TThe economic impact of weather and climate
Richard S.J. Tol a,b,c,d,e,f,1, ∗ a Department of Economics, University of Sussex, Falmer, United Kingdom b Institute for Environmental Studies, Vrije Universiteit, Amsterdam, The Netherlands c Department of Spatial Economics, Vrije Universiteit, Amsterdam, The Netherlands d Tinbergen Institute, Amsterdam, The Netherlands e CESifo, Munich, Germany f Payne Institute for Earth Resources, Colorado School of Mines, Golden, CO, USA
Abstract
I propose a new conceptual framework to disentangle the impacts of weather and climate oneconomic activity and growth: A stochastic frontier model with climate in the productionfrontier and weather shocks as a source of inefficiency. I test it on a sample of 160 countriesover the period 1950-2014. Temperature and rainfall determine production possibilitiesin both rich and poor countries; positively in cold countries and negatively in hot ones.Weather anomalies reduce inefficiency in rich countries but increase inefficiency in poor andhot countries; and more so in countries with low weather variability. The climate effect islarger that the weather effect.
Keywords : climate change; weather shocks; economic growth;stochastic frontier analysis
JEL codes : D24; O44; O47; Q54
1. Introduction
Climate matters to the economy. Not in the way that classical thinkers such as Guan Zhong,Hippocrates or Ibn Khaldun, or modern thinkers such as Huntington (1915) or Diamond(1997) argue it does. Environmental determinism is inconsistent with the observations. ∗ Jubilee Building, BN1 9SL, UK
Email address: [email protected] (Richard S.J. Tol)
URL: (Richard S.J. Tol) Marco Letta expertly assisted with data collection and regressions. Peter Dolton, Jurgen Doornik, BillGreene, David Hendry, Andrew Martinez, Pierluigi Montalbano and Felix Pretis had excellent commentson earlier versions of this work. I also thank seminar participants at the EMCC-III conference, Universityof Sussex, UNU-MERIT University, the 2019 IAERE conference, Cambridge University, London School ofEconomics, Sapienza University of Rome, Shanghai Lixin University, Edinburgh University, University ofSouthern Denmark, Federal Reserve Bank of San Francisco, and CESifo.
Preprint submitted to Elsevier March 1, 2021 a r X i v : . [ ec on . GN ] F e b here are thriving economies in the desert, in the tropics, and in the polar circle. There isdestitution, too, in all these places. Climate is not destiny, but it does matter.The prevailing view among economists, with some exceptions (Bloom and Sachs, 1998,Sachs, 2003, Olsson and Hibbs, 2005, Barrios et al., 2010), is that climate does not matter foreconomic development, only institutions do (Easterly and Levine, 2003, Rodrik et al., 2004).Some argue that climate and geography partly shaped institutions in the past, but havebecome irrelevant since (Acemoglu et al., 2001, 2002, Alsan, 2015). Institutional determinismis just as inconsistent with the observations. The two halves of the Korean Peninsula andof the island of Hispaniola are powerful reminders of the importance of institutions, butclimate matters for agriculture (Mendelsohn et al., 1994, Schlenker et al., 2005), for energydemand (Mansur et al., 2008), for tourism (Lise and Tol, 2002), for transport (Koetse andRietveld, 2009), for labour productivity (Kjellstrom et al., 2009, Zander et al., 2015), andfor health (Sachs and Malaney, 2002)—and thus for the economy as a whole.Climate matters, but it has been an empirical challenge to demonstrate this using countrydata. Climate changes only slowly over time, its signal swamped by confounders, many ofwhich change more quickly than climate. Climate varies substantially over space, but so do agreat many other things that we know are important for development. The insignificance ofclimate variables in cross-country studies may be due to a lack of statistical power. Indeed,a climate association is significant in subnational income data (Nordhaus, 2006, Dell et al.,2009, Henderson et al., 2018, Kalkuhl and Wenz, 2020) and, as is shown below, in long panels.Because of the confounders, this association cannot be given a causal interpretation.Unlike the impact of climate , the impact of weather can be identified—or so people haveargued. Identification rests on the fact that weather is random (Heal and Park, 2016), atleast from the perspective of the economy. The problem with this argument is that bynow many different economic activities have been found to be affected by the weather (seeAuffhammer and Aroonruengsawat, 2011, Barreca et al., 2016, Deschenes and Greenstone,2007, Graff Zivin et al., 2020, Leightner, 1999, Li et al., 2018, Pechan and Eisenack, 2014,Ranson, 2014, Zhang et al., 2018, among others), and these activities impact one another.Causality notwithstanding, these studies show that weather matters to the economy. How-ever, the impacts of weather shocks cannot readily be extrapolated to the impacts of climatechange (Dell et al., 2014, Kolstad and Moore, 2020). Climate is what you expect, weather iswhat you get. Weather are draws from an probability distribution. Climate is that distribu-tion. Climate change shifts the moments of the weather distribution (Auffhammer, 2018b).Weather is unpredictable for more than a few days ahead. Adaptation to weather shocksis therefore limited to immediate responses—put up an umbrella when it rains, close theflood gates when it pours. Adaptation to climate change extends to changes in the capi-tal stock—buy an umbrella, build flood gates. Furthermore, adaptation to climate changedepends on updates of the expectations for weather (Severen et al., 2018, Lemoine, 2017, Andersen et al. (2016) argue that it is UV radiation, rather than climate, that affects development. short-run elasticity,whereas the long-run elasticity is needed to estimate the impact of climate change.Hsiang (2016) and Deryugina and Hsiang (2017) argue that the marginal effect of a weathershock equals the marginal climate effect. Climate change is not marginal but its total impactis an integral of marginals. Their assumptions are quite restrictive, however. Economicagents need to be (1) rational and their adaptation investments (2) optimized. Adaptationneeds to be (3) private and adaptation options (4) continuous. The economy needs to be ina (5) spatial equilibrium and (6) markets complete. Adaptation investments are often long-lived, so both spot and future markets should be complete. Spatial zoning and transporthubs distort the spatial equilibrium. Adaptation is often lumpy, be it air conditioning orirrigation. Some adaptation options, such as coastal protection, are public goods. Otheradaptation options, such as protection against infectious disease, have externalities. Agentsare not always rational, and decisions suboptimal. The result by Deryugina and Hsiang isalmost an impossibility theorem.Weather affects economic activity, and so the measurement of the impact of climate oneconomic activity. Weather can be seen as noise, but that noise may well be correlatedwith climate, the right-hand-side variable of interest. I therefore propose a new way tosimultaneously model the impact of climate and weather, to show that both matter andthat previous work is misspecified.The empirical strategy rests on the following assumptions. Climate affects production pos-sibilities. This is obvious for agriculture: Holstein cows do well in Denmark but jasminerice does not; the reverse is true in Thailand. Climate also affects energy and transport,and thus all other sectors of the economy. Weather affects the realization of the productionpotential. Hot weather may slow down workers, frost may damage crops, floods may disrupttransport and manufacturing. Conceptualized thus, climate affects the production frontier,and weather the distance from that frontier. The econometric specification is therefore astochastic frontier analysis with weather variables in inefficiency and climate variables in thefrontier. Climate affects potential output, weather the output gap.I apply the proposed method to a panel of output per worker, measured at the country level.Dell et al. (2012), Letta and Tol (2018) and Newell et al. (2018) find that weather shocks hitthe economic growth of poorer countries harder. Burke et al. (2015) instead find that hotter countries are hit harder, a specification adopted by Pretis et al. (2018) and Kalkuhl andWenz (2020). Generoso et al. (2020) has a similar result. Within sample, it is difficult todistinguish between these two specifications as hotter countries tend to be poorer. However,out of sample, a hotter, richer world would be more vulnerable to weather shocks accordingto Burke, but less vulnerable according to Dell. Kahn et al. (2019) reject heterogeneity.The results below shed new light on these questions. Kumar and Khanna (2019) estimate the impact of temperature and rainfall on inefficiency in output growth . I here study inefficiency in output. They omit climate from the frontier. In the model below, climate andweather interact at the same level.The cross-validation study of ? finds that weather affects the level of GDP rather than itsgrowth rate, a specification adopted here in line with the intuition sketched above. Fur-thermore, I assume that the economy is affected by unusual weather rather than weather.Frost of -10 ℃ brought Texas to a standstill in February 2021, but is a regular occurrence inNorth Dakota without major consequences. I therefore standardize the weather, expressingtemperature and precipitation in standard deviations from the mean. This introduces aninteraction between weather and climate, and an implicit model of adaptation.The paper proceeds as follows. Section 2 describes methods and data. Section 3 presentsthe baseline results. Section 4 conducts the sensitivity analysis. Section 5 discusses theimplications for climate change. Section 6 concludes.
2. Methods and data
I assume a Cobb-Douglas production function: Y c,t = A c,t K βc,t L − βc,t (1)Total factor productivity A c,t is the Solow residual in country c at time t : It captureseverything that affects output Y c,t that cannot be explained by capital K c,t or labour L c,t .I concentrate Equation (1) by dividing K and L by labour force L , and denote the resultingvariables in lower case.Taking natural logarithms, the equation to be estimated is:ln y c,t = α + β ln k c,t (2)I assume that total factor productivity is a function of moving averages of weather variables(average temperature, ¯ T c,t , and precipitation, ¯ R c,t ). This is loosely based on Nordhaus(1992). Weather shocks affect the variance of the stochastic component of permanent income.Hence, Equation (2) becomes:ln y c,t = β ln k c,t + f (cid:0) ¯ T c,t , ¯ R c,t (cid:1) + µ c + t + v c,t − u c,t (3)where ¯ T c,t and ¯ R c,t are the average temperature c.q. precipitation in country c in the thirtyyears preceding year t , µ c is a full set of country fixed effects, t is a linear time trend, Bigano et al. (2006) use a similar model for tourist destination choice, with climate at destination atthe bottom and climate at origin at the top. c,t ∼ N (0 , σ v ) and u c,t ∼ E ( λ c,t ) = E (cid:18) γ + γ g (cid:18) T c,t − ¯ T c,t τ c,t (cid:19) + γ g (cid:18) R c,t − ¯ R c,t ρ c,t (cid:19)(cid:19) (4)where τ and ρ are the standard deviations of temperature and rainfall, respectively. Insteadof the unwieldy T − ¯ T / τ , I write z ( T ); ditto for R . This is standardized temperature andprecipitation. In the base specification, f (cid:0) ¯ T c,t , ¯ R c,t (cid:1) ≡ β ¯ T c,t + β ¯ T c,t + β ¯ R c,t + β ¯ R c,t + β ¯ T c,t ¯ R c,t and g ( · ) ≡ | · | . I refer to Equation (3) as the frontier or potential output, and toEquation (4) as inefficiency or the output gap.I use the True Fixed-Effect (TFE) model (Greene, 2005) to estimate a one-step stochasticfrontier model in a fixed-effect setting with explanatory variables in the inefficiency pa-rameter. I use the sfmodel package for Stata (Kumbhakar et al., 2015) to estimate themodel.Equation (3) assumes that both error terms are stationary. This is a tall assumption. Iam not aware of any statistical test for stationarity that applies to this particular estimatorand these distributional assumptions. I use three remedies. First, I include a time trend inEquation (3), and try many variants of that trend. Second, I show robustness to differentspecifications. Third, I reformulate the model as an error-correction one. The output gapfollows ∆ ln y c,t = ψ ∆ z ( T c,t ) + ψ ∆ z ( R c,t ) + ψ V c,t + µ c + w c,t (5)where potential output is V c,t = ln y c,t − µ c − µ t − ϑ ln k c,t − f ( ¯ T c,t , ¯ R c,t ) (6)and µ t are time dummies which act as a non-parametric time trend. This alternativeestimation strategy shows that the findings are robust to the inclusion of non-parametrictime trends. This alternative specification is also better suited to explicitly model the path ofconvergence towards the long-term equilibrium in a stochastic setting and provide empiricalevidence for the speed of recovery after weather perturbations. I of course also perform theusual stationarity tests on the error-correction model.I test for heterogeneity by interacting the variables of interest with dummies for poor coun-tries and hot countries. I define a country as “poor” if the World Bank does. Alternative,a country is deemed poor if its GDP per capita was below the 25th percentile of the dis- Taking first differences of all variables may get rid of unit roots in the frontier but would change thedistributional assumptions in inefficiency. Rob Engle (personal communication) suggests that standard stationary tests would roughly apply here. The use of a non-parametric time trend was not possible in the baseline SFA model because the inclusionof so many time dummies causes convergence issues in an already computationally cumbersome maximumlikelihood estimation. The WB classification of high-income economies is available here. A “hot” country is defined as a country whose average annualtemperature is above the 75th percentile of the distribution.
The dataset is an unbalanced panel consisting of 160 countries over the period 1950-2014.Data for this study come from two sources. Economic data on output, capital and labourforce are taken from the Penn World Table (PWT), PWT 9.0 (Feenstra et al., 2015). Weatherdata are from the University of Delaware’s
Terrestrial air temperature and precipitation:1900-2014 gridded time series, (V 4.01) (Matsuura and Willmott, 2015). These griddeddata have a resolution of 0 . × . ×
55 kilometers atthe equator. Following previous literature (Dell et al., 2014, Burke et al., 2015, Auffhammeret al., 2013), we aggregate these grid cells at the country-year level, weighting them bypopulation density in the year 2000 using population data from Version 4 of the
GriddedPopulation of the World . , with the exception of Singapore. We use these weather data toconstruct both the climate and weather variables as defined in Section 2.1. Table 1 presentsdescriptive statistics for the key variables.
3. Results
Table 2 shows the results of the base specification outlined in Equations (3) and (4). Sixvariants are presented. Column 1 reports homogeneous effects in both the frontier andthe inefficiency. In the frontier, capital per worker has a significant impact on output perworker. The output elasticity is around 0.63, in line with previous estimates. This estimateis robust to specification. Long-run temperature (i.e. climate) has a significant impact onthe production frontier, but precipitation does not, as in earlier papers (Dell et al., 2012,Burke et al., 2015, Letta and Tol, 2018). Short-term weather anomalies, either temperatureor precipitation, are insignificant in determining inefficiency.Columns 2 and 3 show heterogeneous impacts between rich and poor countries. The hy-pothesis is that poor countries are disproportionately affected by climate and weather, aseconomic activity is concentrated in agriculture and public investment in protective mea-sures is limited. Column 2 allows heterogeneity only in the production frontier. That is,I interact climate variables with the poor country dummy defined in Subsection 2.1. Theinteraction terms are individually insignificant. Column 3 adds heterogeneity in inefficiency.Results for the production frontier are almost unchanged. Impacts on inefficiency sharply Available here. Singapore has a surface smaller than the size of the weather grids. Given it is one of the few countriesthat are both rich and hot and thus increase the statistical power of the analysis, we kept it in the sampleby attributing to it the weather data of the grid cell in which it is situated. See the Appendix for a complete list of countries and regions in the sample. either poor countries (e.g. Dell et al., 2012, Letta and Tol, 2018) or hot countries (Burkeet al., 2015) particularly vulnerable to weather anomalies. I find both .Columns 1-4 specify that, in the frontier, hot and cold countries respond differently totemperature, and dry and wet countries differently to rainfall. Column 5 adds the interactionbetween rainfall and temperature to the frontier. This interaction is negative, but less so inpoor countries. The rainfall terms are now significant too: Wetter countries are richer, andthis effect is weaker for poor countries.Dropping the insignificant interaction terms between temperature and poverty (column 6)hardly affects the parameter estimates. Column 6 is the preferred specification.Weather anomalies increase inefficiency in poor countries, as expected. Weather anomalies decrease inefficiency in rich countries—that is, unexpectedly much or little water, or unusu-ally hot or cold weather stimulate the economy. This is harder to explain. It may reflect therestoration effort after floods, and crop insurance and government support after droughts.The data are GDP rather than NDP, and thus suffer from Bastiat’s broken window. Thiseffect is not observed in poor countries because restoration after natural disasters is limitedand delayed (Cavallo and Noy, 2011).I interpret the effect size below, after discussing the robustness of the results.
4. Robustness
I implement three different types of robustness checks: sensitivity to different specificationsin the SFA model; an alternative distributional assumption for the inefficiency parameter;and an error-correction model to formally test for non-stationarity. For all these sensitiv-ity tests, with the exception of the error-correction model, I only report estimates of thepreferred specification, column 6 of Table 2.
This first set of robustness checks implements the same baseline model described in Equa-tions (3) and (4) but adopts a broad set of different specification choices for key variablesand interactions.
I test whether the core findings are driven by the somewhat arbitrary discrimination betweenrich and poor countries. I replace the World Bank classification of countries that are rich Available here.
7y the “poorest 25% in 1990”. Results are in column 2 of Table 3. Column 1 repeats thebase specification (column 6) of Table 2.For the production frontier, results are qualitatively the same as in Table 2. The maindifference is that precipitation loses much of its predictive power, highlighting that differenteconomies do respond differently to the availability of water resources. As for the inefficiency,results are again qualitatively similar to the baseline model, but coefficients are closer tozero and less significant. The log-pseudolikelihood is much lower.
Second, I replace absolute weather anomalies in the inefficiency term with squared anomalies.This places a heavier weight on larger anomalies. See column 3 of Table 3. The results forthe production frontier are largely unaffected, and the qualitative results for the inefficiencyare as above. The log-pseudolikelihood falls.
The weather anomalies in Equation (4) are absolute anomalies. Cold and hot weather, wetand dry spells are assumed to equally increase technical inefficiency. Column 4 of Table 3instead use the anomalies. Estimates for the production frontier are almost unaltered. Theparameters for inefficiency become insignificant. Economies are affected by unusual weather,rather than by the weather per se . Adaptation matters.
I also test for asymmetric anomalies, disentangling negative and positive weather shocks oninefficiency. This is the preferred specification of Kahn et al. (2019). Results are in column5 of Table 3. The frontier is not affected. The results are much as above, with anomalousweather being good for rich countries but bad for hot and poor countries. While there issome evidence for asymmetry between the impact of wet and dry spells, cold and hot spells,the increase in the log-pseudolikelihood is minimal (less than 6 points) for the six additionalparameters estimated.
I also look at weather effects on productivity, moving weather anomalies from the inefficiencyparameter to the production frontier. Results are in column 6 of Table 3. The frontier doesnot change. Coefficients of weather variables are individually insignificant and the log-pseudolikelihood is sharply lower. This specification, variations of which are often used inliterature, is not the preferred one.
Equation (4) assumes an exponential half-normal distribution for inefficiency. Column 7 ofTable 3 show results for the half-normal distribution. The estimates for the frontier are as Truncated-normal models with fixed-effects are known to suffer severe convergence issues, and this casewas no exception. It is therefore excluded. √ . π − I find a significant association between climate and economic performance. In the con-centrated Cobb-Douglas production function, Equation (1), there are two determinants ofoutput per worker: climate and capital per worker. In this specification, capital is a defacto substitute for climate, with a constant elasticity. I test that assumption, answeringthe question whether sufficient capital would make a country immune from the influenceof its climate. I therefore interact long-run temperature variables with capital per workerin the production frontier. See Table 4, Columns 2 and 3; column 1 reproduces the basemodel from Table 2. Rainfall is significant and so are its interactions with capital. The in-teractions have the opposite signs. That is, climate’s influence on output shrinks as capitaldeepens. The interaction between temperature, rainfall and capital is insignificant. The log-pseudolikelihood increases by 7 points. However, interactions work both ways. The outputelasticity of capital now depends on rainfall, varying between 0.73 in the driest countriesand 0.93 in the wettest ones. A 5.5% increase in rainfall, well within the climate changeprojections for this century, would lead to increasing returns to scale and explosive economicgrowth. I therefore keep the base specification as is.Column 2 only changes the frontier. In column 3, I replace the interaction with the povertydummy by an interaction with capital per worker. Signs change and the log-pseudolikelihoodfalls. Poverty is more than a lack of capital, and poverty drives vulnerability to weathershocks.
In the debate on the long-run determinants of growth and development, some find thatclimate plays a fundamental role in shaping long-run development, whereas others arguethat the impact of climate disappears when accounting for institutions, although climatemay have shaped those institutions. I test this in column 4 of Table 4. As a proxy forinstitutional quality, I use the
Polity2 Score . This categorical variable is an aggregatescore which ranges from -10 (hereditary monarchy) to 10 (consolidated democracy). Whilethis is not the best indicator for institutional quality, it is correlated with other indicators.Historical depth is the key advantage of Polity2 over other indicators, which are availableonly for recent years. I interact it with long-run precipitation in the production frontier. The Polity Project Database, annual national data for the period 1800-2017, can be downloaded here.
Non-stationarity is a key concern in any long panel of economic data. The residuals of thestochastic frontier model do not pass a stationarity test. See Table A1. Panel stationaritytests require that the residuals of every country are stationary. Equation (3) has a commontrend for all countries. The panel is unbalanced, with fewer observations for hotter andpoorer countries in the early years. It should therefore not come as a surprise that themodel fails the test for panel cointegration.Table A2 shows the results if the model is estimated without a trend, a linear trend (asabove), and a polynomial trend of order two or three; and if a different linear trend is usedfor poor and for other countries. Qualitatively, the impact of climate and weather is thesame. The differences between estimates are not significant. Although the residuals of thealternative models are not stationary (results not shown), the stability of the results suggestthat the regression results are not spurious.Table A1 supports that suggestion. Output and capital per capita are non-stationary, butthe climate and weather variables are. That means that the residuals of the model are non-stationary because output and capital do not cointegrate (after inclusion of a trend). Theimpact of climate and weather on output per worker is not spurious—climate and weatherdo not explain the residual trend in output because there is no trend in the climate andweather data.The rightmost column of Table A1 re-estimates the model in first differences. Note thatthe difference between two exponential distribution is not an exponential distribution; anstochastic frontier model in first differences is a different specification. Reassuringly, theoutput elasticity of capital does not significantly change when the model is estimated in firstdifferences. The impact of weather and climate either becomes insignificant or much smaller.In the frontier, this is because the impact of climate is primarily estimated from cross-sectional variation, which disappears when differencing. The weather impact on inefficiency,rainfall in poor countries excepted, disappears in the noise.
As a further empirical test, I estimate the error-correction model (ECM) defined in Equa-tions (5) and (6). I assume that weather anomalies cause short-term deviations from thelong-run equilibrium, while climate affects the long-run equilibrium growth path of the econ-omy. The error-correction model is dynamic, unlike the stochastic frontier models above,tracking the time needed to absorb the perturbation caused by weather anomalies. TheECM specification allows for country and year fixed-effects, replacing the linear time trendin the stochastic frontier. 10able 5 presents the results for the long-run co-integrating vector, Table 6 for the short-runerror-correction. In the short-run error-correction estimates, V is the residual of Table 6,Column 4, since this specification fits the data best.The output elasticity of capital in the co-integrating vector is much the same as above.The climate variables and their interactions with the poverty dummy are not individuallysignificant, with a few occasional exceptions, but the log-pseudolikelihood reveals that theyare jointly significant: 162 points gain for 10 parameters. This is confirmed by Table A4:Without the climate variables, the Im et al. (2003) test firmly rejects the null-hypothesisthat the residuals are stationary.The cointegrating vector and the stochastic frontier model have the same signs on theclimate variables and on their interactions with poverty. Qualitatively, the above findingsare confirmed.Table A4 shows that the residuals of the short-run equation are stationary. Table 6 showsthe estimates. The cointegrating vector is highly significant. The parameter estimate of0.06 indicates rather fast convergence to the equilibrium relationship. Precipitation is notsignificant but temperature is, in poor countries. This result is qualitatively different fromthe stochastic frontier model—but similar to Dell et al. (2012).Note that the results in Table 6 are for the standardized temperature and precipitation,rather than their absolute values. This is a further deviation from the stochastic frontiermodel. Table A3 shows the results for the absolute anomalies. The results are much thesame, except that temperature now also affects rich countries. The log-pseudolikelihood islower, however.
5. Implications
The impact of climate change is highly nonlinear in this model. The effect size is thereforehard to grasp. Furthermore, there are 160 countries in the database. There are manyscenarios and models of climate change, and many scenarios and models of future economicgrowth. Exploring all possible futures is a combinatorial explosion, and would shed little lighton how the model presented here works. So instead, I used stylized scenarios to illustratethe impact of climate change, according to column 6 in Table 2, on the 2014 population,economy and climate.The production frontier, Equation (3), depends on the thirty-year average of the level oftemperature and precipitation. This is projected to change over time. Inefficiency, Equation(4), depends on the absolute value of the standardized temperature and rainfall. Withoutclimate change, there are weather shocks to inefficiency and hence economic output. Withclimate change, weather shocks are different.I consider warming between 1 ℃ and 6 ℃ , and 0.01 ℃ /year and 0.06 ℃ /year. This is therange shown in the Fifth Assessment Report of the Intergovernmental Panel on Climate11hange (IPCC). I let rainfall increase or decrease by up to 30%, again within the range ofexpectations for this century. The impact of these scenarios on the frontier is immediate.The impact of climate change on inefficiency follows from the deviation of the actual weatherfrom the expected weather. Without climate change, the expected temperature shock is zero.With a 3 ℃ per century warming, the expected temperature shock is 15 × . /τ c per year,where the factor 15 is there because I use the 30-year average and standard deviation fornormalization.Climate affects production possibilities, and anomalous weather the realisation of thosepossibilities. Climate change will affect both. Extrapolating statistical models is alwaystricky. Here, the frontier is estimated on a wide range of climates, while inefficiency dependson time-varying standardization of weather variables. Both help to make extrapolation morereliable.Figure 1 shows the global average impact, separately for changes in temperature and pre-cipitation. The impact on the frontier is not out of line with previous studies (Tol, 2018): A5% loss for 3 ℃ warming. The function is almost linear. The impact on inefficiency is morenon-linear, but smaller and positive because the impact on rich countries dominates.This is confirmed by the second set of graphs in Figure 1. The above results computethe global average output. The two remaining graphs compute the global average utility,expressed in its income equivalent, assuming a rate of risk aversion of one (Fankhauser et al.,1997). At the frontier, these equity-weighted impact are more linear and larger if warmingexceeds 3 ℃ . This is because poorer countries are hit harder by climate change at the frontier.This is more pronounced in inefficiency: The sign flips, and the global average impact issubstantially larger than on the frontier.The right panel of Figure 1 show the impact of changes in precipitation. At the frontier,the impacts are large. Drying would be a loss, wettening a gain. These impacts are lesspronounced if the national impacts are equity-weighted. This follows from Table 2: Poor,hot countries have smaller parameters. For inefficiency, change matters rather than thedirection of change; inefficiency is determined by deviations from experience, regardless ofwhether that deviation is more or less water than expected. The impacts are more modest.Equity-weighting again flips the sign: Poor countries are negatively affected, rich countriespositively.Figure 2 shows the results by country, for a 3 ℃ warming and a 20% increase in precipitationover a century. In all figures, the size of the bubble is proportional to the population size in2014.The top left figure shows the impact of warming on the frontier, plotted against the averagetemperature for 1985-2014. The spread is quite large, ranging from a 90% increase to a 70%decrease. Colder countries see more positive impacts, hotter countries more negative ones.The figure separates poor countries—which are essentially on a continuous lines—and richones—which are more dispersed because the impact of wealth is interacted with precipita-12ion. Richer countries face more negative impacts.The top right figure shows the impact of warming on inefficiency, plotted against the standarddeviation of the temperature for 1985-2014. Effect sizes are smaller than on the frontier,ranging between a 20% decline and a 15% increase, and fall for countries with greater climatevariability. There are three separate graphs, corresponding with the interactions in Table2. Rich countries see benefits, poor but cool countries moderate losses, and poor and hotcountries large losses.The bottom left figure plots the impact of wettening on the frontier against average precipi-tation in 1985-2014. Heterogeneity is again large, ranging from the 15% loss to a 90% gain.There is little structure in the graph.The bottom right figure plots the impact of wettening on inefficiency against the standarddeviation of precipitation in 1985-2014. Effect sizes are smaller than on the frontier, rangingbetween a 10% loss and the 15% gain, and fall with greater climate variability. There areagain three separate graphs. Rich countries see gains, poor and cool countries small losses,and poor and hot countries large losses.
6. Discussion and conclusion
I use stochastic frontier analysis to jointly model the impacts of weather and climate oneconomic activity in most countries over 65 years. I distinguish production potential, affectedby climate, and the realisation of economic output, affected by weather. Weather shocks thushave a transient effect, climate change a permanent impact. Warming affects productionpotential, negatively in cold, positively in hot countries; and more so in rich, wet countries.Changes in precipitation also affect the frontier. The impacts are heterogeneous without anobvious pattern. Climate change also affects inefficiency, particularly in countries with littleclimate variability, reducing the output gap in rich countries but increasing it in poor andhot countries. The weather effect is small compared to the climate effect. These results arequalitatively and quantitatively robust to alternative specifications, controls, and estimators.Dell et al. (2012) find that poor countries are particularly vulnerable to weather shocks,Burke et al. (2015) find that hot countries are. In the Burke (Dell) specification, countrieswould grow more (less) vulnerable to unusual weather in a hotter and richer future. I findthat both are true, and that the impact of heat is about as strong as the impact of poverty.Reduced outdoor work and manual labour, decreased relative importance agriculture inoutput and work force, and greater diffusion of adaptive capital such as air conditioningwould help poorer countries to dampen the negative effects of weather shocks—but only toa degree, as the effort needed to alleviate the heat rises with the temperature.The impact of weather shocks found here cannot directly be compared to previous studies.Letta and Tol (2018) model economic growth as a function of the change in temperature,Dell et al. (2012), Burke et al. (2015), Pretis et al. (2018) and Kalkuhl and Wenz (2020)as a function of the temperature level. Kahn et al. (2019) come closest to my specification,13ut they use (asymmetric) weather anomalies rather than standardized weather. Anotherkey difference with those papers is that, here, the impact of a weather shock is transitory.Unusual weather increases inefficiency, but the economy bounces back the next year, regis-tering higher growth. If my specification is right, then previous studies that excluded laggedtemperature effects are wrong. Previous studies, Barrios et al. (2010) and Generoso et al. (2020) excepted, did not finda significant impact of precipitation. This is a puzzling result, as droughts and floods aremore devastating than heat and cold. The same result is found here, in the frontier, unlessI interact precipitation with temperature and poverty. Net water—rainfall minus evapo-ration—matters rather than gross water—rainfall—and more so in countries that dependmore on agriculture. Precipitation also has a significant effect on inefficiency, one that variesstrongly with its variability. Previous studies did not standardize weather variables.The impact on the frontier is larger than in previous studies of the impact of climate change(Tol, 2018). Compared to some previous empirical studies (Easterly and Levine, 2003, Ro-drik et al., 2004), climate has a significant effect, also when controlling for institutionalquality, perhaps because I used more data (as did Nordhaus, 2006, Dell et al., 2009, Hender-son et al., 2018, Kalkuhl and Wenz, 2020), perhaps because I modelled heteroskedasticity.Previous studies did not do this and therefore their estimators would be inefficient and, ifweather-related heteroskedasticity correlates with climate, may be biased.Higher income, more capital nor better institutions fully insulate countries from the influenceof their climate. This contradicts earlier studies (Acemoglu et al., 2001, 2002, Alsan, 2015).I do not include all impacts of climate change. I omit direct impacts on human welfare, suchas biodiversity and health. The model does not capture the range of events which could betriggered by climate change but lie outside the current range of historical experience, suchas thawing permafrost(Wirths et al., 2018), a thermohaline circulation shutdown (Anthoffet al., 2016) or unprecedented sea level rise (Nordhaus, 2019). Because of data availability,I use democracy as a proxy for high-quality government. I limit the attention to aggregateeconomic activity. Adaptation and expectations are implicit in the model, as are productionrisks and risk preferences. The projections with respect to climate change are static, notdynamic.The numerical results are therefore far from final. The methodological advancement in thiswork is more important: the joint, simultaneous estimation of the impact of two different,but often confused, phenomena: weather and climate. I defer to future research the taskof refining the theoretical and empirical framework proposed here, and applying it to othermacro contexts and, crucially, household and firm data. The lags in Dell et al. (2012) are insignificant. eferences D. Acemoglu, S. Johnson, and J. A. Robinson. The colonial origins of comparative development: Anempirical investigation.
American Economic Review , 91(5):1369–1401, December 2001. doi: 10.1257/aer.91.5.1369.D. Acemoglu, S. Johnson, and J. A. Robinson. Reversal of fortune: Geography and institutions in the makingof the modern world income distribution.
The Quarterly Journal of Economics , 117(4):1231–1294, 2002.M. Alsan. The effect of the tsetse fly on African development.
American Economic Review , 105(1):382–410,January 2015. URL .T. B. Andersen, C.-J. Dalgaard, and P. Selaya. Climate and the emergence of global income differences.
The Review of Economic Studies , 83(4):1334–1363, 2016. URL http://dx.doi.org/10.1093/restud/rdw006 .D. Anthoff, F. Estrada, and R. S. J. Tol. Shutting down the Thermohaline Circulation.
American Eco-nomic Review , 106(5):602–06, May 2016. URL .M. Auffhammer. Climate adaptive response estimation: Short and long run impacts of climate changeon residential electricity and natural gas consumption using big data. Working Paper 24397, NationalBureau of Economic Research, March 2018a. URL .M. Auffhammer. Quantifying economic damages from climate change.
Journal of Economic Perspectives ,32(4):33–52, 2018b.M. Auffhammer and A. Aroonruengsawat. Simulating the impacts of climate change, prices and populationon california’s residential electricity consumption.
Climatic Change , 109(1):191–210, Dec 2011. URL https://doi.org/10.1007/s10584-011-0299-y .M. Auffhammer, S. M. Hsiang, W. Schlenker, and A. Sobel. Using weather data and climate model outputin economic analyses of climate change.
Review of Environmental Economics and Policy , 7(2):181–198,2013.A. Barreca, K. Clay, O. Deschenes, M. Greenstone, and J. Shapiro. Adapting to climate change: Theremarkable decline in the us temperature-mortality relationship over the twentieth century.
Journal ofPolitical Economy , 124(1):105–159, 2016. doi: 10.1086/684582.S. Barrios, L. Bertinelli, and E. Strobl. Trends in rainfall and economic growth in Africa: A neglectedcause of the African growth tragedy.
The Review of Economics and Statistics , 92(2):350–366, 2010. URL https://doi.org/10.1162/rest.2010.11212 .A. Bigano, J. M. Hamilton, and R. S. J. Tol. The impact of climate on holiday destination choice.
ClimaticChange , 76:389–406, 2006.D. Bloom and J. Sachs. Geography, demography, and economic growth in Africa.
Brookings Papers onEconomic Activity , (2):207–295, 1998. doi: 10.2307/2534695.C. Bogmans, G. Dijkema, and M. van Vliet. Adaptation of thermal power plants: The (ir)relevance ofclimate (change) information.
Energy Economics , 62:1–18, 2017. doi: 10.1016/j.eneco.2016.11.012.M. Burke, S. M. Hsiang, and E. Miguel. Global non-linear effect of temperature on economic production.
Nature , 527(7577):235–239, 2015.E. Cavallo and I. Noy. Natural disasters and the economy - a survey.
International Review of Environmentaland Resource Economics , 5(1):63–102, 2011. doi: 10.1561/101.00000039.M. Dell, B. F. Jones, and B. A. Olken. Temperature and income: Reconciling new cross-sectional andpanel estimates.
American Economic Review , 99(2):198–204, May 2009. URL .M. Dell, B. F. Jones, and B. A. Olken. Temperature shocks and economic growth: Evidence from thelast half century.
American Economic Journal: Macroeconomics , 4(3):66–95, July 2012. URL .M. Dell, B. F. Jones, and B. A. Olken. What do we learn from the weather? The new climate-economyliterature.
Journal of Economic Literature , 52(3):740–798, 2014.T. Deryugina and S. Hsiang. The marginal product of climate. Working Paper 24072, National Bureau ofEconomic Research, November 2017. URL . . Deschenes and M. Greenstone. The economic impacts of climate change: Evidence from agriculturaloutput and random fluctuations in weather. American Economic Review , 97(1):354–385, 2007.J. Diamond.
Guns, Germs and Steel: A short history of everybody for the last 13,000 years . W.W. Norton,New York, 1997.W. Easterly and R. Levine. Tropics, germs, and crops: how endowments influence economic develop-ment.
Journal of Monetary Economics , 50(1):3–39, 2003. ISSN 0304-3932. doi: https://doi.org/10.1016/S0304-3932(02)00200-3.S. Fankhauser, R. Tol, and D. Pearce. The Aggregation of Climate Change Damages: a Welfare TheoreticApproach.
Environmental & Resource Economics , 10(3):249–266, 1997. doi: 10.1023/A:1026420425961.R. C. Feenstra, R. Inklaar, and M. P. Timmer. The next generation of the penn world table.
AmericanEconomic Review , 105(10):3150–82, 2015.R. Generoso, C. Couharde, O. Damette, and K. Mohaddes. The growth effects of El Ni˜no and La Ni˜na:Local weather conditions matter.
Annals of Economics and Statistics , (140):83–126, 2020. URL .J. Graff Zivin, Y. Song, Q. Tang, and P. Zhang. Temperature and high-stakes cognitive performance:Evidence from the national college entrance examination in China.
Journal of Environmental Economicsand Management , 104, 2020. doi: 10.1016/j.jeem.2020.102365.W. Greene. Fixed and random effects in stochastic frontier models.
Journal of productivity analysis , 23(1):7–32, 2005.G. Heal and J. Park. Reflections—Temperature stress and the direct impact of climate change: A reviewof an emerging literature.
Review of Environmental Economics and Policy , 10(2):347–362, 2016. URL http://dx.doi.org/10.1093/reep/rew007 .J. V. Henderson, T. Squires, A. Storeygard, and D. Weil. The global distribution of economic activity:Nature, history, and the role of trade.
The Quarterly Journal of Economics , 133(1):357–406, 2018. URL http://dx.doi.org/10.1093/qje/qjx030 .S. Hsiang. Climate econometrics.
Annual Review of Resource Economics , 8(1):43–75, 2016. URL https://doi.org/10.1146/annurev-resource-100815-095343 .E. Huntington.
Civilization and climate . Yale University Press, New Haven, 1915. URL https://jscholarship.library.jhu.edu/bitstream/handle/1774.2/33468/31151027498744.pdf .K. S. Im, M. Pesaran, and Y. Shin. Testing for unit roots in heterogeneous panels.
Journal ofEconometrics , 115(1):53–74, 2003. URL .M. E. Kahn, K. Mohaddes, R. N. C. Ng, M. H. Pesaran, M. Raissi, and J.-C. Yang. Long-term macroeconomiceffects of climate change: A cross-country analysis. Working paper, University of Cambridge, 2019.M. Kalkuhl and L. Wenz. The impact of climate conditions on economic production. Evidence from aglobal panel of regions.
Journal of Environmental Economics and Management , 103:102360, 2020. URL .T. Kjellstrom, R. Kovats, S. Lloyd, T. Holt, and R. Tol. The direct impact of climate change on regionallabor productivity.
Archives of Environmental and Occupational Health , 64(4):217–227, 2009.M. Koetse and P. Rietveld. The impact of climate change and weather on transport: An overview ofempirical findings.
Transportation Research Part D: Transport and Environment , 14(3):205–221, 2009.doi: 10.1016/j.trd.2008.12.004.C. D. Kolstad and F. C. Moore. Estimating the economic impacts of climate change using weather obser-vations.
Review of Environmental Economics and Policy , 14(1):1–24, 2020. URL https://doi.org/10.1093/reep/rez024 .S. Kumar and M. Khanna. Temperature and production efficiency growth: empirical evidence.
ClimaticChange , 156(1-2):209–229, 2019. doi: 10.1007/s10584-019-02515-5.S. C. Kumbhakar, H. Wang, and A. P. Horncastle.
A practitioner’s guide to Stochastic Frontier Analysisusing Stata . Cambridge University Press, 2015.J. Leightner. Weather-induced changes in the tradeoff between SO and NO ( x ) at large power plants. EnergyEconomics , 21(3):239–259, 1999. doi: 10.1016/S0140-9883(98)00018-8. . Lemoine. Expect above average temperatures: Identifying the economic impacts of climate change.Working Paper 23549, National Bureau of Economic Research, June 2017. URL .M. Letta and R. S. J. Tol. Weather, climate and total factor productivity. Environmental and ResourceEconomics , 2018. URL https://doi.org/10.1007/s10640-018-0262-8 .J. Li, L. Yang, and H. Long. Climatic impacts on energy consumption: Intensive and extensive margins.
Energy Economics , 71:332–343, 2018. doi: 10.1016/j.eneco.2018.03.010.W. Lise and R. Tol. Impact of climate on tourist demand.
Climatic Change , 55(4):429–449, 2002.E. Mansur, R. Mendelsohn, and W. Morrison. Climate change adaptation: A study of fuel choice andconsumption in the US energy sector.
Journal of Environmental Economics and Management , 55(2):175–193, 2008.K. Matsuura and C. Willmott. Terrestrial air temperature and precipitation: 1900–2014 gridded time series(version 4.01), 2015.R. Mendelsohn, W. D. Nordhaus, and D. Shaw. The impact of global warming on agriculture: A ricardiananalysis.
The American Economic Review , 84(4):753–771, 1994. URL .R. G. Newell, B. C. Prest, and S. Sexton. The GDP temperature relationship: Implications for climatechange damages. Technical report, Resources for the Future, 2018. URL .W. Nordhaus. Economics of the disintegration of the greenland ice sheet.
Proceedings of the NationalAcademy of Sciences , 116(25):12261–12269, 2019. URL .W. D. Nordhaus. An optimal transition path for controlling greenhouse gases.
Science , 258(5086):1315–1319,1992. URL https://science.sciencemag.org/content/258/5086/1315 .W. D. Nordhaus. Geography and macroeconomics: New data and new findings.
Proceedings of the Na-tional Academy of Science , 103(10):3510–3517, 2006. URL .O. Olsson and D. J. Hibbs. Biogeography and long-run economic development.
European Economic Review ,49(4):909–938, 2005.A. Pechan and K. Eisenack. The impact of heat waves on electricity spot markets.
Energy Economics , 43:63–71, 2014. doi: 10.1016/j.eneco.2014.02.006.F. Pretis, M. Schwarz, K. Tang, K. Haustein, and M. R. Allen. Uncertain impacts on economic growth whenstabilizing global temperatures at 1.5 ℃ or 2 ℃ warming. Philosophical Transactions of the Royal Society A:Mathematical, Physical and Engineering Sciences , 376(2119):20160460, 2018. doi: 10.1098/rsta.2016.0460.M. Ranson. Crime, weather, and climate change.
Journal of Environmental Economics and Management ,67(3):274 – 302, 2014. ISSN 0095-0696. URL .D. Rodrik, A. Subramanian, and F. Trebbi. Institutions rule: The primacy of institutions over geographyand integration in economic development.
Journal of Economic Growth , 9(2):131–165, 2004. URL https://doi.org/10.1023/B:JOEG.0000031425.72248.85 .J. Sachs and P. Malaney. The economic and social burden of malaria.
Nature , 415(6872):680–685, 2002.J. D. Sachs. Institutions don’t rule: Direct effects of geography on per capita income. Working Paper 9490,National Bureau of Economic Research, February 2003. URL .W. Schlenker, W. M. Hanemann, and A. C. Fisher. Will us agriculture really benefit from global warming?accounting for irrigation in the hedonic approach.
American Economic Review , 95(1):395–406, 2005.C. Severen, C. Costello, and O. Deschˆenes. A forward-looking Ricardian approach: Do land markets capital-ize climate change forecasts?
Journal of Environmental Economics and Management , 89:235–254, 2018.ISSN 0095-0696. doi: https://doi.org/10.1016/j.jeem.2018.03.009.R. S. J. Tol. The economic impacts of climate change.
Review of Environmental Economics and Policy , 12(1):4–25, 2018. URL http://dx.doi.org/10.1093/reep/rex027 .H. Wirths, J. Rathmann, and P. Michaelis. The permafrost carbon feedback in DICE-2013R modelingand empirical results.
Environmental Economics and Policy Studies , 20(1):109–124, January 2018. doi: Nature Climate Change , 5(7):647–651, 2015.P. Zhang, O. Deschenes, K. Meng, and J. Zhang. Temperature effects on productivity and factor reallocation:Evidence from a half million Chinese manufacturing plants.
Journal of Environmental Economics andManagement , 88:1 – 17, 2018. ISSN 0095-0696. doi: https://doi.org/10.1016/j.jeem.2017.11.001. URL . able 1: Descriptive statisticsVariable Unit Symbol Mean Std Dev Min Max ObsOutput per worker ln( $ ) ln( y ) 9.768 1.183 6.047 13.318 7753Capital per worker ln( $ ) ln( k ) 10.831 1.392 5.650 14.524 7753Temperature ℃ ¯ T R | z ( T ) | | z ( R ) | P H G able 2: Baseline results.Dependent variable: ln(output per worker). (1) (2) (3) (4) (5) (6)frontierln( k ) 0.628 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (79.19) (79.35) (77.16) (77.18) (76.24) (77.76) T ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (6.28) (3.76) (3.78) (4.02) (5.23) (8.58) T -0.00380 ∗∗∗ -0.00662 ∗∗ -0.00750 ∗∗ -0.00730 ∗∗ -0.00320 -0.00457 ∗∗∗ (-5.87) (-2.75) (-2.89) (-3.15) (-1.47) (-5.42) R ∗∗∗ ∗∗∗ (1.77) (1.34) (1.01) (1.42) (6.75) (6.89) R -0.000236 -0.00133 -0.000973 -0.00148 0.00554 ∗∗ ∗∗ (-0.69) (-0.97) (-0.70) (-1.09) (2.94) (2.85) P × T P × T P × R -0.00236 0.00565 0.000946 -0.140 ∗∗ -0.146 ∗∗ (-0.08) (0.19) (0.03) (-2.68) (-3.05) P × R ∗∗ -0.00614 ∗∗ (0.68) (0.42) (0.69) (-3.08) (-2.96) T × R -0.0198 ∗∗∗ -0.0199 ∗∗∗ (-8.55) (-8.45) P × T × R ∗∗∗ ∗∗∗ (6.20) (6.50)inefficiency | z ( T ) | ∗ -0.189 ∗∗ -0.200 ∗∗∗ -0.202 ∗∗∗ (1.27) (1.30) (-2.10) (-3.17) (-3.47) (-3.54) | z ( R ) | ∗ -0.229 ∗∗∗ -0.269 ∗∗∗ -0.266 ∗∗∗ (0.15) (0.19) (-2.39) (-3.86) (-4.56) (-4.53) P × | z ( T ) | ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (3.24) (3.70) (3.95) (4.05) P × | z ( R ) | ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (3.06) (3.48) (4.02) (4.01) H × | z ( R ) | ∗∗ ∗∗∗ ∗∗∗ (3.11) (3.50) (3.49) H × | z ( R ) | ∗∗ ∗∗ ∗∗ (2.83) (3.26) (3.25)Observations 7753 7753 7753 7753 7753 7753LpL 2441.1 2454.8 2487.0 2507.9 2557.9 2556.5 t statistics in parentheses ∗ p < . ∗∗ p < . ∗∗∗ p < . able 3: Robustness checks. Dependent variable: ln(output per worker). (1) (2) (3) (4) (5) (6) (7)frontierln( k ) 0.631 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (77.76) (79.66) (78.92) (78.62) (77.05) (79.67) (79.73) T ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (8.58) (7.43) (8.32) (8.14) (8.46) (8.14) (10.89) T -0.00457 ∗∗∗ -0.00337 ∗∗∗ -0.00446 ∗∗∗ -0.00449 ∗∗∗ -0.00451 ∗∗∗ -0.00452 ∗∗∗ -0.00666 ∗∗∗ (-5.42) (-4.90) (-5.18) (-5.01) (-5.30) (-5.14) (-8.26) R ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (6.89) (5.90) (6.54) (6.22) (6.96) (6.16) (7.58) R ∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗∗∗ (2.85) (0.59) (2.83) (2.61) (2.78) (2.68) (5.44) T × R -0.0199 ∗∗∗ -0.00553 ∗∗∗ -0.0192 ∗∗∗ -0.0189 ∗∗∗ -0.0202 ∗∗∗ -0.0187 ∗∗∗ -0.0245 ∗∗∗ (-8.45) (-7.24) (-7.92) (-7.52) (-8.67) (-7.34) (-10.70) P × R -0.146 ∗∗ -0.148 ∗ -0.135 ∗∗ -0.134 ∗∗ -0.151 ∗∗ -0.130 ∗∗ -0.205 ∗∗∗ (-3.05) (-2.16) (-2.79) (-2.69) (-3.12) (-2.63) (-4.56) P × R -0.00614 ∗∗ -0.000814 -0.00613 ∗∗ -0.00582 ∗∗ -0.00605 ∗∗ -0.00599 ∗∗ -0.00911 ∗∗∗ (-2.96) (-0.94) (-2.94) (-2.67) (-2.92) (-2.75) (-5.40) P × T × R ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (6.50) (3.31) (6.05) (5.67) (6.63) (5.67) (9.29) | z ( T ) | -0.000452(-0.13) | z ( R ) | -0.000418(-0.12) P × | z ( T ) | -0.00469(-0.88) P × | z ( R ) | H × | z ( R ) | -0.00892(-1.35) H × | z ( R ) | f ( z ( T )) -0.202 ∗∗∗ -0.0641 -0.0778 ∗∗∗ -0.0508 -0.224 ∗∗∗ (-3.54) (-1.52) (-3.52) (-1.17) (-4.08) f ( z ( R )) -0.266 ∗∗∗ -0.120 ∗∗ -0.0961 ∗∗∗ -0.0195 -0.254 ∗∗∗ (-4.53) (-2.73) (-4.64) (-0.48) (-4.15) P × f ( z ( T )) 0.259 ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ (4.05) (2.96) (4.52) (1.79) (4.31) P × f ( z ( R )) 0.282 ∗∗∗ ∗∗∗ ∗∗∗ -0.0307 0.292 ∗∗∗ (4.01) (3.65) (3.58) (-0.60) (3.72) H × f ( z ( T )) 0.229 ∗∗∗ ∗ ∗ ∗ H × f ( z ( R )) 0.231 ∗∗ ∗ ∗∗∗ ∗ (3.25) (2.21) (3.73) (1.09) (2.08) ( T ) + -0.148 ∗ (-2.25) z ( T ) − ∗∗∗ (3.94) z ( R ) + -0.306 ∗∗∗ (-4.62) z ( R ) − ∗∗ (3.28) P × z ( T ) + ∗∗ (2.86) P × z ( T ) − -0.351 ∗∗∗ (-3.41) P × z ( R ) + ∗∗∗ (3.37) P × z ( R ) − -0.326 ∗∗∗ (-3.49) H × z ( T ) + ∗∗ (3.04) H × z ( T ) − -0.349 ∗∗ (-2.83) H × z ( P ) + ∗∗∗ (3.57) H × z ( P ) − -0.114(-1.32)Observations 7753 7753 7753 7753 7753 7753 7753LpL 2556.5 2517.9 2527.5 2499.8 2561.2 2496.2 2185.8 P = Poverty dummy, World Bank definition except in Column (2): 25%ile of 1990 income distribution.Columns (1), (2), (6), (7): f ( z ( · )) ≡ | z ( · ) | ; column (3): f ( z ( · )) ≡ z ( · ) ; column (4): f ( z ( · )) ≡ z ( · )Columns (1)-(6): Exponential distribution; column (7): Half-normal distribution. t statistics in parentheses ∗ p < . ∗∗ p < . ∗∗∗ p < . able 4: Robustness checks. Dependent variable: ln(output per worker). (1) (2) (3) (4)frontierln( k ) 0.631 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (77.76) (27.15) (29.57) (76.47) T ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (8.58) (7.69) (7.95) (11.66) T -0.00457 ∗∗∗ -0.00307 ∗∗∗ -0.00342 ∗∗∗ -0.00457 ∗∗∗ (-5.42) (-4.52) (-4.90) (-6.93) R ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (6.89) (6.73) (6.43) (7.25) R ∗∗ -0.0103 ∗∗∗ -0.00950 ∗∗∗ T × R -0.0199 ∗∗∗ -0.00730 ∗∗∗ -0.00630 ∗∗∗ -0.00611 ∗∗∗ (-8.45) (-4.57) (-4.02) (-9.06) P × R -0.146 ∗∗ (-3.05) P × R -0.00614 ∗∗ (-2.96) P × T × R ∗∗∗ (6.50)ln( k ) × R -0.0277 ∗∗∗ -0.0236 ∗∗∗ (-5.26) (-4.92)ln( k ) × R ∗∗∗ ∗∗∗ (6.75) (6.50)ln( k ) × T × R × R -0.000467(-0.99)Polity2 × R × T × R ∗ (2.38)inefficiency | z ( T ) | -0.202 ∗∗∗ -0.233 ∗∗∗ ∗∗∗ -0.201 ∗∗∗ (-3.54) (-3.88) (3.64) (-3.61) P × | z ( T ) | ∗∗∗ ∗∗∗ ∗∗∗ (4.05) (4.38) (4.06)ln( k ) × | z ( T ) | -0.0792 ∗∗∗ (-3.72) | z ( R ) | -0.266 ∗∗∗ -0.237 ∗∗∗ ∗ -0.180 ∗∗∗ (-4.53) (-3.91) (1.96) (-3.40) P × | z ( R ) | ∗∗∗ ∗∗∗ ∗∗ (4.01) (3.57) (2.87) n( k ) × | z ( R ) | -0.0531 ∗ (-2.21) H × | z ( T ) | ∗∗∗ ∗∗ ∗∗ ∗∗∗ (3.49) (2.69) (2.60) (4.27) H × | z ( R ) | ∗∗ ∗ ∗ ∗∗ (3.25) (2.25) (2.11) (3.07)Observations 7753 7753 7753 7099LpL 2556.5 2563.6 2547.1 2621.2 t statistics in parentheses ∗ p < . ∗∗ p < . ∗∗∗ p < . able 5: Cointegrating vector.Dependent variable: ln(output per worker) (1) (2) (3) (4) (5)ln( k ) 0.609 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (14.65) (14.42) (15.36) (17.30) (17.28) T ∗ ∗∗ (1.66) (1.18) (2.11) (2.68) T -0.00697 ∗∗ -0.0112 ∗ -0.00783 -0.00810 ∗∗∗ (-3.26) (-2.41) (-1.63) (-3.37) R -0.0234 0.0196 0.261 0.287 ∗ (-0.47) (0.22) (1.84) (2.12) R P × T P × T P × R -0.0392 -0.222 -0.270(-0.37) (-1.35) (-1.78) P × R T × R -0.0235 ∗ -0.0262 ∗ (-1.98) (-2.51) P × T × R ∗ (1.65) (2.34)Observations 7753 7753 7753 7753 7753LpL 2186.2 2250.7 2295.4 2342.6 2339.1 t statistics in parentheses ∗ p < . ∗∗ p < . ∗∗∗ p < . able 6: Short-run error-correction.Dependent variable: ∆ln(output per worker) (1) (2) (3) (4) (5)Output gap 0.0634 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (7.54) (7.52) (7.50) (7.50) (7.51)∆ z ( T ) -0.00134 ∗ z ( R ) 0.0000500 -0.000546 -0.000463(0.09) (-0.85) (-0.78)∆ P × z ( T ) -0.00253 ∗∗ -0.00246 ∗ -0.00266 ∗∗ (-2.71) (-2.29) (-2.95)∆ P × z ( R ) 0.000845 0.000949(0.81) (0.81)∆ H × z ( T ) -0.000258(-0.20)∆ H × z ( R ) -0.000629(-0.43)Observations 7591 7591 7591 7591 7591LpL 10223.0 10225.8 10228.7 10228.8 10228.3 t statistics in parentheses ∗ p < . ∗∗ p < . ∗∗∗ p < . igure 1: The change in global average output per worker due to changing temperature (left panel) andprecipitation (right panel). igure 2: The change in national average output per worker due to changing temperature (top panels) andprecipitation (bottom panels), in the frontier (left panels) and inefficiency (right panels). The bubble sizeis proportional to population size. able A1: Stationarity tests series statistic value p-valuedependent variableln( y ) Z ¯˜ t k ) Z ¯˜ t T Z ¯˜ t -5.9142383 1.667e-09 T Z ¯˜ t -4.8816254 5.261e-07 R Z ¯˜ t -4.0643271 .00002409 R Z ¯˜ t -4.827023 6.929e-07 T × R Z ¯˜ t -6.5881907 2.226e-11 | z ( T ) | Z ¯˜ t -47.886946 0 | z ( R ) | Z ¯˜ t -49.250722 0residualsFrontier Z ¯˜ t -.19176587 .4239628Inefficiency Z ¯˜ t -.42786821 .33437354Frontier + inefficiency Z ¯˜ t -.205783 .41848021The null hypothesis is stationarity for each country (Im et al., 2003).29 able A2: Robustness. Dependent variable: ln(output per worker). no trend linear quadratic cubic split first diff.frontierln( k ) 0.692 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (117.70) (77.76) (77.91) (74.82) (75.62) (21.21) T ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗ (9.46) (8.58) (7.71) (6.11) (8.08) (2.17) T -0.00244 ∗∗ -0.00457 ∗∗∗ -0.00463 ∗∗∗ -0.00440 ∗∗∗ -0.00392 ∗∗∗ -0.00233(-3.10) (-5.42) (-5.31) (-6.01) (-4.65) (-1.60) R ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ R ∗∗ ∗∗ ∗∗ ∗ ∗∗ T × R -0.0185 ∗∗∗ -0.0199 ∗∗∗ -0.0199 ∗∗∗ -0.0200 ∗∗∗ -0.0221 ∗∗∗ -0.00209(-6.82) (-8.45) (-8.44) (-8.75) (-9.00) (-1.79) P × R -0.138 ∗ -0.146 ∗∗ -0.146 ∗∗ -0.197 ∗∗∗ -0.212 ∗∗∗ -0.119 ∗ (-2.47) (-3.05) (-3.05) (-4.82) (-4.55) (-2.11) P × R -0.00438 ∗ -0.00614 ∗∗ -0.00617 ∗∗ -0.00461 ∗ -0.00664 ∗∗ -0.00121(-2.22) (-2.96) (-2.95) (-2.41) (-2.99) (-1.25) P × T × R ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗ (4.45) (6.50) (6.50) (7.42) (7.57) (2.68)inefficiency | z ( T ) | -0.216 ∗∗∗ -0.202 ∗∗∗ -0.203 ∗∗∗ -0.207 ∗∗∗ -0.187 ∗∗∗ -0.118(-3.76) (-3.54) (-3.53) (-3.48) (-3.33) (-1.28) | z ( R ) | -0.215 ∗∗∗ -0.266 ∗∗∗ -0.268 ∗∗∗ -0.299 ∗∗∗ -0.282 ∗∗∗ -0.106(-3.75) (-4.53) (-4.50) (-4.74) (-4.88) (-0.82) P × | z ( T ) | ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ P × | z ( R ) | ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗ (3.41) (4.01) (3.98) (4.13) (4.35) (2.05) H × | z ( T ) | ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ H × | z ( R ) | ∗∗ ∗∗ ∗∗ ∗∗∗ ∗∗∗ -0.0492(2.70) (3.25) (3.25) (3.33) (3.73) (-0.28) L. | z ( T ) | L. | z ( R ) | -0.0305(-0.29) P × L. | z ( T ) | -0.201(-1.75) P × L. | z ( R ) | H × L. | z ( T ) | -0.0466(-0.39) H × L. | z ( R ) | bservations 7753 7753 7753 7753 7753 7591LpL 2439.0 2556.5 2556.7 2775.4 2585.7 11337.1 t statistics in parentheses ∗ p < . ∗∗ p < . ∗∗∗ p < . able A3: Short-run error-correction.Dependent variable: ∆ln(output per worker) (1) (2) (3) (4) (5)Output gap 0.0634 ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ (7.54) (7.54) (7.54) (7.54) (7.54)∆( | z ( T ) | ) -0.00107 0.00120 ∗ ∗∗ ∗ (-1.51) (2.32) (2.77) (2.40)∆( | z ( R ) | ) -0.00121 -0.000721 -0.00119(-1.76) (-0.65) (-1.18)∆( P × | z ( T ) | ) -0.00329 ∗∗ -0.00297 ∗ -0.00338 ∗∗ (-2.92) (-2.43) (-3.05)∆( P × | z ( R ) | ) -0.000632 -0.000951(-0.45) (-0.61)∆( H × | z ( T ) | ) -0.00171(-1.07)∆( H × | z ( R ) | ) 0.00251(1.35)Observations 7591 7591 7591 7591 7591LpL 10223.0 10225.3 10227.6 10229.1 10226.3 t statistics in parentheses ∗ p < . ∗∗ p < . ∗∗∗ p < . able A4: Stationarity tests—Error-Correction Model model statistic value p-valueLong-run(1) Z ¯˜ t -.62047954 .26747106(2) Z ¯˜ t Z ¯˜ t -2.7926792 .00261368(4) Z ¯˜ t -1.9118767 .027946(5) Z ¯˜ t -.32010646 .37444381Short-run(4) + (1) Z ¯˜ t -41.963767 0(4) + (2) Z ¯˜ t -42.106363 0(4) + (3) Z ¯˜ t -41.939126 0(4) + (4) Z ¯˜ t -41.941283 0(4) + (5) Z ¯˜ t -41.95429 0The null hypothesis is stationarity for each country (Im et al., 2003).“model” refers to columns in Tables 5 and 6.33 ist of countries AlbaniaAlgeriaAngolaArgentinaArmeniaAustraliaAustriaAzerbaijanBahamasBangladeshBelarusBelgiumBelizeBeninBhutanBoliviaBosnia and HerzegovinaBotswanaBrazilBruneiBulgariaBurkina FasoBurundiCabo VerdeCambodiaCameroonCanadaCentral African RepublicChadChileChinaColombiaComorosCongoCosta RicaCroatiaCyprusCzech RepublicD.R. of the CongoDenmarkDjibouti 34ominican RepublicEcuadorEgyptEl SalvadorEquatorial GuineaEstoniaEthiopiaFijiFinlandFranceGabonGambiaGeorgiaGermanyGhanaGreeceGuatemalaGuineaGuinea-BissauHaitiHondurasHungaryIcelandIndiaIndonesiaIranIraqIrelandIsraelItalyIvory CoastJamaicaJapanJordanKazakhstanKenyaKuwaitKyrgyzstanLao People’s DRLatviaLebanonLesothoLiberia 35ithuaniaLuxembourgMacedoniaMadagascarMalawiMalaysiaMaliMauritaniaMauritiusMexicoMongoliaMontenegroMoroccoMozambiqueMyanmarNamibiaNepalNetherlandsNew ZealandNicaraguaNigerNigeriaNorwayOmanPakistanPanamaParaguayPeruPhilippinesPolandPortugalQatarRepublic of KoreaRepublic of MoldovaRomaniaRussian FederationRwandaSao Tome and PrincipeSaudi ArabiaSenegalSerbiaSierra LeoneSingapore 36lovakiaSloveniaSouth AfricaSpainSri LankaSt. Vincent and the GrenadinesSudan (Former)SurinameSwazilandSwedenSwitzerlandSyriaTaiwanTajikistanTanzaniaThailandTogoTrinidad and TobagoTunisiaTurkeyTurkmenistanUgandaUkraineUnited Arab EmiratesUnited KingdomUnited StatesUruguayUzbekistanVenezuelaVietnamYemenZambiaZimbabwe