Panel D. A.3 ASI-NSS firm-level data A.3.1 Data sources Apart from the Economic Census, which does not collect revenue and expenditure data, there is no firm-level survey that collects representative data on both informal micro-enterprises and larger, more formalized firms in India. As a result, past studies such as (Nataraj, 2011; Hsieh and Klenow, 2014; Hsieh and Olken, 2014) have relied on a combination of two surveys 44 to estimate labor productivity differentials between informal and formal firms, an approach that we also follow. Our data source for the formal sector is the 2015 round of the Annual Survey of Industries (ASI). The ASI, conducted annually by India’s Central Statistical Organi- zation, is a detailed plant-level survey of manufacturing establishments registered under The Factory Act (1948). The ASI consists of a census of plants with employment above a threshold of 300 and a random sample of smaller plants.21 Our data source for the informal sector is the 2015-16 (73rd Round) Survey of Non-Agricultural Unincorporated Enterprises conducted by the National Sample Survey Organization. Until 2020, this survey was conducted about once every five years. The NSS survey covers man- ufacturing establishments not registered under the Factories Act (1948), as well as service sector establishments, but excludes construction firms. A.3.2 Sample selection From each dataset, we obtain firm-level ARPL, calculated as total value added divided by the number of employees. We define “formal” firms as those with 8 or more workers. For consistency with ASI, we drop non-manufacturing firms from the NSS dataset before combining them. We also drop establishments with employment greater than 9 from the NSS dataset and less than 10 from the ASI dataset.22 Lastly, we winsorize value added at the top and bottom 0.1 percentile level in both ASI and NSS data to mitigate the influence of extreme outliers in value added.23 After combining the ASI and NSS datasets, we obtain 134,480 firm- level observations for which we can observe value-added per worker and employment size. A.4 Other data sources A.4.1 District-Level NREGA Data We incorporate district-level data on the National Rural Employment Guarantee Act (NREGA) from Garg et al. (2020). These data provide annual measures of NREGA enrollment at the dis- 21 The cutoff for the "census" segment of the ASI has varied over the years. 22 Bydesign, since all manufacturing establishments with 10 or more workers (and using electricity, which almost all firms do) are required to register under the Factories Act, there should be no overlap between the NSS and ASI beyond the employment cutoff of 10. In practice, outdated sampling frames and potential regulatory non-compliance lead to overlaps (Kanbur, 2017). 23 Our estimates are not sensitive to varying the winsorization cutoff between 0.01 and 1 percentile. 45 trict level, as well as a corresponding harmonized district shapefile which fix the geographic boundaries of districts at one moment in time. Each village in our dataset is assigned to a district using the NREGA district shapefile. This mapping allows us to merge firm-level Eco- nomic Census records with district-level NREGA statistics through the SHRID village identi- fier. The NREGA data contains the following district-year participation variables: the number of households, person-days, labor expenditure, and materials expenditure of the program. Since NREGA enrollment is reported annually from 2006-2014, the Economic Census rounds that overlap this period are 2005 and 2013. We construct smoothed measures by taking the three-year average of NREGA enrollment surrounding each census year. Thus, the NREGA measure for 2005 reflects the average over 2004–2006, and the measure for 2013 reflects the av- erage over 2012–2014. In total, there are 75,268,119 firm-years observed in the 2005 and 2013 Economic Census rounds. Of these, 44,825,100 firm-level observations are located in villages that can be mapped to NREGA districts using the shapefile. This reflects the fact that NREGA was not rolled out nationwide simultaneously, and certain urban and non-program districts fall outside its coverage. A.4.2 Trade Exposure Data We construct sector-level trade exposure measures using the Asian Development Bank Multi-Regional Input–Output Tables (ADB MRIOT). The ADB MRIOT provides trade-to- output ratios for India across detailed sectors, including imports-to-gross output (MtGO), exports-to-gross output (XtGO), and total trade-to-gross output (XMtGO). These indicators quantify the extent to which industries are integrated into international markets. To integrate the ADB MRIOT trade exposure measures with the Economic Census, we first map MRIOT industry codes to ISIC Rev. 3 (three-digit) industries using the concordance provided in the MRIOT documentation. We then translate ISIC Rev. 3 codes into India’s Na- tional Industrial Classification (NIC-08), which is the industry classification system used for the EC SHRIC. During this process, certain industry codes were consolidated (for example, NIC 2711–2719 were recoded to 2710) to ensure consistency. Once the MRIOT sectors are mapped into SHRIC, we assign to each NIC code its corresponding values of MtGO, XtGO, and XMtGO. These SHRIC-level exposure measures are then merged directly with the firm- level Economic Census using each establishment’s reported industry code. Thus, every firm 46 inherits the trade exposure values of its industry. Importantly, trade exposure is not con- structed at the SHRID (village) level; rather, it remains at the SHRIC (industry) level and is common across all firms in the same industry. A.4.3 Migration Data We use migration statistics from the Population Census of India for the years 2001 and 2011. The census reports origin district-level tabulations of migrants by gender, age, reason for migration, and years since migration. For the purposes of this study, we focus exclusively on individuals who reported work or business as their reason for migration. For both males and females, the census distinguishes work/business migrants by years since migration into three groups: migrated within the last 0–5 years, migrated 6–10 years ago, and migrated more than 10 years ago. We construct total male and female work-related migrants by summing across all age groups and genders for each migration period. To merge migration outcomes with climate exposures, we aggregate the SHRID-level weather data to the district level annually for the period 1980–2021. For each district and year, we collapse maximum temperature anomalies, weighting by SHRID population. Migration is reported only in 2001 and 2011, while weather data are annual. However, since each mi- gration census contains a count of migrants who migrated in two distinct periods (within 5 years, and 5-10 years), we construct four periods in a pseudo-panel: 1992-1996, 1997-2001, 2002-2006, 2007-2011. We then merge to weather data at the district-period level, taking the time-averaged temperature anomaly in each period. All district identifiers are harmonized to the 2011 census codes. This involves standardiz- ing district names, correcting spelling inconsistencies, and mapping the 2001 district codes to 2011 boundaries. Adjustments are made for districts that were split or renamed between the two rounds. After harmonization, the dataset covers a total of 640 districts in a total of 26800 observation from 1980 to 2021. For 26800 observations, 12290 observations are matched with district-level information and migration data, such that one row is one district at one year. 47 Table A1: Summary statistics Census year Mean SD P25 P50 P75 P95 P99 Count Panel A: Firm size, all firms 1990 2.28 45.11 1 1 2 5 14 15.1 1998 2.46 30.08 1 1 2 6 15 22.5 2005 2.21 11.45 1 1 2 5 10 31.0 2013 2.27 24.35 1 1 2 5 11 42.0 Panel B: Firm size, services firms 1990 2.05 43.73 1 1 2 5 12 11.6 1998 2.25 29.41 1 1 2 5 13 18.0 2005 2.08 10.01 1 1 2 5 9 24.8 2013 2.19 24.34 1 1 2 5 11 33.0 Panel C: Firm size, manufacturing firms 1990 3.06 50.54 1 2 2 7 23 3.4 1998 3.32 33.22 1 2 3 8 25 4.2 2005 2.74 16.31 1 1 2 6 15 5.9 2013 2.59 24.17 1 1 2 6 14 8.0 Panel D: Maximum temperature anomaly 1990 -0.31 0.17 -0.43 -0.31 -0.20 -0.003 0.09 352146 1998 -0.55 0.38 -0.82 -0.59 -0.31 0.05 0.16 402907 2005 0.10 0.20 -0.02 0.13 0.24 0.39 0.40 478115 2013 -0.22 0.29 -0.41 -0.17 -0.03 0.16 0.45 510262 Note: Figure shows summary statistics for firm-level economic census variables (Panel A-C), as well as village-level climate shocks (Panel D), by census year. Count is in millions of firms for Panels A-C, and in number of villages for Panel D. B Estimating temperature trends We begin by assessing aggregate warming trends across India over our sample period, in part replicating work done by Kumar et al. (2023). Figure A1 plots trends in average an- nual and seasonal surface temperature anomalies in degrees celsius. We disaggregate an- nual anomalies into four seasons corresponding with standard seasonal demarcations in In- dia: pre-monsoon (March–April–May or MAM), monsoon (June–July–August–September or JJAS), post-monsoon (October–November or ON), and winter (December–January–February or DJF). Seasonal anomalies are calculated relative to the long run mean for that season. We 48 find evidence for a clear aggregate warming trend in the post-monsoon period of October and November, but not in other seasons. Figure A1: Seasonal temperature trends Yearly MAM JJAS 1 2 2 Temperature Anomaly(Celsius) Temperature Anomaly(Celsius) Temperature Anomaly(Celsius) .5 1 1 0 0 0 -.5 -1 -1 -2 -1 1980 1985 1990 1995 2000 2005 2010 2015 2020 1980 1985 1990 1995 2000 2005 2010 2015 2020 1980 1985 1990 1995 2000 2005 2010 2015 2020 Year Year Year ON DJF 2 Temperature Anomaly(Celsius) Temperature Anomaly(Celsius) 1 1 0 0 -1 -1 -2 1980 1985 1990 1995 2000 2005 2010 2015 2020 1980 1985 1990 1995 2000 2005 2010 2015 2020 Year Year Maximum Temperature Minimum Temperature Mean Temperature Note: Figure shows estimated annual surface temperature anomalies in degrees celsius relative to the 1980-2021 temperature average, for all of India. Anomalies are estimated for both annual and seasonal temperature averages. MAM - March, April, May; JJAS - June, July, August, September; ON - October, November; DJF - December, January, February. However, this aggregate trend masks substantial spatial heterogeneity across India. We estimate long-run annual trends in temperature change across Indian districts using the fol- lowing time trend regression yit = αi + β i t + ϵit where yit is average annual temperature in district i at year t. We estimate location-specific linear trends for each of the 5868 subdistricts in India for annual min, mean, and max temper- ature readings. Figure A2 then plots the distribution district-specific coefficients across India by temperature reading and region. The map indicates substantial spatial variation across In- dia in warming trends. In some seasons and locations, warming trends are actually negative, 49 whereas in others they are as large as 0.5 degrees every decade. Generally most warming is concentrated in the northern part of the country, where trends are consistently positive. More universal warming trends are observed for minimum and mean than for maximum temper- ature. The post-monsoon and winter seasons also see the most uniform warming trends, consistent with the results of Kumar et al. (2023). Warming in India is uneven, seasonal, and spatially heterogeneous. Figure A2: Histograms of district-level temperature trends Max temperature trend - All regions Mean temperature trend - All regions Min temperature trend - All regions .1 .1 .15 Fraction Fraction Fraction .08 .08 .1 .06 .06 .04 .04 .05 .02 .02 0 0 0 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 Max temperature trend Mean temperature trend Min temperature trend Max temperature trend - NW region Mean temperature trend - NW region Min temperature trend - NW region .1 .1 .15 Fraction Fraction Fraction .08 .08 .1 .06 .06 .04 .04 .05 .02 .02 0 0 0 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 Max temperature trend Mean temperature trend Min temperature trend Max temperature trend - NE region Mean temperature trend - NE region Min temperature trend - NE region .1 .15 .3 Fraction Fraction Fraction .08 .1 .2 .06 .04 .05 .1 .02 0 0 0 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 Max temperature trend Mean temperature trend Min temperature trend Max temperature trend - SW region Mean temperature trend - SW region Min temperature trend - SW region .08 .08 .08 Fraction Fraction Fraction .06 .06 .06 .04 .04 .04 .02 .02 .02 0 0 0 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 Max temperature trend Mean temperature trend Min temperature trend Max temperature trend - SE region Mean temperature trend - SE region Min temperature trend - SE region .1 .15 Fraction Fraction Fraction .08 .06 .08 .1 .04 .06 .04 .05 .02 .02 0 0 0 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 -0.021 -0.011 -0.001 0.009 0.019 0.029 0.039 0.049 Max temperature trend Mean temperature trend Min temperature trend Note: Figure shows histograms of estimated annual subdistrict-level surface temperature trends in degrees Celsius of average temperature change per year, by temperature reading and region. Sample is all 5868 subdistricts in India from 1980-2021. Trends are estimated from a regression of temperature on a linear time trend separately for each subdistrict. This substantial heterogeneity in heat shocks over time and space will form the basis of our identifying variation. Appendix Figure A3 plots histograms of temperature anomalies at the village level for each census year. There is substantial cross-sectional variation across villages in census-year anomalies, which drive identification in models without village fixed effects. As the census round becomes more recent, there is also a rightward shift in the distribution of annual anomalies, consistent with moderate warming, though this shift is far from universal. 50 In models with village fixed effects, differential within-village shifts in anomalies over time will be the primary identifying variation. The key statistics of the temperature distribution are in Panel D of Appendix Table A1. Figure A3: Histograms of village-level temperature anomalies Maximum, 1990 Maximum, 1998 Maximum, 2005 Maximum, 2013 .08 .06 .08 .06 .06 .06 Fraction Fraction Fraction Fraction .04 .04 .04 .04 .02 .02 .02 .02 0 0 0 0 -1 -.5 0 .5 -1.5 -1 -.5 0 .5 -1 -.5 0 .5 1 -1 -.5 0 .5 1 Anomaly, degrees celsius Anomaly, degrees celsius Anomaly, degrees celsius Anomaly, degrees celsius Mean, 1990 Mean, 1998 Mean, 2005 Mean, 2013 .08 .08 .1 .1 .06 .06 .08 .08 Fraction Fraction Fraction Fraction .06 .06 .04 .04 .04 .04 .02 .02 .02 .02 0 0 0 0 -.8 -.6 -.4 -.2 0 .2 -2 -1.5 -1 -.5 0 .5 -.5 0 .5 1 -1 -.5 0 .5 Anomaly, degrees celsius Anomaly, degrees celsius Anomaly, degrees celsius Anomaly, degrees celsius Minimum, 1990 Minimum, 1998 Minimum, 2005 Minimum, 2013 .08 .15 .15 .1 Fraction Fraction Fraction Fraction .06 .1 .1 .04 .05 .05 .05 .02 0 0 0 0 -1 -.5 0 .5 -3 -2 -1 0 1 -.5 0 .5 1 1.5 -1 -.5 0 .5 Anomaly, degrees celsius Anomaly, degrees celsius Anomaly, degrees celsius Anomaly, degrees celsius Note: Figure shows histograms of estimated annual village-level surface temperature anomalies in degrees celsius relative to the 1980-2021 temperature average, for each round of the Economic Census of India (1990, 1998, 2005, and 2013). Anomalies are calculated at the village level as the deviation between the average annual temperature in the two years before the census year and the long-run average annual temperature from 1980-2021. Sample is all villages in India for which Economic Census data is available. C Robustness tests C.1 Normalized quantile effects The estimates in Figure 2 are not adjusted for firm size. In Figure A4 we present normalized quantile effects. We normalize the quantile coefficients by dividing by the average firm size in each quantile. The axis now represents percent changes in firm size. The downward slope 51 of the quantile curve becomes even more pronounced, particularly in the SHRID and district- year fixed effects specification Figure A4: Quantile regressions: normalized changes Effect on firm size (%) Effect on firm size (%) 20 20 10 10 0 0 -10 -10 -20 -20 -30 -30 -40 -40 5 25 45 65 85 5 25 45 65 85 Quantile Quantile District FE District-Year FE Effect on firm size (%) Effect on firm size (%) 20 20 10 10 0 0 -10 -10 -20 -20 -30 -30 -40 -40 5 25 45 65 85 5 25 45 65 85 Quantile Quantile SHRID FE SHRID and District-Year FE Estimate 95% CI Note: Figure shows estimates and 95% confidence intervals of β τ from quantile regressions of firm size on temperature anomalies and geographic fixed effects. The estimates are normalized by quan- tiles of firm size to obtain percent changes. Standard errors clustered at the village level. Anomalies are calculated at the village level as the deviation between the average annual temperature in the two years before the census year and the long-run average annual temperature from 1980-2021. Sample is the unit level data of the last four Economic Censuses of India (1990, 1998, 2005, and 2013). C.2 Sector effects Section 4.3 presents an analysis of sector-specific reallocation effects. Here, we expand on specific robustness tests. Appendix Table A2 estimates the main firm-level linear regression, 52 splitting the sample into large macro-sectors of services and manufacturing. For services, in Panel B, the estimated effect again remains significant in three out of five specifications, with magnitudes similar to those in Table 1. Instead, Panel A reveals that the effect sizes are substantially larger in absolute magnitudes and more significant in the manufacturing sector. Combining these estimates with the subsample means in Appendix Table A1 reveals that in our preferred specification (column 5), a 1 degree increase in maximum temperature reduces firm size by 23.5% in manufacturing and 9.7% in services. While the effect obtains in both sectors, it is larger in both absolute and proportional terms in the manufacturing sector. It is therefore possible that the quantile effects are driven in part by larger negative responses among larger manufacturing firms relative to smaller services firms. Table A2: Heat shocks and firm size by macro-sector Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Panel A: Manufacturing Max temp anomaly -0.456** -0.294*** -0.234 -0.296*** -0.671*** (0.171) (0.046) (0.166) (0.051) (0.157) Observations 21475981 21475973 21475971 21428980 21428977 Panel B: Services Max temp anomaly -0.040 -0.126*** -0.067 -0.109*** -0.209* (0.069) (0.030) (0.105) (0.028) (0.093) Observations 87499077 87499074 87499074 87480090 87480090 District FE No Yes No No No District × Year FE No No Yes No Yes SHRID FE No No No Yes Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Outcome is the total number of employees of the firm. Anomalies are calculated as the difference between the average annual maximum temperature of t and t − 1 and the long-run average temperature from 1980-2021. Fixed effect specification given in table footer. *** p < 0.01, ** p < 0.05, * p < 0.1. The most rigorous test of whether sector effects are driving the firm size response to heat shocks is to include disaggregated sector fixed effects in the main model (Appendix Table A3). Panel A reveals that, after conditioning on 90 3-digit subsectors, the estimated average firm size effects hardly change, remaining significant at 1% in the same specifications. Panel 53 B interacts the sector fixed effects with year indicators to account for not just sector-specific effects, but also potential sectoral trends in firm size and exposure to climate shocks. The results in columns (1)-(4) seem to reveal that the average firm size effect vanishes conditional on sector-specific trends. However, this is because these specifications do not control suffi- ciently for heterogeneous spatial trends. Instead, in our preferred specification of column (5), the estimate is nearly identical to that of Table 1 1. Table A3: Heat shocks and firm size, sectoral fixed effects Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Panel A: Sector fixed effects Max temp anomaly -0.101 -0.142*** -0.130 -0.131*** -0.293*** (0.070) (0.032) (0.093) (0.033) (0.085) Panel B: Sector-by-year fixed effects Max temp anomaly 0.033 0.009 -0.107 0.018 -0.279** (0.096) (0.038) (0.093) (0.041) (0.085) District FE No Yes No No No District × Year FE No No Yes No Yes SHRID FE No No No Yes Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Outcome is the total number of employees of the firm. Anomalies are calculated as the difference between the average annual maximum temperature of t and t − 1 and the long-run average temperature from 1980-2021. Fixed effect specification given in table footer. Sector fixed effects refer to 90 NIC87 3-digit subsectors of economic activity. *** p < 0.01, ** p < 0.05, * p < 0.1. Lastly, Appendix Figure A5 plots quantile curves conditional on sector fixed effects. The results are nearly identical to, and if anything somewhat more pronounced than, those of Fig- ure 2, implying nearly a 2-employee reduction in firm size at the right tail of the distribution. In sum, we find evidence of some between-sector reallocation, as well as quantitatively mean- ingful heterogeneity in effects across manufacturing and services. It is indeed highly likely that sectors are not equally exposed to, or equally able to adapt to, increasing heat shocks. However, the results of the sectoral fixed effects specification suggest that these sectoral het- erogeneities do not appear to be driving effects across the firm size distribution. Instead, it 54 appears that large firms contract employment most regardless of the sector in which they operate. Figure A5: Quantile regressions with sector fixed effects Effect on firm size 1 Effect on firm size 1 .5 .5 0 0 -.5 -.5 -1 -1 -1.5 -1.5 -2 -2 -2.5 -2.5 5 25 45 65 85 5 25 45 65 85 Quantile Quantile District FE District-Year FE Effect on firm size 1 Effect on firm size 1 .5 .5 0 0 -.5 -.5 -1 -1 -1.5 -1.5 -2 -2 -2.5 -2.5 5 25 45 65 85 5 25 45 65 85 Quantile Quantile SHRID FE SHRID and District-Year FE Estimate 95% CI Note: Figure shows estimates and 95% confidence intervals of β τ from quantile regressions of firm size on temperature anomalies and geographic fixed effects. Standard errors clustered at the village level. Anomalies are calculated at the village level as the deviation between the average annual temperature in the two years before the census year and the long-run average annual temperature from 1980-2021. Sample is the unit level data of the last four Economic Censuses of India (1990, 1998, 2005, and 2013). C.3 Long differences Section 6 estimates the long-run impacts of rising temperature on the distribution of firm size. Here, we present several auxiliary results to Table 4. First, in Figure A6, we present binned scatterplots of the relationship between long-run change in firm size and temperature 55 across 8, 15, and 23-year periods. The top panel shows the unconditional long-difference relationship, while the bottom panel shows the relationship after controlling for state-by- period fixed effects. Overall, the results reveal robust negative relationships, which tend to be weakest for the 15-year period but relatively consistent overall. The relationships are not meaningfully sensitive to the inclusion of fixed effects. Figure A6: Heat and firm size: long differences .2 .2 .2 Change in firm size Change in firm size Change in firm size .1 .1 .1 0 0 0 -.1 -.1 -.1 -.2 -.2 -.2 -1 -.5 0 .5 1 1.5 -.5 0 .5 1 -1 -.5 0 .5 1 Change in temperature (C) Change in temperature (C) Change in temperature (C) 8-year difference, no FE 15-year difference, no FE 23-year difference, no FE .2 .2 .2 Change in firm size Change in firm size Change in firm size .1 .1 .1 0 0 0 -.1 -.1 -.1 -.2 -.2 -.2 -.5 0 .5 -.2 0 .2 .4 .6 .8 -.5 0 .5 Change in temperature (C) Change in temperature (C) Change in temperature (C) 8-year difference, state-by-period FE 15-year difference, state-by-period FE 23-year difference, state-by-period FE Note: Figure shows binned scatterplots of the relationship between the change in firm size and change in temperature over periods of 8 (1990-1998, 1998-2005, 2005-2013), 15 (1990-2005, 1998- 2013), and 23 (1990-2013) years. Temperature at each endpoint of a long difference period is calcu- lated as a 3-year lagged average. Bottom panel includes state-by-period fixed effects, where period refers to the long difference period. Next, we test the robustness of measurement choices in Table 4 by using the long-log- differences in Table A4. In this case, the outcome variable is the log-difference of firm size instead of its level, so the coefficients can be interpreted as percent changes for small changes. Overall, the results are highly consistent with the main estimates. 56 Table A4: Heat shocks and firm size: long log difference estimates Dependent variable Change in firm size (employment) Difference 23-year 15-year 8-year (1) (2) (3) (4) (5) (6) Temperature change (C) -0.075*** -0.060*** -0.010*** -0.027*** -0.046*** -0.026*** (0.003) (0.005) (0.002) (0.003) (0.001) (0.003) State × Period FE No Yes No Yes No Yes Observations 331998 331998 697146 697146 1097327 1097327 R2 0.002 0.015 0.000 0.026 0.003 0.024 Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013), collapsed to the village level. Outcome is the log change in average number of employees of private firms in a village over the long difference period t, k. Temperature shocks are calculated as the change in the 2-year lagged average annual maximum temperature over the long difference period t, k. Fixed effect specification given in table footer. *** p < 0.01, ** p < 0.05, * p < 0.1. C.4 Heat shock measurement Our results may be sensitive to specification choices for the heat shock. We consider ro- bustness to several different measurements of heat. First, it is possible that the results are sensitive to use of the average daily maximum temperature for the calculation of anomalies. In Table A5, we re-estimate the main models of Table 1, calculating the temperature anomaly as the deviation of the average daily mean or minimum temperature from its long-run value. We find similar results when defining the temperature anomaly using average or minimum daily temperatures, though they are somewhat less pronounced for the minimum tempera- ture specification. Throughout, we use the average of temperature anomalies in t and t − 1 as our main independent variable, to reduce the influence of year-to-year noise in the temperature series. We test whether the results are sensitive to the choice of lag period. In Table A6, we calculate the anomaly over lags of one and three years instead of two. We find that the effect holds for three-year lags but not for only the census-year (that is, the one-year lag), consistent with increased noise from year-to-year temperature fluctuation and limited response time for firms when the shock is measured in the contemporaneous year. Finally, we consider an entirely different specification of temperature that does not use anomalies. Rather, we implement a model following Burgess et al. (2017) and Caggese et al. (2024), among many others, in which temperature enters semi-parametrically as the number 57 Table A5: Heat shocks and firm size, robustness to temperature measurement Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Panel A: Mean temperature Mean temp anomaly (2-year lag) -0.151 -0.237*** -0.003 -0.207*** -0.301** (0.115) (0.042) (0.130) (0.029) (0.109) Panel B: Min temperature Min temp anomaly (2-year lag) -0.167 -0.250** 0.043 -0.177*** -0.252* (0.138) (0.081) (0.167) (0.041) (0.113) District FE No Yes No No No District × Year FE No No Yes No Yes SHRID FE No No No Yes Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Outcome is the total number of employees of the firm. Anomalies are calculated as the difference between the average annual maximum temperature of the given lag period and the long-run average temperature from 1980-2021. Fixed effect specification given in table footer. *** p < 0.01, ** p < 0.05, * p < 0.1. Table A6: Heat shocks and firm size, robustness to lag Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Panel A: Current year Max temp anomaly (census year) -0.007 -0.048 0.089 -0.025 -0.065 (0.048) (0.034) (0.083) (0.035) (0.053) Panel B: 3-year lag Max temp anomaly (3-year lag) -0.136 -0.165** -0.070 -0.182*** -0.246** (0.183) (0.051) (0.149) (0.024) (0.090) District FE No Yes No No No District × Year FE No No Yes No Yes SHRID FE No No No Yes Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Outcome is the total number of employees of the firm. Anomalies are calculated as the difference between the average annual maximum temperature of the given lag period and the long-run average temperature from 1980-2021. Fixed effect specification given in table footer. *** p < 0.01, ** p < 0.05, * p < 0.1. 58 of days in specific temperature interval. Specifically, for firm i in village v in census year t, we estimate the model yivt = α + ∑ β j days jvt + δv + δt + δvt + ϵivt j Where j are bins of the distribution of temperatures, and days jvt indicates the number of days that village v experienced in temperature bin j in round t. As before, we average this number over t and t − 1. We set bins in intervals of 10C from 0 to 40, with endpoints at < 0 and ≥ 40, and average the number of days in each temperature bin in the census year and t − 1.24 The estimates in Table A7 reveal that firm size remains a downward sloping function of temperature, consistent with the main results. The largest and most significant effects on firm size occur in the ≥ 40 degree bin. These results are robust to most fixed effect specifications and the use of an exponential Poisson pseudo maximum likelihood (PPML) model to account for the right-skewness and count nature of the firm size outcome (Wooldridge, 1999). Finally, it is possible that our linear specification of temperature anomalies is unable to capture more complex relationships between firm outcomes and heat. This has important economic implications, as Barreca et al. (2015) argue that differential impacts of heat shocks between initially hotter vs. colder places can be interpreted as evidence of long-run adapta- tion. We therefore test whether effects of temperature vary by initial temperature. We interact the temperature anomaly in our main specification with an indicator variable equaling one if the long-term average daily maximum temperature is 30C or greater. Table A8 finds mixed evidence of systematic differences in firm size effects for hotter and cooler places. In some specifications (columns 2 and 3), hotter locations have significantly smaller firm size reduc- tions in response to heat, consistent with adaptation. However, in the most rigorous specifi- cation (column 5), these differences shrink to zero, suggesting limited long-run adaptation on average. C.5 Model specification We consider robustness to two key model specification choices—the use of controls, and the choice of estimation method. Table A9 tests the robustness of the main result to several specification choices. In Panel A, we control for precipitation anomalies, constructed iden- 24 All estimates omit the bin between 10 and 20C. 59 Table A7: Firm size and heat: number of days Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Panel A: Linear regression Average number of days Temp < 0 0.000279 0.00433 0.00217** 0.00246*** 0.0142* (0.00145) (0.00304) (0.000947) (0.000820) (0.00768) Average number of days 0 ≤ Temp < 10 -0.0113* 0.00345 0.000683 0.000905 0.00689 (0.00588) (0.00243) (0.00267) (0.00210) (0.00446) Average number of days 10 ≤ Temp < 20 Average number of days 20 ≤ Temp < 30 -0.00233 -0.00442 0.00146 0.00168 -0.00278 (0.00257) (0.00403) (0.00158) (0.00121) (0.00477) Average number of days 30 ≤ Temp < 40 -0.00113 -0.00552 0.00134 0.00176 -0.00349 (0.00221) (0.00399) (0.00143) (0.00110) (0.00487) Average number of days Temp ≥ 40 -0.00976*** -0.0153*** -0.00356** 0.00242 -0.00833 (0.00169) (0.00264) (0.00151) (0.00190) (0.00554) Panel B: Poisson regression Average number of days Temp < 0 0.000228 0.00287* 0.00120** 0.00139*** 0.00878** (0.000636) (0.00165) (0.000499) (0.000447) (0.00370) Average number of days 0 ≤ Temp < 10 -0.00550** 0.00230* 0.0000734 0.000125 0.00442** (0.00258) (0.00135) (0.00128) (0.00103) (0.00223) Average number of days 10 ≤ Temp < 20 Average number of days 20 ≤ Temp < 30 -0.00104 -0.00179 0.000664 0.000771 -0.00101 (0.00109) (0.00203) (0.000724) (0.000555) (0.00189) Average number of days 30 ≤ Temp < 40 -0.000526 -0.00223 0.000624 0.000804 -0.00135 (0.000933) (0.00202) (0.000661) (0.000512) (0.00194) Average number of days Temp ≥ 40 -0.00432*** -0.00679*** -0.00164** 0.00111 -0.00367* (0.000718) (0.00139) (0.000687) (0.000865) (0.00221) Observations 110537617 110523664 110537617 110537617 110523664 District FE No No Yes No No District × Year FE No No No Yes Yes SHRID FE No Yes No No Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Temperature bins are calculated as the number of days in which the maximum temperature fell within the indicate range, averaged between the census year and t − 1. *** p < 0.01, ** p < 0.05, * p < 0.1. tically to our temperature anomalies. The results are essentially unchanged. Next, Panel B estimates an exponential model with a fixed effects Poisson regression estimator, to account for outliers given the skewness of the firm size distribution in India, with a large mass at 1, as well as the discrete count nature of the outcome. Estimates from the Poisson regression, 60 Table A8: Heat shocks and firm size: heterogeneity by long-run temperature Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Max temp anomaly -0.110 -0.275*** -0.295*** -0.018 -0.205* (0.126) (0.035) (0.061) (0.114) (0.097) Hot location 0.072 0.000 0.003 -0.026 0.000 (0.094) . (0.054) (0.063) . Max temp anomaly × Hot location 0.012 0.128** 0.148* -0.048 -0.004 (0.154) (0.046) (0.065) 0.073 0.066 Observations 1.105e+08 1.105e+08 1.105e+08 1.105e+08 1.105e+08 District FE No No Yes No No District × Year FE No No No Yes Yes SHRID FE No Yes No No Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Anomalies are calculated as the difference between the average annual maximum temperature of the given lag period and the long-run average temperature from 1980-2021. A hot location is defined as a village whose average annual maximum temperature exceeds 30C. *** p < 0.01, ** p < 0.05, * p < 0.1. interpreted as percent changes, are similar in relative magnitude to the main estimates, and again significant in the same three out of five specifications. C.6 Urban heat islands Our results may also be driven by the urban heat island effect. Because of the UHI, ur- ban areas tend to be hotter, and large firms cluster in urban areas. In a naive specification, this would lead us to underestimate the negative effect of heat; however, our use of temper- ature anomalies and fixed effect strategy address this concern. A more subtle issue is that a measured 1°C anomaly may feel hotter in urban areas, which may not be captured in the relatively coarse temperature data, which can underestimate temperature extremes. This cor- related measurement error could generate the illusion of larger negative effects for large firms, when in fact these large urban firms simply experience larger unmeasured heat shocks. We test this hypothesis by splitting the sample into urban and rural SHRIDs and re-running both the fixed effects (Table A10) and quantile regressions (Figure A7) on each subsample. The main results hold in both subsamples. 61 Table A9: Heat shocks and firm size, robustness to specification Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Panel A: Controlling for precipitation Max temp anomaly (2-year lag) -0.064 -0.178*** -0.093 -0.176*** -0.278** (0.099) (0.049) (0.093) (0.042) (0.085) Panel B: Poisson regression Max temp anomaly (2-year lag) -0.047 -0.072*** -0.026 -0.068*** -0.097** (0.034) (0.014) (0.042) (0.014) (0.031) District FE No Yes No No No District × Year FE No No Yes No Yes SHRID FE No No No Yes Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Outcome is the total number of employees of the firm. Anomalies are calculated as the difference between the average annual maximum temperature of the given lag period and the long-run average temperature from 1980- 2021. Panel A includes a control for the 2-year lag annual precipitation anomaly- the difference between the average annual precipitation of years t and t − 1 and the long-run average annual precipitation for 1980-2021. Fixed effect specification given in table footer. *** p < 0.01, ** p < 0.05, * p < 0.1. 62 Figure A7: Heat and firm size: quantile regressions by urban location 2 2 1 1 Effect on firm size Effect on firm size 0 0 -1 -1 -2 -2 -3 -3 -4 -4 -5 -5 5 25 45 65 85 5 25 45 65 85 Quantile Quantile Urban Rural Estimate 95% CI Note: Figure shows estimates and 95% confidence intervals of β τ from quantile regressions of firm size on temperature anomalies and geographic fixed effects for subsamples of urban and rural vil- lages. Standard errors clustered at the village level. Anomalies are calculated at the village level as the deviation between the average annual temperature in the two years before the census year and the long-run average annual temperature from 1980-2021. Sample is the unit level data of the last four Economic Censuses of India (1990, 1998, 2005, and 2013). 63 Table A10: Heat shocks and firm size: urban and rural Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Panel A: Urban Max temp anomaly -0.192* -0.191*** -0.174*** -0.138 -0.234* (0.094) (0.039) (0.036) (0.118) (0.117) Observations 73605742 73602707 73605742 73605741 73602706 Panel B: Rural Max temp anomaly 0.030 -0.107*** -0.164*** 0.165 -0.181* (0.031) (0.011) (0.041) (0.177) (0.079) Observations 36931875 36920957 36931875 36931874 36920956 District FE No No Yes No No District × Year FE No No No Yes Yes SHRID FE No Yes No No Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Anomalies are calculated as the difference between the average annual maximum temperature and the long-run average temperature from 1980-2021, averaged between t and t − 1. Relevant sample of villages given in panel headers. *** p < 0.01, ** p < 0.05, * p < 0.1. 64 D Estimating the productivity impact of labor reallocation We estimate the ARPL gap between formal and informal firms through the following OLS regression: ARPLi = α + β In f ormali + Xi + ϵi (8) where ARPL is value added per worker and In f ormal is a dummy for firms with less than 8 workers. The baseline version does not include any controls; thus, the constant term of the baseline specification measures the mean ARPL in formal firms. The robustness variants of this specification include Xi controls, such as industry and state fixed effects. The regressions are weighted using the inverse sampling probabilities included in the ASI and NSS datasets. Table A11 shows the estimated labor productivity differentials between firm types, in USD. Each column contains an estimate for the informal-formal differential as well as the average in the formal sector. Column (1) shows the raw differential, while additional spec- ifications in columns (2) and (3) also adjust for industry and state fixed effects to adjust for unobserved differences in workers’ human capital across firms. The underlying concern is that part of the gap in observed labor productivity between formal and informal firms could be due to a higher average human capital level of workers in formal firms; this component should be excluded when assessing the change in aggregate productivity from worker real- location. Therefore, we estimate gaps in productivity within narrow 5-digit industry groups and state. As expected, using within-industry variation industry reduces the estimated labor pro- ductivity gap between formal and informal firms. The state fixed effects do not make a sub- stantial additional difference. However, these adjustments will be overly conservative, since a substantial share of the cross-sector variance in labor productivity will be due to firm rather than worker effects. Columns (4) and (5) present two additional estimates of the ARPL differential. It has been estimated by de Mel et al. (2009) that micro-enterprises in low- and middle-income countries may under-report revenue by about 30 percent. As this would tend to bias the productivity gap upwards, we present an estimate where the reported value added of informal firms has been adjusted for 30 percent under-reporting. Lastly, we present the ARPL gap estimated by using 10 workers as the threshold for categorizing firms as formal. Finally, our baseline 8- 65 Table A11: Productivity gaps between formal and informal sectors: NSS-ASI Survey Esti- mates Outcome variable Value-added per worker (1) (2) (3) (4) (5) Informal -2172.315*** -1576.889*** -1494.859*** -1594.732*** -5283.343*** (111.472) (108.057) (107.113) (112.280) (257.116) Formal average 3520.006*** 6636.827*** (110.691) (256.779) Sector FE No Yes Yes No No State FE No No Yes No No Observations 134480 134449 134449 134480 134480 R2 0.005 0.139 0.157 0.002 0.007 Note: Estimates based on combined National Sample Survey of Unincorporated Non-agricultural Enterprises (2015-16) and the Annual Survey of Industries (2015) establishment-level dataset. Value added per worker is expressed in 2015 USD terms. The employment size cutoff used to define the formal and informal sectors is 8 in columns 1-4 and 10 in column 5. Column 4 uses an alternative version of the outcome variable based on adjusting the reported value added of informal firms for 30 percent under-reporting. *** p < 0.01, ** p < 0.05, * p < 0.1. worker cutoff—which follows the Economic Census’s categorization scheme—is a relatively inclusive definition of formal firms in the Indian context. For example, 10 workers is a com- monly used cutoff for formal manufacturing firms (Kanbur, 2017). By including borderline informal firms in the formal segment, the 8-worker cutoff may lead to an underestimation of the latter’s productivity advantage. Indeed, the ARPL gap in column (5) is the largest of all specifications. Finally, Table A12 shows the calculation details that translate productivity differentials and the formal-informal reallocation effect into changes in aggregate ARPL, following the formula in equation (5). The first section of the table reprints all the estimates of Table A11, which come from the NSS-ASI data. Instead, the second section takes the estimates from Table 2, column (5), as well as estimates of the share of formal and informal firms, both from the Economic Census. 66 Table A12: Labor reallocation and aggregate productivity change accounting Industry Underreporting- 10-worker Raw Industry FE & State FE adjusted cutoff (1) (2) (3) (4) (5) NSS-ASI Informal ARPL 1347.69 1943.12 2025.15 1925.27 1353.48 Formal ARPL 3520.01 3520.01 3520.01 3520.01 6636.83 ARPL ratio 0.38 0.55 0.58 0.55 0.20 Economic Census Baseline informal share (%) 77 77 77 77 77 Baseline formal share (%) 23 23 23 23 23 ∆ informal employment share (%) 2.10 2.10 2.10 2.10 2.10 ∆ ARPL -2.47 -1.44 -1.33 -1.46 -4.32 Note: Estimates based on the combined National Sample Survey of Unincorporated Non-agricultural Enterprises (2015-16) and Annual Survey of Indus- tries (2015) dataset (NSS-ASI) and the 6th Economic Census of India. ARPL is value added per worker, expressed in USD terms using the 2015 exchange rate. The ARPL ratio is the ratio of value added per worker in informal and formal establishments. The "raw" or baseline ARPL ratio is estimated by regressing value added per worker on a dummy for informality of the establishment. The “industry FE” and “industry + state FE” ARPL ratios are based on including 5-digit industry and state fixed effects in the baseline regression. The “underreporting-adjusted” estimate is based on correcting value added per worker in informal establishments for an assumed under-reporting by 30%. The “10-worker cutoff” estimate is based on defining the formal sector using a minimum size of 10 workers, as compared to a baseline cutoff of 8 workers. The baseline shares of formal and informal firms in total non-agricultural employment in India are estimated using the 2013 Economic Census data. The change in these shares due to a 1 degree heat anomaly increase is based on the regression estimate reported in column 5 of Table 2. E Alternative explanations E.1 Firm exit A plausible explanation for the short-run results in Section 4.2 is that small and large respond identically to heat, but small firms exit the market while large firms contract but remain active. This explanation is mechanical, rather than behavioral given the large density of one-person firms in the data. To test this theory, we estimate subdistrict-level regressions over firm size bins similar to equation (2). However, instead of employment shares in each firm size bin as the outcomes, we use the total number of firms. The logic behind this test is that if our results are driven by small firm exit, we should expect that the number of small firms reduces substantially in response to heat shocks. Table A13 contains the results. We estimate the model using two different methods— fist, a log-linear regression (Panel A), and second, an exponential Poisson pseudo-maximum likelihood (PPML) model to account for the highly right-skewed count outcome (Panel B). In Panel A, there is a small but significant 4% reduction in the number of firms, but this is accompanied by an even larger 6-7% reduction in the number of large firms, as well as an increase in the number of mid-sized firms. Panel B reconciles these conflicting results by re-estimating the regressions with a PPML model, using an exponential functional form to 67 Table A13: Heat shocks and firm counts, subdistrict level Dependent variable Number of firms Firm size 1 2-7 8 and above (1) (2) (3) (4) (5) (6) Panel A: Log-linear regression Max temp anomaly -0.036** -0.040** 0.066*** 0.059*** -0.059*** -0.073*** (0.018) (0.017) (0.015) (0.014) (0.021) (0.020) District FE Yes No Yes No Yes No Year FE Yes Yes Yes Yes Yes Yes Subdistrict FE No Yes No Yes No Yes Observations 22072 22072 22032 22032 21176 21176 Panel B: Poisson regression Max temp anomaly -0.013 -0.025* 0.031 0.013 0.015 -0.015 (0.017) (0.015) (0.024) (0.020) (0.036) (0.035) District FE Yes No Yes No Yes No Year FE Yes Yes Yes Yes Yes Yes Subdistrict FE No Yes No Yes No Yes Observations 22095 22093 22095 22093 22095 22095 Note: Standard errors in parentheses clustered at the subdistrict level. Sample is all private sector firms surveyed in four Economic Census rounds (1990, 1998, 2005, and 2013). Outcome is the number of firms of a given size in the subdistrict, indicated in table header. Anomalies are calculated as the difference between the average annual maximum temperature and the long-run average temperature from 1980- 2021, averaged between t and t − 1. Panel B uses an exponential Poisson pseudo maximum likelihood (PPML) estimator. *** p < 0.01, ** p < 0.05, * p < 0.1. account for the highly right-skewed count outcome. This model yields small and statistically insignificant effects for all firm size groups. The PPML approach is likely to yield a more reliable estimator of the model if the true relationship is exponential (Wooldridge, 1999). E.2 Migration It is well-established in the literature that rising temperatures spur migration (Mueller et al., 2014; Baez et al., 2017; Lang et al., 2025). Out-migration of working-age men might force firms to contract due to labor shortage. This is a labor supply explanation of contraction, 68 rather than the labor demand story we favor.25 We test this explanation by estimating the relationship between temperature anomalies and migration in data from the Census of India. Specifically, we take district-level migration census data for 2001 and 2011, where migrant origin districts are collected. The data are presented as origin district-level counts of migrants, who are categorized based on the recency of their migration in bins of 0-5 years and 6-11 years. We then calculate the average temperature anomaly for the origin district in the period in question, yielding a pseudo panel of four periods—two census rounds ten years apart, each with two five-year periods. Finally, we regress the log of total district migrants in a given period on the average temperature anomaly for that period. Figure A8: Heat and migration 4 2 Log migration 0 -2 -4 -.8 -.4 0 .4 Temperature anomaly (C) Note: Figure shows residualized scatterplot relationship between average temperature anomaly and migration at the district-period level. Sample is 640 migrant origin districts from the Census of India, observed over four 5-year periods in a pseudo panel from 1991-2011. Migration is defined as the log of the number of migrants who left the origin district for work over a given 5-year period. Temperature anomaly defined as the average difference between the annual max temperature and its long-run mean over the same period. Both variables are residualized of district and period fixed effects. The results are in Figure A8, which shows a residualized scatterplot after controlling for 25 Although presumably declining labor demand and wages would serve as a cause of climate-related migration. 69 district and period fixed effects. The results reveal a precisely estimated zero relationship between heat shocks and migration levels. This zero relationship may appear inconsistent with the existing literature that identifies migration as a common response to heat (Rexer and Sharma, 2024). However, recall that our regression is identified based on short-run shocks, while climate-induced migration is likely to be long run phenomenon. This perspective is consistent with other results throughout the paper that show limited adaptation in the short run. Overall, the results suggest that migration is unlikely to be the mechanism by which large, formal firms shed labor. E.3 NREGA Another labor supply explanation for our results is that the Indian state targets welfare programs to villages most affected by heat, increasing the opportunity cost of private sector work. One such program in India is NREGA, an employment guarantee workfare scheme. Garg et al. (2020) show that NREGA participation increases in heat-affected areas, potentially explaining our results. We address this potential confounder in two ways. First, we use district-level data to replicate the results of Garg et al. (2020) on the relationship between our temperature anomalies and uptake of NREGA employment at the district level. Second, we estimate a version of our main specification in which we interact the temperature anomaly with a measure of NREGA uptake at the district level to show that firm size effects are not larger for areas with NREGA. Table A14: Heat shocks and NREGA uptake Dependent variable Log households Log person-days Log labor exp Log material exp (1) (2) (3) (4) Max temp anomaly 0.115 0.040 0.101** 0.174** (0.099) (0.041) (0.046) (0.087) District FE Yes Yes Yes Yes Year FE Yes Yes Yes Yes Observations 3511 3511 3511 3511 R2 0.631 0.946 0.785 0.641 Note: Standard errors in parentheses clustered at the district level. *** p < 0.01, ** p < 0.05, * p < 0.1. Table A14 estimates the impact of heat on NREGA uptake. To do so, we construct a 70 district-level panel from 2006-2014. We then regress various NREGA take-up outcomes on a population-weighted district-level average of our maximum temperature anomaly, control- ling for district and year fixed effects. The results confirm the findings of Garg et al. (2020), with evidence of meaningful increases in NREGA activity—particularly labor and material expenditure—in response to heat shocks. Table A15: Heat shocks and firm size: NREGA uptake Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Max temp anomaly -2.407 0.355 0.284 0.774 1.231 (1.419) (2.253) (1.960) (0.771) (0.860 ) NREGA 3 year average -0.028 -0.014 -0.019 -0.110 0.024 (0.043 ) ( 0.016 ) ( 0.014 ) (0.057 ) ( 0.053 ) Max temp anomaly × NREGA 3 year average 0.162 -0.032 -0.032 -0.077 -0.071 (0.101) (0.156 ) ( 0.134 ) (0.056 ) (0.062 ) Observations 44665890 44641946 44665887 44665885 44641944 District FE No No Yes No No District × Year FE No No No Yes Yes SHRID FE No Yes No No Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. *** p < 0.01, ** p < 0.05, * p < 0.1. In Table A15, we interact the district-level log of NREGA person-days with the temper- ature anomaly in our main regression specification. Since the data do not perfectly overlap with the Economic Census, we use only the latter two rounds and assign the NREGA value of 2006 for the 2005 round and the average of 2012-2014 for the 2013 round. The interaction between NREGA participation and the temperature anomaly is generally small and insignif- icant. There is no evidence that villages with higher NREGA enrollment observe larger firm size effects, suggesting that labor supply effect does not drive our results. An important caveat is the limited overlap between the Economic Census and the advent of NREGA, which began in 2006, severely limiting the sample for Table A15. That being said, the lack of NREGA in roughly half our sample should imply, prima facie, that this is unlikely to drive the results. E.4 Demand-side effects One concern is that output market demand drive the quantile effects of heat shocks, rather than labor demand effects. This would be the case if, for example, demand for the products made by large firms is differentially sensitive to heat. Caggese et al. (2024) find that demand 71 effects only explain a small amount of the total firm-level effect of heat for Italian firms. Fol- lowing the approach in Caggese et al. (2024), we test for demand effects by leveraging dif- ferences across sectors in tradeability, based on the logic that traded sectors are less sensitive to local demand shocks. Therefore, if demand effects are driving the contraction, rather than labor productivity effects, we should expect to see larger effects in the non-traded sectors. We test this hypothesis by interacting the heat anomaly with the log of sectoral export expo- sure. Export exposure is measured at the sector level as the share of total domestic sectoral output that is exported, averaged from 2000-2013, accounting for intermediate imported in- puts. Data on sectoral exports and output comes from the regional ADB input-output tables is transformed into 3-digit NIC08 industry codes using available crosswalks. Table A16: Heat shocks and firm size: heterogeneity by sectoral export exposure Dependent variable Firm size (employment) (1) (2) (3) (4) (5) Max temp anomaly -0.079 -0.120 -0.133* -0.018 -0.156 (0.048) (0.066) (0.066) (0.098) (0.083) Log export exposure -0.134*** -0.137*** -0.135*** -0.137*** -0.138*** (0.008) (0.008) (0.008) (0.008) (0.007) Max temp anomaly × Log export exposure -0.020 -0.029 -0.023 -0.025 -0.029 (0.024) (0.026) (0.027) (0.024) (0.024) Observations 110537617 110523664 110537617 110537617 110523664 District FE No No Yes No No District × Year FE No No No Yes Yes SHRID FE No Yes No No Yes Note: Standard errors in parentheses clustered at the SHRID (village) level. Sample is all private sector firms surveyed in four Eco- nomic Census rounds (1990, 1998, 2005, and 2013). Anomalies are calculated as the difference between the average annual maximum temperature and the long-run average temperature from 1980-2021, averaged between t and t − 1. Export exposure is calculated as the share of domestic production that is exported, defined at the sector-level. *** p < 0.01, ** p < 0.05, * p < 0.1. The results are in Table A16. The results show no evidence of differential firm size ef- fects based on tradeability, with interaction effects being uniformly negative and insignificant across all specifications. This suggests that the employment effects operate through dimin- ished labor productivity, consistent with Somanathan et al. (2021). E.5 Capital-labor substitution Another potential explanation for our quantile effects results is that large firms respond to heat shocks by switching to more capital-intensive production technologies. This might be an optimal response on their part if heat shocks lower labor productivity more than they 72 do capital productivity —as suggested in Somanathan et al. (2021)— and more heat-resilient, commercially viable capital-intensive technologies exist in the market. Smaller firms might be unable to do this to the same extent because of credit constraints or a lack of capital-intensive technologies suitable to their scale. We cannot explore this alternative explanation directly using Economic Census data, but present descriptive evidence that runs counter to it from a recent survey of firms in South Asia. Figure A9: Adaptation Responses by Firm Size 60 60 60 Changed business practises Upgraded building Hired workers 40 40 40 20 20 20 0 0 0 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 Employment (in logs) Employment (in logs) Employment (in logs) 60 60 60 Invested in cooling Protective capital Fired workers 40 40 40 20 20 20 0 0 0 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 Employment (in logs) Employment (in logs) Employment (in logs) % firms on x-axis Note: Based on data from the World Bank South Asia Climate Adaptation Survey (Lang et al., eds, 2025). The charts represent binned scatter plots of adaptation actions versus firm size (log employment). The x-axis depicts the adaptation action; specifically, the percentage of firms that report undertaking the specified action in the past five years to prepare for episodes of extreme heat. We use the World Bank South Asia Climate Adaptation Survey, a survey conducted by the World Bank in 2024-25 to understand how firms in South Asia are adapting to climate change (Lang et al., eds, 2025). The survey was implemented in selected districts of Bangladesh, Pakistan and three large states of India (Gujarat, Maharashtra and Tamil Nadu), and cov- ered registered non-farm enterprises.26 As part of a module on weather shock expectations 26 Details about survey content and sampling design are discussed in Chapter 4 of Lang et al., eds (2025). 73 and responses, the survey asked respondents (firm managers) if they expected their firm to be affected adversely by episodes of extreme heat in the next five years. Respondents who replied in the affirmative were then asked to select from a list those actions that they under- took in the past 5 years to protect against heat shocks. While the list of potential actions did not specifically include capital-labor switching, the responses suggest that large firms are no more predisposed than small firms to prepare for heat shocks by replacing workers with ma- chines. Figure A9 presents binned scatter plots of firm size versus the percent of firms that report having undertaken particular adaptation actions. Larger firms are more likely than smaller firms to have "upgraded buildings", "changed business practices", installed "protec- tive capital", and most clearly of all "invested in cooling" to protect themselves from heat. However, they are also more likely than smaller firms to have prepared for extreme heat by hiring workers. Firing workers appears to be a relatively uncommon response to expected heat shocks, and it does not vary with firm size. Note, however, that the relative lack of firing does not necessarily conflict with our overriding labor demand explanation, since the sur- vey questions are phrased in such a way as to elicit ex-ante adaptations rather than ex-post responses to heat, as we estimate. In short, these responses indicate that while larger firms are more likely to upgrade capital for heat adaptation, the upgrades do not involve labor replacement. 74