## wpiea2022156-print-pdf ## Source details **Canonical URL:** [wpiea2022156-print-pdf](https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022156-print-pdf.pdf) ## Other formats - [Markdown version](/-/media/files/publications/wp/2022/english/wpiea2022156-print-pdf.pdf.md) - [Structured JSON version](/-/media/files/publications/wp/2022/english/wpiea2022156-print-pdf.pdf.json) --- ### Data and methods - Weather dataset: - ERA5 hourly reanalysis aggregated to daily level via Google Earth Engine (GEE); grid resolution up to 30×30 km at the equator; time span 1979–2019 (41 years). - Resulting raw ERA5-derived temperature and precipitation datapoints: 279,936×365×41×4 = 16,756,968,960 («17 billion). - PDSI observations: 10,110,022 land cells × 12 × 41 = 4,974,130,917 («5 billion). - Combined starting point ~21,715,195,392 («22 billion) datapoints; merged dataset produces 9,621,976 («10 million) country-matched grid cells for core variables; full climate-variable dataset uses 580,705,495,552 («600 billion) datapoints; population-weighted additions ≈ 14,517,637,388,800 («15 trillion). - Constructed climate variables: - 45 temperature variables, 29 precipitation variables, 8 wetness-drought variables → 82 unique climate variables. - After dropping collinear variables and adding first and second lags and population-weighted counterparts → 480 climate variables used in estimation. - Temperature constructs include absolute thresholds (e.g., TX35: max T above 35 ˚C), 3 ˚C intervals from -9 ˚C to 30 ˚C, heat/cold wave durations and counts. - Precipitation constructs include wet day definitions (>1 mm), PX1 and PX5 (1- and 5-day maxima), Longest Dry Spell (LLDS) at thresholds .50, .65, .80, .95, MaxP>10, MaxP>20, etc. - Wetness/drought uses PDSI with categories PDSI < −3, PDSI < −4, PDSI > 3, PDSI > 4 and intensity/maximum-extent measures (MPDSI metrics). - Macro-fiscal and macro variables: - GDP per capita: World Bank WDI "GDP per Capita constant 2015 US$" (NY.GDP.PCAP.KD); fiscal variables: IMF WEO government revenue/expenditure ratios (GGRNGDP, GGXNGDP); gross debt from Global Debt Database ("Central government debt, % of GDP"). - Empirical strategy: - Baseline regressions on first differences of macro variables on first differences of selected weather variables; country and year fixed effects; include first two lags of dependent variable; country-specific polynomial trends used in levels specifications; local projections (Jorda, 2005) for impulse responses up to horizon 7. - Variable selection: - LASSO applied to residualized data (partial-out approach following Belloni et al., 2014). - Hyper-parameter λ chosen via two-stage process: - Stage 1: random search k-fold cross-validation (five-fold), 200 random λ draws (half-normal variance 0.05) → 1,000 R-squared values; pick λ with largest average out-of-sample R-square (tends to select very small λ keeping ~20 variables). - Stage 2: grid search increasing λ by 0.001 increments until no variable dropped for three consecutive increments (+0.003 stability rule); AIC and BIC used to assess selected models. - Final LASSO-selected variables re-estimated by standard OLS for unbiased effect estimates. ### Sample, summary statistics, and trends - Sample sizes: - GDP analysis: 199 countries, 6,550 country-year observations (countries with ≥7-year time series; excluded Libya, Iraq, Equatorial Guinea, Bahrain). - Fiscal analysis: 159 countries, 3,859 country-year observations (after additional exclusions and removing observations with missing fiscal values). - Macroeconomic summary statistics (first differences): - GDP per capita growth (largest sample): Mean = 1.6 percent per year; within standard deviation = 4.7 percentage points. - GDP per capita growth (fiscal sample): Mean = 2.0 percent per year; within standard deviation = 3.7 percentage points. - Government revenue and expenditure ratios to GDP: average growth = 0.1 percentage points per year; within standard deviations range 3.2 to 4.0 percentage points. - Government debt-to-GDP: stable on average; within standard deviation = 11.1 percentage points. - Weather variable trends: - Average annual temperature trends across countries range from 0.07 to 0.6 ˚C per decade (noted elsewhere as average annual temperature positive trend = 0.03 ˚C per year in Appendix trend table). - Standard deviation of average annual temperature usually about 0.5 ˚C. - Many weather variables show significant country-specific trends (temperature-related trends more widespread than precipitation trends). - Unit-root tests on first differences of weather variables reject unit roots with p-values close to zero. ### Key GDP findings - LASSO selection and parsimonious baseline: - Initial random-search LASSO selected 30 variables with λ = 0.0166; after grid search final preferred selection includes three climate variables plus two lags of dependent variable. - Preferred specification improves within R-square by 7 percent relative to a specification with only the 2 lags of the dependent variable; the 30-variable random-search set yielded a 17 percent improvement in within R-square. - Baseline within R-square reference reported as within R2 = ".0937" for the specification with only the first two lags of the dependent variable. - BIC minimized for selections of 4 to 6 variables; AIC declines as selection grows up to the random-search set; preferred selection of 5 variables supported by criteria. - Three selected climate variables (baseline): - Harsh Drought Prevalence — PDSI < −4 (share of grid-days with harsh drought conditions) (population-weighted in LASSO). - Max T ˚C above 35 — TX35 (share of grid-days with maximum daily temperature above 35 ˚C) (population-weighted in LASSO). - Mean T ˚C in [9; 12) — TS9,12 (share of grid-days with mean temperature in the interval 9–12 ˚C) (unweighted chosen in final set). - Estimated standardized impacts (baseline OLS with country fixed effects; coefficients for first-differences standardized): - First difference in Harsh Drought Prevalence (W): coefficients reported −0.211***, −0.251***, −0.231***, −0.191***, −0.257*** with standard errors (0.0544), (0.0527), (0.0552), (0.0568), (0.0749). - Max T ˚C above 35 (W): coefficients −0.201***, −0.244***, −0.175**, −0.215***, −0.254*** with standard errors (0.0745), (0.0727), (0.0745), (0.0752), (0.0929). - Mean T ˚C in [9; 12): coefficients 0.155***, 0.174***, 0.153***, 0.142*** with standard errors (0.0395), (0.0402), (0.0380), (0.0400). - Interpretable shock magnitudes (baseline column A, standardized one-standard-deviation shocks): - One standard deviation positive shock in the first difference of Harsh Drought Prevalence reduces GDP growth by 0.21 percentage points. - One standard deviation positive shock in the first difference of Max T ˚C above 35 reduces GDP growth by 0.20 percentage points. - One standard deviation positive shock in the first difference of Mean T ˚C in [9; 12) increases GDP growth by 0.16 percentage points (reported text: 0.16 percentage points; table entries show 0.155–0.179 depending on column). - Relative explanatory power vs. prior literature: - The selected climate variables perform much better than annual average temperature and precipitation specifications (e.g., Burke et al., 2015; Kahn et al., 2021). - Adding selected climate variables can double or triple the increase in within R-square relative to specifications using averages; e.g., adding four selected climate variables increased within R-square by 102 percent in one comparison. - Nevertheless, total variation in GDP per capita attributable to weather is small: selection can at most increase the within R-square by a few percentage points. - Dynamics and persistence (local projections up to horizon 7): - One-standard-deviation shocks in the selected variables produce persistent effects on GDP levels that remain broadly significant over the 7-year horizon. - Harsh Drought Prevalence and Max T ˚C above 35 have adverse effects on GDP levels that are stable and significant over most horizons. - Mean T ˚C in [9; 12) has a positive effect that remains stable and significant over the 7-year horizon. - Effects are persistent on levels but there is no evidence of persistent effects on growth rates (no growing absolute coefficient with horizon). - Findings support using climate variables in first differences rather than levels. ### Heterogeneity and robustness - Heterogeneity across country groups and regions: - Effects larger in countries more oriented to agriculture and in poorer countries. - Harsh Drought Prevalence harmful in almost all groups; particularly strong negative impacts in MENA and SSA; significant in ECA; not significant in LAC and EAP. - Max T ˚C above 35 harmful and significant in most groups; especially harmful in SSA and EAP, and in agricultural hot and poor countries; effects not significant in cold countries where such temperatures are rare. - Mean T ˚C in [9; 12) beneficial especially in agricultural cold countries and in ECA region; not significant in hot, agricultural hot, and poor countries. - Robustness checks: - Entering each selected variable separately yields similar coefficients. - Alternative specifications (world GDP growth instead of year effects; country-specific quadratic trends; balanced subsample of 135 countries for 1983–2019) retain Harsh Drought Prevalence and Max T ˚C above 35 as robust influencers; Mean T ˚C in [9; 12) sometimes replaced in balanced sample by cold-spell or precipitation measures but Harsh Drought and Max T>35 remain robust. - Alternative climate definitions and LASSO tuning produce broadly similar substantive conclusions. ### Macro-fiscal findings - LASSO for fiscal dependent variables selected additional parsimonious climate variables (in addition to the three GDP-selected variables): - Length of the longest dry spell (LLDS, specifically LLDS.80 in the baseline fiscal selection). - Wetness intensity (MPDSI > 3). - Total precipitation in the longest period of continued intense precipitation (lagged Continued Heavy Precipitation, PC95WD). - Empirical patterns (estimates from Table 3; standardized climate variables with standard deviations equal to 100; controls include first two lags of dependent variable, year and country fixed effects; standard errors clustered by country): - General: - Within R-square for fiscal models is at least as large as for GDP growth models. - Estimated magnitudes of climate effects on fiscal variables are somewhat similar to GDP effects but often noisier. - Specific findings (selected coefficients and standard errors reported exactly as in source): - Lag(1) coefficients on dependent variables: - Column (A) ∆ ln GDP POP p.c.: 0.169*** (0.0562) - Column (B) ∆ Revenue GDP p.p.: -0.256*** (0.0494) - Column (C) ∆ Expenditure GDP p.p.: -0.250*** (0.0502) - Column (D) ∆ Balance GDP p.p.: -0.310*** (0.0315) - Column (E) ∆ Debt GDP p.p.: 0.074** (0.0292) - Column (F) ∆ ln Revenue p.c.: -0.199*** (0.0534) - Column (G) ∆ ln Expenditure p.c.: -0.102*** (0.0340) - Column (H) ∆ ln Debt p.c.: 0.136*** (0.0318) - Harsh Drought Prevalence (first difference, W): - Column (A): -0.146** (0.0661) - Column (C): 0.131** (0.0525) - Column (G): 0.327* (0.1891) - Max T ˚C above 35 (W): - Column (B): -0.150** (0.0636) - Column (D): -0.133* (0.0745) - Column (F): -1.094** (0.4881) - Mean T ˚C in [9; 12): - Column (A): 0.146*** (0.0455) - Column (D): 0.107** (0.0516) - Column (F): 0.247** (0.1207) - Longest Dry Spell (.80) (W): - Column (E): 0.343* (0.1974) - Lag(1) of Continued Heavy PPT: - Column (G): 0.522* (0.2655) - Wetness Intensity: - Column (C): 0.128** (0.0596) - Column (E): 0.358* (0.1949) - Column (G): 0.387* (0.1995) - Column (H): 0.829*** (0.2748) - Interpretation of macro-fiscal patterns: - Droughts (Harsh Drought Prevalence) tend to increase expenditure-to-GDP ratio (counter-cyclical spending) and reduce GDP concurrently. - High extreme heat (Max T>35) associated with declines in revenue levels and negative effects on revenue-to-GDP and balance-to-GDP (pro-cyclical revenue declines). - Mean T in [9; 12) raises revenue and GDP roughly proportionally; no significant expenditure response → positive effect on fiscal balance. - Wetness intensity and prolonged heavy precipitation associated with increased expenditure and increased debt and debt-to-GDP ratio (counter-cyclical fiscal support following flood-like conditions). - Longest Dry Spell associated with a sizeable increase in debt-to-GDP (0.343*), suggesting borrowing to finance responses. - Interaction with fiscal space: - Fiscal space proxied by standardized deviation of debt-to-GDP from country mean; evidence that limited fiscal space constrains counter-cyclical responses for Continued Heavy Precipitation: countries with debt-to-GDP above their average show muted increase in debt but experience a significant reduction in GDP per capita (Table A.17). - Robustness: - Alternative definitions (LLDS.95, PDSI ≥4, averaging annual conditions) can yield significant negative GDP effects and retain fiscal significance (Table A.16). - LASSO selected variables for each fiscal outcome separately; close alternative climate measures sometimes better explain other fiscal outcomes. ### Synthesis of substantive conclusions - A small number of infra-annual, high-frequency, locally measured weather variables (drought prevalence, extreme high-temperature days, and mid-range mild-temperature frequency) explain GDP growth variations better than simple annual averages of temperature and precipitation. - One-standard-deviation shocks in Harsh Drought Prevalence and Max T above 35 reduce GDP growth by roughly 0.20–0.21 percentage points; a one-standard-deviation increase in Mean T in [9; 12) raises GDP growth by about 0.16 percentage points. - Selected climate shocks have persistent effects on GDP levels over at least a 7-year horizon but no evidence they change growth trends persistently. - Fiscal policy responses are heterogeneous and often counter-cyclical: governments raise expenditure and debt in response to droughts and intense precipitation, and revenue responses are often pro-cyclical (especially to extreme heat). - Limited fiscal space can constrain effective counter-cyclical fiscal responses, amplifying adverse GDP impacts for some shocks. - Although selected variables improve explanatory power relative to annual averages, the overall share of GDP variation explained by weather remains small on average. ### Policy implications and recommendations - Prioritize monitoring and policy attention to the specific high-frequency, local weather measures identified as relevant: drought measures (PDSI < −4), extreme high-temperature days (TX35), prolonged heavy precipitation and wetness intensity, and Longest Dry Spell metrics (e.g., LLDS.80). - Design fiscal frameworks to preserve fiscal space and rapid financing mechanisms so governments can deploy counter-cyclical spending in response to droughts and flood-like events. - Recognize that fiscal policy can mitigate GDP impacts from adverse weather, but effectiveness depends on shock characteristics and pre-existing fiscal space constraints. - Project the selected high-frequency, high-resolution weather variables under climate scenarios with comparable spatial and temporal granularity to estimate short-term macro-fiscal impacts of climate change (avoid extrapolating short-term random-shock identification directly to long-term slow climate change without caution). *Source: wpiea2022156-print-pdf (content as provided).* ### 1. Introduction ### 1. Introduction ### Context and motivation - Climate is changing and is expected to continue changing in the forthcoming decades even with sharp reductions in greenhouse gas emissions (IPCC, 2021b). - Existing macroeconomic literature on weather and climate splits into two strands: - Reduced-form damage functions linking global mean temperature to total output losses. - Econometric analyses directly estimating impacts of random changes in temperature (and sometimes precipitation) on GDP per capita or Total Factor Productivity. - Most prior work focuses on annual average weather and GDP, while public debate emphasizes extreme weather events. ### Contributions of this study - Leverages a global dataset of daily temperature and precipitation with 30-Km resolution from ERA5 to construct a rich set of infra-annual and local weather variables. - Uses the Least Absolute Shrinkage and Selection Operation (LASSO) to select weather variables that best explain macroeconomic and fiscal outcomes from a large candidate set. - Extends analysis beyond GDP to important fiscal aggregates (government revenue, expenditure, debt) to study composition and cyclicality of fiscal responses to weather shocks. ### Data and constructed variables - Dataset: dozens of billions of daily temperature and precipitation measurements on a global grid with a 30-Km resolution from the ERA5 dataset. - Constructed variables: 164 weather variables including center and tails of temperature and precipitation distributions, heat and cold waves, droughts, intense precipitation. - Population-weighted spatial aggregations considered as alternatives. - Merged dataset coverage: 199 countries annually over the 1979-2019 period. ### Empirical specification (overview) - Baseline approach: regress first differences of macroeconomic variables on first differences of selected weather variables, including country and year fixed effects, and lags to allow for rich dynamics and autocorrelation control. - Use of country-specific polynomial trends in levels specification; estimation carried out in first differences to address serial correlation and non-stationarity of country GDP levels. - Local projection method (following Jorda, 2005) used to estimate impulse response functions up to horizon 7. ### Variable selection: LASSO and model selection strategy - Reason for machine learning: 164 variables would create over-fitting and multicollinearity issues in standard regressions, especially with multiple lags. - LASSO formulation: minimizes sum of squared errors plus λ times the L1 norm of coefficients. Hyper-parameter λ selected via a two-stage process. - Stage 1: Random search k-fold cross validation to maximize out-of-sample R-square. - Procedure details: panel randomly divided into five equal sets (five-fold cross validation). For each randomly drawn λ from a half normal distribution (variance 0.05) apply LASSO on training sets and compute R-square on test sets. - Repeated using 200 randomly selected values of λ for a total of 1,000 R-squared values. Select λ that gives the largest average R-square. - This random search tends to select a very small λ keeping about 20 variables or more. - Stage 2: Grid search — increase λ in small increments until no variable is dropped for three consecutive increments to refine to a parsimonious set. AIC and BIC used to assess selected models. - Implementation detail: follow Belloni et al. (2014) partial-out approach by regressing dependent and independent variables on fixed effects/trends, then applying LASSO to residuals. - Final estimation: coefficients from LASSO-selected variables re-estimated by standard OLS to obtain unbiased effect estimates. ### Key empirical findings (GDP per capita) - A handful of weather variables significantly impact GDP per capita; selected variables include measures capturing droughts and very high temperatures. - An increase in occurrence of such adverse weather shocks is estimated to have a detrimental effect on GDP; conversely, increases in mild temperatures have beneficial effects. - Impulse response analysis (local projections) finds shocks have permanent effects on the level of GDP per capita: - A shock of one standard deviation in the selected variables leads to impacts of around 0.2 percentage points of GDP that appear constant over time. - Evidence found for persistent impact on GDP levels but no evidence of a persistent effect on growth rates. - Relative explanatory power: - The selected climate variables perform much better in explaining GDP variations than average temperature and precipitation measures used in prior literature (e.g., Burke et al., 2015; Kahn et al., 2021). - Adding the selected climate variables can double or triple the increase in the within R-square relative to specifications using averages. - Nevertheless, the total variation in GDP per capita attributable to weather is small: selection can at most increase the within R-square by a few percents, suggesting weather is not the main driver of GDP variations globally on average. ### Heterogeneity and robustness - Results robust to alternative specifications and controls. - Heterogeneity across country groups: - Effects of weather shocks larger in countries more oriented towards agriculture and in poorer countries. - Positive effects of more frequent mild temperatures particularly relevant for agricultural and cold countries, possibly reflecting fewer freezing days. - Regional patterns: sub-Saharan Africa comparatively most affected by high temperatures; Middle-East and Northern Africa most affected by droughts. ### Macro-fiscal analysis and findings - LASSO applied to fiscal dependent variables selects three parsimonious variables (in addition to the three variables selected for GDP): - Length of the longest dry spell. - Wetness intensity. - Total precipitation in the longest period of continued intense precipitation (last two typically associated with flood-like conditions). - Macro-fiscal responses: - Weak evidence that these precipitation-related variables have a negative impact on GDP. - Counter-cyclical and often significant increase in government spending and debt in response to these shocks — fiscal responses that support the economy and may explain why negative GDP effects are not always significant. - Revenue responses are mostly significant to high temperatures and are pro-cyclical. - Overall conclusion: fiscal responses to weather shocks are rich, complex, and dependent on shock characteristics and country context. ### Identification, interpretation, and limitations - Identification relies on annual and infra-annual weather shocks treated as exogenous to annual economic shocks; authors argue reverse causality is not a concern for these short-term weather shocks. - Potential biases: - Measurement error would induce attenuation bias. - Omitted variable bias remains a concern but authors do not identify variables likely to simultaneously cause weather shocks and macro-fiscal impacts. - Limitations for extrapolation: - Identification is based on random weather shocks and is not exactly equivalent to long-term climate change responses. - Short-term responses to transitory 1 °C changes are not necessarily equal to responses to slow permanent 1 °C increases; extrapolating long-term impacts from short-term responses requires caution and is not attempted. - Additional numeric notes from methods and data: - Temperature trends cited range from 0.07 to 0.6 ̋C per decade across countries. - Standard deviation of average annual temperature usually about 0.5 ̋C; anticipated warming even with strong mitigation could be in the range of +2 to +4 ̋C in many countries. - Local projection horizons considered up to horizon 7. ### Organization of the paper - Subsequent sections: empirical specification and selection algorithm; construction and summary of weather variables; results for GDP per capita; results for fiscal variables; conclusion. *Source: wpiea2022156-print-pdf - 1. Introduction* ### 3. Data ### wpiea2022156-print-pdf - 3. Data ### Weather data sources and aggregation over time and space - Weather dataset: - Temperature and precipitation data from the ERA5 dataset compiled by the European Centre for Medium-Range Weather Forecasts (ECMWF) (Hersbach et al., 2018). - ERA5 provides hourly reanalysis weather data on a global grid from 1979 to 2019. - Grid resolution varies with latitude with cells of 30ˆ30 km at most (at the equator). - Processing and aggregation: - Raw ERA5 weather data aggregated over space and time using Google Earth Engine (GEE) (Gorelick et al., 2017). - Aggregation purpose: reduce millions of weather measurements per country-year into a manageable set of potentially meaningful climate variables while preserving local and infra-year extremes. - Rationale for high spatial/temporal resolution vs country-year averages: - Averages can miss local and infra-year extreme events when opposite events cancel out (e.g., drought in one region and intense precipitation in another). - Relationship between weather shocks and the economy can be non-linear and dependent on duration, spatial coverage, and intensity (e.g., prolonged high temperatures vs short-lived very high temperatures). - Deviations from local seasons can be relevant even if not reflected in annual averages (e.g., unusually high summer precipitation in central Europe). - Population weighting: - Both unweighted and population-weighted country-level aggregates are constructed. - Population weights use year 2000 population weights. - Grid-cell population data from Socioeconomic Data and Application Center’s UN WPP-Adjusted Population count dataset. - Both weighted and unweighted variables included in the LASSO selection to avoid biasing characterization of climate toward either low-population or high-population areas. ### Variable definitions - General approach: - Construct variables describing distributions and extremes for temperature, precipitation, and wetness/drought following climate literature (Kim et al., 2020; Perkins and Alexander, 2013). - Include both relative and absolute threshold definitions of extremes. - Count occurrences (share of grid-cells and days) and measure intensity (average or maximum intensity) for extreme events. - Include first and second lags and population-weighted counterparts in empirical estimation. - Temperature variables: - Consider average temperature, variance of daily temperature, and average diurnal temperature range (difference between minimum and maximum temperature in a day). - Count cold nights, cold days, warm nights, warm days using relative thresholds based on the 1979-2019 distributions for every 5-day window centered on each day of the year. - Heatwave and coldwave variables follow Perkins and Alexander (2013): various thresholds for daytime and nighttime; measure length of longest wave, number of waves per year, number of days, and average maximum/minimum temperatures during waves. - Additional duration definitions for cold and warm spells following Kim et al. (2020). - Compute annual minimum of night temperatures and maximum of daytime temperatures per grid cell, then average over space. - Absolute thresholds: count instances where daily temperature crosses thresholds (e.g., below 0 ̋ C, above 35 ̋ C and 40 ̋ C) over 365 days and across grid cells. - Define 3 ̋ C-wide temperature intervals from -9 ̋ C and below to 30 ̋ C and above; count frequency in these intervals to capture non-linear temperature effects. - Precipitation variables: - Precipitation includes rain and snow converted to rain equivalents. - Define “wet days” as days with 1 mm precipitation or more; “dry days” with precipitation below 1 mm. - Construct country-year averages and variance of daily precipitation twice: on all calendar days and on wet days. - Measure precipitation on very wet and extremely wet days using relative thresholds. - Capture extended wet/dry periods: largest number of consecutive dry days, wet days, very wet days, extremely wet days; measure precipitation in longest wet spells. - Flood-related measures: maximum annual rainfall over 1-day and 5-day intervals. - Local extreme precipitation: total monthly precipitation per grid cell; country average of maximum and minimum monthly precipitation. - Absolute thresholds: number of consecutive days with a minimum percentage of country area experiencing a dry day using different percentage thresholds. - Four precipitation intervals divided by 1, 10, and 20 mm thresholds; measure frequency and maximum spatial extent where precipitation exceeds 10 mm and 20 mm. - Construct indicator of deviation from balanced precipitation: absolute deviation from having precipitation between 1 and 10 mm half the time over space and time. - Wetness and drought variables: - Use Palmer Drought Severity Index (PDSI) to combine temperature and precipitation into cumulative soil moisture deviations. - PDSI ranges from -10 to +10; values below -4 and above +4 are very rare. - Build variables measuring share of total grid-months subject to droughts (PDSI < -3) and harsh droughts (PDSI < -4), and periods with high (PDSI > 3) and very high moisture (PDSI > 4). - For each category compute maximum geographical extent (share of affected grid-cells in the month where share is maximal). - Variable counts and final dataset used in estimation: - 45 different temperature variables. - 29 precipitation variables. - 8 wetness-drought variables. - Total of 82 unique climate variables. - After removing perfectly collinear variables and adding first and second lags plus population-weighted counterparts, obtain a set of 480 climate variables. - Macro-fiscal variables: - GDP per capita from World Bank WDI: variable “GDP per Capita constant 2015 US$” (NY.GDP.PCAP.KD). - Fiscal variables from IMF World Economic Outlook (WEO): government revenue and expenditure as percentage of GDP (GGRNGDP and GGXNGDP for general government). - Gross debt from Global Debt Database: “Central government debt, % of GDP”. - Note: for some countries fiscal variables cover only central government. ### Summary statistics - Sample selection and sizes: - Countries with at least 7-year time series for GDP per capita and climate variables; exclude Libya, Iraq, Equatorial Guinea and Bahrain because of outliers. - Result: 199 countries for a total of 6,550 country-year observations for GDP analysis. - Fiscal sample: remove Kuwait and S ̃ao Tom ́e and Pr ́ıncipe and all observations with any missing fiscal value; result: 159 countries for a total of 3,859 country-year observations for fiscal analysis. - Variance decomposition approach: - Empirical approach relies on inter-annual variation within country; decompose variance into between- and within-country components and use within variation for identification. - Key macroeconomic summary statistics: - GDP per capita growth: - Grew by 1.6 percent per year on average in the largest sample used for GDP analysis. - Grew by 2.0 percent per year on average in the smaller fiscal-policy sample. - Within standard deviation of GDP per capita growth: 4.7 percentage points (larger sample) and 3.7 percentage points (smaller sample). - Government revenue and expenditure ratios to GDP: - Grew at the same average rate of 0.1 percentage points per year. - Inter-annual variation with within standard deviations ranging from 3.2 to 4.0 percentage points. - Government debt-to-GDP: - Ratios stable on average but with large within standard deviation of 11.1 percentage points. - Climate variable variation and trends: - The within-country standard deviation of climate variables in first differences is typically much higher than the between-country standard deviation. - Many weather variables exhibit trends over 1979–2019; positive and statistically significant country-specific trends found in a majority of countries for temperature-related variables; trends less common for other variables. - Model specification assumes time-invariant trends; structural break tests with unknown break dates generally do not reject no-break null for most variables in most countries, supporting the time-invariant trends assumption. - Unit-root tests on first differences of weather variables reject unit roots in all cases with p-values close to zero. *Source: wpiea2022156-print-pdf - 3. Data* ### 4. GDP Results ### 4. GDP Results ### 4.1. Climate variable selection - Baseline specification uses country and year fixed effects; collinearity test of Belloni et al. (2013) does not suggest dropping any of the 480 weather variables. - Random search process selects 30 variables with λ = 0.0166; selected variables include the first two lags of GDP per capita growth and a mix of temperature and precipitation variables (unweighted and weighted, contemporaneous and lagged). The full list is reported in the appendix Table A.7. - First phase reduces total variables by approximately 95 percent, but many selected climate variables are statistically insignificant, prompting further reduction via the grid search algorithm (Section 2.3). - Final selection includes three climate variables plus the first two lags of the dependent variable. - Within R-square outcomes: - Preferred specification improves the within R-square by 7 percent relative to a specification with just the 2 lags of the dependent variable. - Specification using all 30 variables selected by the LASSO after the random search yielded a 17 percent improvement in within R-square. - The within R-square reported baseline reference is within R2 = ".0937" (note: reported as the within R-square obtained with a specification that only includes the first two lags of the dependent variable). - Information criteria: - Bayesian information criterion (BIC) is minimized for selections of 4 to 6 variables (Figure 2 panel b). - Akaike information criterion (AIC) declines monotonously as the selection of variables increases up to the set chosen by the random search. - Preferred selection of 5 variables is supported by these criteria. - Hyperparameter tuning note: λ increased by increments equal to 0.001 and the smallest λ is selected such that variable selection is stable for at least three consecutive increments (+0.003). ### 4.2. The effect of weather variables on GDP growth - Three selected weather variables in the baseline specification: - Harsh Drought Prevalence — PDSI < −4 (share of grid-days with harsh drought conditions) - Max T̋C above 35 — TX35 (share of grid-days with maximum daily temperature above 35 ̋C) - Mean T̋C in [9; 12) — TS9,12 (share of grid-days with mean temperature in the interval 9-12 ̋C) - LASSO selection characteristics: - LASSO selects population-weighted variables except for Mean T̋C in [9; 12). - Correlation and summary statistics: - Correlation between first differences of selected climate variables and GDP growth ranges from 0.04 to 0.05 in absolute value. - Correlation between first differences of average annual temperature (Mean Temperature — T) and GDP growth is −0.006. - Largest correlation among climate variables: Harsh Drought Prevalence and Max T̋C above 35 = 0.183. - Harsh Drought Prevalence correlated with Mean Temperature = 0.159. - Max T̋C above 35 and Mean Temperature correlation = 0.356. - Mean Temperature is not retained by the LASSO. - Standardized-variable OLS findings (Table 1 summarized): - Increases in Harsh Drought Prevalence and in Max T̋C above 35 have adverse effects on contemporaneous GDP growth. - Increases in Mean T̋C in [9; 12) have beneficial effects on GDP growth. - These effects are highly significant and robust across alternative model specifications. - Baseline specification (column A: unbalanced panel and year fixed effects) quantified impacts: - A positive shock equal to one standard deviation of the first difference of Harsh Drought Prevalence in a country reduces GDP growth by 0.21 percentage points. - A positive shock equal to one standard deviation in the first difference of Max T̋C above 35 reduces GDP growth by 0.20 percentage points. - A positive shock equal to one standard deviation in Mean T̋C in [9; 12) leads to an increase in the growth rate equal to *Source: wpiea2022156-print-pdf - 4. GDP Results* ### 0.16 percentage points. In this case, the unweighted variable was chosen over the population-weighted one. ### wpiea2022156-print-pdf - 0.16 percentage points. In this case, the unweighted variable was chosen over the population-weighted one. ### Selected climate variables and estimated effects on GDP per capita growth - Variables selected and their estimated coefficients from OLS with country fixed effects (Table 1): - First difference in Harsh Drought Prevalence (W): - Coefficients: -0.211***, -0.251***, -0.231***, -0.191***, -0.257*** - Standard errors: (0.0544), (0.0527), (0.0552), (0.0568), (0.0749) - Max T ˚C above 35 (W): - Coefficients: -0.201***, -0.244***, -0.175**, -0.215***, -0.254*** - Standard errors: (0.0745), (0.0727), (0.0745), (0.0752), (0.0929) - Mean T ˚C in [9; 12): - Coefficients: 0.155***, 0.174***, 0.153***, 0.142*** - Standard errors: (0.0395), (0.0402), (0.0380), (0.0400) - Sample and fit statistics (Table 1): - Observations: 6,550 (columns A-F), 6,504,860 (column G?) — Table shows "6,5506,5506,5506,5506,5506,5504,860" as presented in source. - Year fixed effects: Yes (columns A-D, F-G), No (column E) - World GDP growth: included only in Column E - Country quadratic trends: included in Column F - Balanced sample: Column G is estimated on a balanced subsample of 135 countries for 1983-2019 - R-square: 0.264, 0.261, 0.261, 0.260, 0.253, 0.350, 0.224 (as reported) - Within R-square: 0.0998, 0.0967, 0.0965, 0.0952, 0.149, 0.0375, 0.0790 (as reported) - Notes on variables and estimation: - Results based on first differences of log real GDP per capita expressed in constant 2015 USD. - All regressions include controls such as two lags of the dependent variables. - (W) indicates population-weighted variables. - Standard errors are clustered by country. - Significance: *pă0.10, **pă0.05, ***pă0.01 ### Robustness and alternative specifications - Individual-variable entry: - Entering each selected climate variable separately (columns B–D) yields coefficient estimates very similar to baseline (column A), indicating main effects are largely independent. - LASSO-based selection robustness: - Column E (world GDP per capita growth instead of year effects): - Initial random search selects 36 variables; 26 variables common with baseline; only 4 of 30 baseline variables not re-selected. - Narrowed to 8 variables (including two lags of dependent variable and world growth); five climate variables include the same 3 as baseline plus duration of cold spells and lag of spatial average of grid-cell maximum temperatures. - Coefficients of the 3 baseline climate variables remain robust. - Specification with country-specific quadratic trends: - Random search selects 24 variables; 22 common with baseline. - After increasing penalty, stable set of 7 variables includes the same 3 climate variables and 2 lags of dependent variable. - Coefficient estimates remain very similar (Column F). - Balanced subsample (135 countries, 1984–2018): - Random search selects 18 variables, including the 3 baseline variables. - Grid search loses Mean T ˚C in [9; 12) and adds first lag of Cold Spell Duration and 1-Day PPT Maximum (PXp1). - Effects of Harsh Drought Prevalence and Max T ˚C above 35 remain robust (Column G). ### Comparisons with the literature - Burke et al. (2015) specification (Table 2): - Regress GDP growth on annual average temperature, precipitation, their squares, two lags of GDP growth, country quadratic trends and year effects. - Estimated optimal temperature: 13.3 ˚C (source finds 13.3 ˚C vs. Burke et al.’s 13.1 ˚C). - Relative to no-climate-variable baseline (column A), introducing annual average temperature and precipitation and their squares improves within R-square by 10 percent. - Adding the four selected climate variables increases within R-square by 102 percent (column C). - AIC and BIC show unambiguous improvement with the selected variables; annual averages and squares are not clearly supported by information criteria. - Kahn et al. (2021) specification: - Baseline focuses on absolute deviations of annual average temperatures and precipitations relative to 30-year moving averages; includes four lags for each variable. - Introducing Kahn et al. climate variables improves within R-square by 4 percent (columns D–E). - Additionally including the selected variables improves within R-square by 12 percent. - Information criteria again favor the selected variables over annual averages and deviations. - Overall conclusion: - A range of performance indicators suggests the selected climate variables explain GDP growth variations better than annual average temperature and precipitation measures. ### Dynamic effects (persistence) - Method: local projection method to estimate impulse responses to a one standard deviation shock in each climate variable over a 7-year horizon. - Main findings (Figure 3): - All selected climate shocks have very persistent effects on GDP levels that remain broadly significant over the 7-year horizon. - Mean T ˚C in [9; 12): - One-year increase has a positive effect that remains stable and significant throughout the 7-year period (Panel (c)). - Harsh Drought Prevalence and Max T ˚C above 35: - Both have adverse effects that are stable and remain significant over most horizons, albeit with slightly more uncertainty (Panels (a) and (b)). - All effects show persistent impacts on GDP levels but no effect on GDP growth (i.e., no increasing absolute coefficient magnitude with horizon). - The findings support introducing climate variables in first differences rather than levels. ### Heterogeneity across country groups and regions - Grouping strategy: - Countries separated into rich vs. poor, hot vs. cold, agricultural vs. non-agricultural, agricultural cold vs. agricultural hot, and six macro-regions. - Baseline LASSO-selected variables used for all subgroup estimations. - Distributional notes: - Use of absolute thresholds (e.g., 35 ˚C) yields different frequencies by local climate; occurrence and standard deviation of Max T ˚C above 35 vary across countries. - Mean T ˚C in [9; 12) standard deviation larger at higher latitudes, particularly Europe. - Harsh Drought Prevalence measured via PDSI reflects relative dryness/wetness, thus spatial distribution follows regional precipitation and temperature variation. - Subgroup findings (Figure 5): - General: - Main results are generally confirmed across groups; significant coefficients tend to remain significant, with no significant sign changes. - Coefficients are almost never significantly different between groups. - Harsh Drought Prevalence: - Significantly harmful in almost all groups. - Harmful in both agricultural and non-agricultural countries, and in rich and poor countries. - Strong negative impacts in MENA and SSA; significant also in ECA; not significant in LAC and EAP. - Max T ˚C above 35: - Harmful and significant in most groups, except cold countries and agricultural cold countries where such temperatures are rare. - SSA and EAP are regions with significant harmful effects. - Particularly harmful in agricultural hot countries and in poor countries—suggestive of agricultural vulnerability to extreme heat. - Mean T ˚C in [9; 12): - Not significant in hot, agricultural and hot, and poor countries (low frequency of these temperatures). - Positive impact is particularly large in agricultural cold countries—consistent with a beneficial effect of longer growing seasons in cold agricultural settings. - Among regions, positive and significant only in EAC region (many cold agricultural countries). - Group definitions (as provided in Figure 5 notes): - Hot (N=3,472): 1979-2019 average temperature ≥ 22 ˚C. - Cold (N=3,078): 1979-2019 average temperature < 22 ˚C. - Agricultural (N=3,130): share of “Agriculture, forestry, and fishing, value added (% of GDP)” in 2002 is above median. - Non Agricultural (N=3,062): not Agricultural. - Agricultural Cold (N=1,124): agricultural and cold. - Agricultural Hot (N=2,006): agricultural and hot. - Rich (N=3,823): “High Income” and “Upper Middle Income” in WDI. - Poor (N=2,727): “Low Income” and “Lower Middle Income” in WDI. - SSA (N=1,629): Sub-Saharan Africa. - MENA (N=538): Middle-East and North Africa. - LAC (N=1,372): Latin America and the Caribbean. - ECA (N=1,638): Eastern and Central Asia. - EAP (N=1,013): Eastern Asia and Pacific. - Note: None of the coefficients is significant for North America and South Asia due to low number of countries in these regions. *Italic: Content derived from wpiea2022156-print-pdf (source PDF content as provided).* ### 5. Macro-Fiscal Outcomes ### 5. Macro-Fiscal Outcomes ### 5.1. A systematic empirical approach to macro-fiscal impacts - Channels through which weather shocks affect fiscal outcomes: - Automatic stabilizers: climate effects on GDP alter government finances when fiscal variables are measured as GDP ratios. - Operational disruption: extreme events (e.g., floods) can prevent government workers from working or impair public infrastructure. - Active fiscal responses: governments may provide support, foster recovery, or address fiscal sustainability issues. - Main fiscal indicators analyzed: - Government revenue (measured as GDP ratio and percent change in constant 2011 USD). - Government expenditure (measured as GDP ratio and percent change in constant 2011 USD). - Government debt (measured as GDP ratio and percent change in constant 2011 USD). - Fiscal balance (revenue minus expenditure, measured as GDP ratio). - Methodological note: - Examine both ratios and percent changes; also analyze first differences of the logarithm of the three main fiscal variables to separately identify numerator and denominator effects. - Expected directional effects and mechanisms: - Revenue-to-GDP ratio: - Can increase if shocks disproportionately reduce activity in low-taxed sectors (tax losses smaller than GDP losses). - Can decrease if active policy reduces taxes in vulnerable sectors or if revenue falls more than GDP. - Expenditure-to-GDP ratio: - Expected to increase under counter-cyclical fiscal policy (numerator rises, denominator falls). - Examples: increased social transfers, unemployment insurance, ad-hoc compensation, capital expenditure for reconstruction. - May decrease after severe shocks in low-capacity governments due to disorganization and delayed expenditure. - Debt: - Expected to increase when shocks worsen the fiscal balance (higher deficits) and create financing needs. - Timing issues: borrowing can precede expenditure implementation. - Other channels: government loans or bail-outs increase debt without affecting fiscal balance; exchange rate effects can alter valuation of foreign debt. ### 5.2. Estimates of the impact of climate shocks on macro-fiscal outcomes - Estimation approach: - Empirical OLS estimates reported in Table 3 using a compact selection of climate variables: the three climate variables from the GDP baseline and the first weather variable selected by the LASSO for each fiscal ratio. - All climate variables are standardized with standard deviations equal to 100. (W) indicates population-weighted variables. - Controls include the first two lags of the dependent variable, and year and country fixed effects. Standard errors clustered by country. - Sample restricted to observations with non-missing values for all fiscal variables; this reduces sample size relative to GDP-only analysis and affects some coefficients' significance. - Climate variables selected by LASSO for fiscal outcomes: - Revenue-to-GDP ratio: contemporaneous population-weighted Longest Dry Spell (.80) — LLDS.80 (uninterrupted period where at least 80% of country’s grid cells have precipitation < 1 mm/day). - Expenditure-to-GDP ratio: first lag of Continued Heavy Precipitation (PC95WD) — largest number of consecutive days with precipitation above the 95th percentile. - Debt-to-GDP ratio: Wetness Intensity (MPDSI>3) — maximum monthly share of grid cells with moisture conditions far exceeding typical level, proxying flood-like conditions. - General empirical patterns: - Within R-square: fraction of within variation in fiscal variables explained by the model is at least as large as for GDP growth. - Magnitudes of estimated effects of new climate variables are somewhat similar to effects on GDP reported earlier, but often not statistically significant. - Evidence of non-neutral fiscal responses varies by shock type: - Harsh Drought Prevalence: - Increases expenditure-to-GDP ratio. - Appears driven by combined increase in expenditure levels and decline in GDP. - Droughts do not significantly affect other fiscal variables. - Max T̈C above 35: - Associated with a decline in revenue collection due to a sharp reduction in the level of revenue. - Affects the revenue-to-GDP ratio and yields a weakly significant negative effect on the balance-to-GDP ratio. - Mean T̈C in [9; 12): - Raises revenue significantly and raises GDP in roughly the same proportion. - Leads to a positive but insignificant impact on revenue-to-GDP ratio. - In absence of a significant expenditure response, indicates an unambiguous positive effect on the fiscal balance. - Wetness Intensity (proxy for flood-like conditions): - Leads to a significant increase in public expenditure (column G) and in the expenditure-to-GDP ratio (column C). - Debt and debt-to-GDP ratio significantly increase (columns E and H). - Has a negative but insignificant effect on GDP, suggesting a counter-cyclical fiscal policy response that may mitigate GDP impact. - Longest Dry Spell (0.80): - Associated with a sizeable but insignificant increase in the fiscal deficit (column D) and a significant increase in the debt-to-GDP ratio (column E). - Lag of Continued Heavy Precipitation: - Large and significant positive response of expenditure (column G); negative but insignificant impact on the fiscal balance. - Overall: systematic evidence of economic support from fiscal policy in response to likely adverse weather shocks, especially via increased public spending and debt, though many estimates are noisy. - Robustness and alternative definitions: - Using alternative definitions (e.g., Longest Dry Spell affecting 95 percent — LLDS.95, Very Wet Conditions — PDSI ≥4, or averaging annual conditions rather than peak month) can yield significant and negative GDP effects while retaining significance for fiscal outcomes (Table A.16). - LASSO selects variables relevant for one fiscal variable at a time; alternative close definitions may provide better explanatory power for other fiscal outcomes. - Interaction with fiscal space: - Fiscal space proxied by the standardized deviation of debt-to-GDP from its country mean; the cube of the deviation retained as informative for non-linearity. - Significant interaction found only for Continued Heavy Precipitation: countries with debt-to-GDP above their average show a muted increase in debt but experience a significant reduction in GDP per capita (Table A.17), supporting hypothesis that limited fiscal space constrains counter-cyclical responses. - Quantitative magnitudes and statistical highlights from Table 3 (selected coefficients and statistics): - Lag(1) of fiscal variable coefficients: - Column (A) ∆ ln GDP POP p.c.: 0.169*** (0.0562) - Column (B) ∆ Revenue GDP p.p.: -0.256*** (0.0494) - Column (C) ∆ Expenditure GDP p.p.: -0.250*** (0.0502) - Column (D) ∆ Balance GDP p.p.: -0.310*** (0.0315) - Column (E) ∆ Debt GDP p.p.: 0.074** (0.0292) - Column (F) ∆ ln Revenue p.c.: -0.199*** (0.0534) - Column (G) ∆ ln Expenditure p.c.: -0.102*** (0.0340) - Column (H) ∆ ln Debt p.c.: 0.136*** (0.0318) - First difference in Harsh Drought Prevalence (W): - Column (A): -0.146** (0.0661) - Column (C): 0.131** (0.0525) - Column (G): 0.327* (0.1891) - Max T̈C above 35 (W): - Column (B): -0.150** (0.0636) - Column (D): -0.133* (0.0745) - Column (F): -1.094** (0.4881) - Mean T̈C in [9; 12): - Column (A): 0.146*** (0.0455) - Column (D): 0.107** (0.0516) - Column (F): 0.247** (0.1207) - Longest Dry Spell (.80) (W): - Column (E): 0.343* (0.1974) - Lag(1) of Cont’d Heavy PPT: - Column (G): 0.522* (0.2655) - Wetness Intensity: - Column (C): 0.128** (0.0596) - Column (E): 0.358* (0.1949) - Column (G): 0.387* (0.1995) - Column (H): 0.829*** (0.2748) - Constant terms (selected): - Column (A) constant: 1.606*** (0.1134) - Column (F) constant: 5.041*** (0.2936) - Column (G) constant: 4.456*** (0.2411) - Column (H) constant: 3.458*** (0.1628) - Sample and fit: - Observations: 3849.000 for all columns reported. - R-square ranges from 0.078 to 0.267 across columns. - Within R-square ranges from 0.018 to 0.094 across columns. - Significance notation: * p < 0.10, ** p < 0.05, *** p < 0.01. - Synthesis of empirical conclusions: - Weather shocks have rich and sometimes large effects on fiscal aggregates, though many estimates lack statistical significance. - Revenue responses are pro-cyclical for some temperature shocks (Max T̈C above 35 reduces revenue; Mean T̈C in [9; 12) increases revenue proportionally to GDP). - Government expenditure and debt respond counter-cyclically to droughts and intense precipitation, mitigating GDP impacts. - Limited fiscal space can constrain counter-cyclical responses and amplify adverse GDP impacts for certain shocks. ### Key policy implications and recommendations (from empirical findings) - Decision-makers should pay better attention to the specific shocks identified as relevant (drought measures, extreme high temperatures, intense precipitation/wetness intensity) rather than relying solely on annual average temperature measures. - Ensure adequate capacity to address identified shocks, including by creating fiscal space to enable counter-cyclical fiscal responses. - Recognize that fiscal policy can mitigate GDP impacts of adverse weather shocks, but effectiveness depends on the nature of the shock and country fiscal space. - Consider projecting the selected high-frequency, high-resolution weather variables under climate scenarios with comparable spatial and temporal granularity to estimate short-term macro-fiscal impacts of climate change. *Source: 5. Macro-Fiscal Outcomes, wpiea2022156-print-pdf* ### References ### wpiea2022156-print-pdf - References ### Major data sources and datasets - Weather data: ERA5 dataset from 1979 to 2021. The original ERA5 dataset has hourly data but the authors use data aggregated at daily level by Google Earth Engine (GEE). - Palmer Drought Severity Index (PDSI): TerraClimate (Abatzoglou et al., 2018) accessed using GEE; monthly intervals with spatial resolution equal to 0.0416 degrees. - Population grid for weighted variables: resolution «1 kmˆ1 km at the equator. - Country projection and zonal statistics computed via Google Earth Engine. ### Dataset spatial and temporal resolution and observation counts - ERA5 spatial resolution: 0.25 degrees. Global grid cells = 1,036,800. - Approximate share of Earth’s surface covered by land (excluding Antarctica and Greenland): 27%. - Approximate land grid cells used from ERA5: 1,036,800ˆ0.27 = 279,936. - Time span used for daily ERA5-derived weather: 41 years (1979 to 2019) with four weather data points per grid-day (T, TN, TX, and P). - Resulting raw ERA5 temperature and precipitation data points: 279,936ˆ365ˆ41ˆ4 = 16,756,968,960 («17 billion). - PDSI spatial resolution 0.0416 degrees gives approximately 10,110,022 land cells (excluding Antarctica and Greenland). - PDSI observations from January 1979 to December 2019: 10,110,022ˆ12ˆ41 = 4,974,130,917 («5 billion). - Combined starting point: 21,715,195,392 («22 billion) data points on temperature, precipitation, and the PDSI. - Merged ERA5 and PDSI into one geospatial dataset using PDSI higher resolution, producing 9,621,976 («10 million) country-matched grid cells for each of five core climate variables (Tj,d, TNj,d, TXj,d, Pj,d, and PDSIj,d). - Full set of climate variables uses 580,705,495,552 («600 billion) data points (daily data for 41 years across country-matched grid cells). - Weighted variables using population gridded at «1 kmˆ1 km add 25 additional points per raw weather grid cell, adding 25ˆ580,705,495,552 = 14,517,637,388,800 («15 trillion) data points to the dataset for zonal statistics. ### Construction and aggregation procedures - Daily grid-cell temperature measures: daily mean (Tj,d), minimum (TNj,d), maximum (TXj,d). Total daily precipitation Pj,d as sum of hourly precipitation measures per day. - Country-level aggregation: country-wide averages of daily average, minimum, and maximum temperature (respectively Td, TNd, and TXd) computed from grid-level daily data; analogous averaging for precipitation Pd. - Merging approach: use higher resolution PDSI (~5 kmˆ5 km at the equator) for country projection and zonal statistics. ### Temperature variables and extreme-event constructions - Annual and distributional statistics: - Annual average temperature (T). - Variance of daily temperature (TVar). - Average diurnal temperature range (DTR) from daily TN and TX. - Country averages of grid-level annual minimum of minimum daily temperature (TNn) and grid-level maximum of maximum daily temperature (TXx). - Relative-threshold extreme indicators (using 1979–2019 distribution and small moving windows): - Cold nights (CN10), cold days (CD10), warm nights (WN90), warm days (WD90) using 10th and 90th percentiles of TNd and TXd in a 5-day window centered on each day. - Spell metrics (Kim et al., 2020; Perkins and Alexander, 2013): - Cold (CSD) and warm spell duration (WSD): number of days with TNd (TXd) below (above) the 10th (90th) percentile in a 5-day window for at least six consecutive days. - Eight additional indicators of day/night heat waves: exceedance of 90th percentile of TXd (TNd) in a 15-day window centered on each day for at least three consecutive days. For each: number of days with day (night) heat wave, length of the longest day (night) heatwave, number of day (night) heatwaves during a year, and average maximum (minimum) temperature during day (night) heatwaves. - Day and night cold wave analogues using the 10th percentile. - Absolute-threshold extreme indicators (grid-level sums, then normalized): - Frost: minimum daily temperature below 0 ̋C — TN0. - TX35: maximum temperature above 35 ̋C. - TX40: maximum temperature above 40 ̋C. - Temperature-response functional form: - Distribution divided into 3 ̋C-wide intervals; measure share of grid-day observations in each interval (Schlenker and Roberts, 2009). Extreme terminal intervals group days below -9 ̋C and at or above 30 ̋C. ### Precipitation variables and extreme-event constructions - Aggregate statistics: - Annual average precipitation (P) and annual variance of daily precipitation (PVar). - Focus on “wet days” (days with precipitation > 1 mm). - Number of wet days (W), average daily precipitation in wet days (PWA), and wet days precipitation variance (PWVar). - Total precipitation in very wet (PW95T) and extremely wet days (PW99T) using the 95th and 99th percentiles of wet-day distribution. - Spell and intensity metrics: - Largest number of consecutive dry days (CDD: precipitation < 1 mm) and largest number of consecutive wet days (CWD); total precipitation during longest wet-days period (PCWD). - Consecutive very wet (PC95WD) and extremely wet (C99WD) days in the longest periods with daily precipitation above the 95th and 99th percentiles; total precipitation in consecutive very (PC95W) and extremely wet days (PC99WD). - Annual maxima of rainfall: PX1 (maximum 1-day rainfall), PX5 (maximum 5-day rainfall). - Local extreme precipitation from monthly totals per grid cell: country average of maximum monthly precipitation (PX(p1Month)) and minimum monthly precipitation (PN(p1Month)). - Spatial-extent and absolute-threshold precipitation metrics (grid-level calculations, then aggregated): - Length of the longest dry spell (LLDS): uninterrupted series of days where a minimum percent of the country area has daily precipitation < 1 mm. Thresholds used: 50%, 65%, 80%, 95% of country area. - Share of grid-days with total precipitation in intervals: <1 mm; 1–10 mm; 10–20 mm; >20 mm. - Maximum extent of heavy (MaxP>10) and very heavy (MaxP>20) precipitation: maximum daily share of country with precipitation >10 mm and >20 mm. - Imbalance indicator BP110(0.5): absolute deviation from having 50% of grid-days with precipitation between 1 and 10 mm. ### Wetness and drought (PDSI) variables - PDSI ranges from -10 to +10; values below -4 and above +4 are rare. - Category indicators and extent measures: - Share of total grid-months with extreme drought (PDSIă ́4). - Share with extreme and severe drought (PDSIă ́3). - Share with extreme moisture (PDSIą4). - Share with very high and extreme moisture (PDSIą3). - For each category, also compute the maximum extent (share of affected grid-cells in the month of maximum share). ### Mathematical notation and methodological notes - Indexing and notation: - d denotes calendar days, m denotes months, j = 1,...,J denotes grid cells in every country. - For ease of notation, variables are not indexed by country and year in the mathematical exposition; each year is assumed to have the same number of days. - Iverson brackets are used to define indicator functions: - rXs = 1 if X is true; 0 otherwise. - Summations of Iverson brackets (e.g., ř d rXs) are used to count days meeting condition X. - Percentile and distributional references use the whole time horizon of the dataset from 1979 onward for long-term distributions where specified. *Italic: Content unit wpiea2022156-print-pdf - References (source PDF: wpiea2022156-print-pdf - References).* ### 2019. This is represents a 41-year time window that is well-suited to capture extreme realizations of ### Estimating Macro-Fiscal Effects of Climate Shocks From Billions of Geospatial Weather Observations — Appendix (Selected Tables and Figures) ### Data and variable construction - Time window: 1979–2019 (41-year window) used to capture extreme realizations of temperature and precipitation. - Data source: Data downloaded from Google Earth Engine. See http://www.climatologylab.org/terraclimate.html and https://developers.google.com/earth-engine/datasets/catalog/IDAHOEPSCORTERRACLIMATE. - Temperature sampling: - Daily precipitation: distribution uses all days of the calendar year (1979–2019). - Daily mean temperature: distributions restricted to moving windows centered on the day of interest (5-day and 15-day windows as in Kim et al. (2020); Perkins and Alexander (2013)). - Example: For August 16, 2000, precipitation extreme check vs distribution over all days 1979–2019; temperature extreme check vs August 14–18 over 1979–2019 (5-day window). ### Definitions of climate variables (selected descriptors) - Temperature (grid/day/country/year level): T_{j,d}, T_d, ř_j T_{j,d}, ř_d T_d{365, TVarL = ř_d pT_d́Tq^2{(p365 ́1)q} - Daily minimum/maximum: TN_{j,d}, TX_{j,d}; country-day aggregates TN_d, TX_d = ř_j TN_{j,d}, ř_j TX_{j,d}. - Diurnal T Range: DTR_d = TX_d ́TN_d; DTRL = ř_d DTR_d{365. - Percentiles and extreme counts: - TN_{p(k)}_d, TX_{p(k)}_d: pth percentile of 1979-2019 distribution in k-day window centered on d. - CN10L # of Cold Nights: ř_d rTN_dăTN10p5q_d sa (CountryYear) - WD90L # of Warm Days: ř_d rTX_dąTX90p5q_d sa (CountryYear) - Heat and cold waves (country-year): - DDHWL Heat Wave Days: number of days with TX_dąTX90p15q_d for at least 3 consecutive days. - DNHWL Heat Wave Nights: nights with TN_dąTN90p15q_d for at least 3 consecutive nights. - DD C WL Cold Wave Days: days with TX_dăTX10p15q_d for at least 3 consecutive days. - DNCWL Cold Wave Nights: nights with TN_dăTN10p15q_d for at least 3 consecutive nights. - Longest and counts: LDHWL, LNHWL, LDCWL, LNCWL, NDHWL, NNHWL, NDCWL, NNCWL (definitions in table). - Absolute temperature thresholds (country-year share of grid-days): - TN0L Frost prevalence: ř_d ř_j “TN_{j,d}ă0”‰{(Jˆ365)q - TX35L Max T above 35: ř_d ř_j “TX_{j,d}ą35”‰{(Jˆ365)q - TX40L Max T above 40: ř_d ř_j “TX_{j,d}ą40”‰{(Jˆ365)q - TSă ́9 Mean T below -9: ř_d ř_j “T_{j,d}ă ́9”‰{(Jˆ365)q - TSě30 Mean T above 30: ř_d ř_j “T_{j,d}ě30”‰{(Jˆ365)q - TS_{rx1,x2} Mean T in [x1,x2): increments of 3 ̊C from -9 ̊C to 30 ̊C for x1,x2. - Precipitation variables (grid/day/country/year): - P_{j,d}, P_d, PL Mean Precipitation ř_d P_d{365. - Wet day: PW_d = P_d rP_dě1s; WL # of Wet Days ř_d rP_dě1s. - PWp and PWp_j percentiles: pth percentile of 1979-2019 distribution using only wet days. - P95WTL Very Wet Day PPT: ř_d P_d rP_dě1 and PW_dąPW95sa. - P99WTL Extremely Wet Day PPT: ř_d P_d rP_dě1 and PW_dąPW99sa. - Cont'd dry/wet spells: CDDL, CWDL, PCWDL, C95WDL, C99WDL, PC95WDL, PC99WDL, PXp5q (5-day max), PXp1q (1-day max), PXML (max 1-month), PNML (min 1-month). - Longest Dry Spell thresholds: LLDS_.5, LLDS_.65, LLDS_.80, LLDS_.95 defined as maximum number of consecutive days with ř_j “P_{j,d}ă1”‰/{J} greater than thresholds .50, .65, .80, .95. - Precipitation share bins: PS_ď1 (less than 1 mm), PS_110 (1 to 10 mm), PS_1020 (10 to 20 mm), PS_ą20 (above 20 mm). - MaxP_ą10, MaxP_ą20: heavy/extreme PPT maxima. - Droughts (PDSI-based, grid-month/country-year): - PDSI_{j,m}: Palmer Drought Severity Index (grid-month). - PDSIă ́3 L Drought Prevalence: ř_m ř_j rPDSI_{j,m}ă ́3s{(Jˆ12)q. - PDSIă ́4 Harsh Drought Prevalence (L, R). - PDSIą3 Wet Conditions Prevalence, PDSIą4 VeryWetConditions Prevalence. - MPDSI metrics: MPDSIă ́3 Drought Intensity (max over months), MPDSIă ́4 Harsh Drought Intensity, MPDSIą3 Wetness Intensity, MPDSIą4 High Wetness Intensity. Notes: The letter L in the second column indicates variables used in the LASSO procedure. The letter R indicates variables used in OLS regressions. References: a = Kim et al. (2020), b = Perkins and Alexander (2013), c = Palmer (1965), d = IPCC (2021c). ### Trends in weather variables (Table A.2 — trend tests by country) - Regression: w_t = α + β t + u_t; Column A reports average β across countries; Column B percent of countries with β significant at 5 percent; Column C percent of countries with both significant trend and structural break (supremum Wald test at 95 percent). - Selected exact values (A Average trend; (B) % of countries significant trend; (C) % significant trend and break): - Mean Temperature: 0.029 299% 49% - # of Warm Days: 1.3430 96% 59% - # of Warm Nights: 1.4725 96% 69% - # of Cold Days: -1.0482 95% 49% - # of Cold Nights: -1.1931 93% 59% - # of Night Heat Waves: 0.1165 89% 62% - Heat Wave Days: 0.6758 88% 55% - Heat Wave Nights: 0.7412 88% 58% - Mean T in [27; 30): 0.0018 80% 46% - Day T C Maximum: 0.0322 75% 32% - Max T above 35: 0.0000875% 35% - Warm Spell Duration: 0.4277 74% 47% - Cold Spell Duration: -0.3637 72% 42% - Frost Prevalence: -0.000969% 27% - Drought Intensity: 0.003934% 29% - Very Wet Day PPT: 1.5521 27% 14% - Precipitation Variance: 0.1201 26% 14% - Many precipitation variables show low percentages of countries with significant trends (often ≤ 24%). - Commentary in text: Positive trend for average annual temperature equal to 0.03 ̊C per year, noted as remarkably in line with IPCC WG I average decadal increase of 0.3 ̊C. ### Summary statistics of macro-fiscal variables (Table A.3) - Summary Statistics of First Differences (N, Mean, St. Dev., Between St. Dev., Within St. Dev.): - ∆ ln pGDP/POP q in GDP Growth Sample (p.c.) N=6,550 Mean = 1.640% St. Dev. = 4.99% Between = 2.15% Within = 4.68% - ∆ ln pGDP/POP q in Fiscal Sample (p.c.) N=3,849 Mean = 1.977% St. Dev. = 4.01% Between = 1.66% Within = 3.69% - ∆ Revenue-to-GDP (p.p.) N=3,849 Mean = 0.105% St. Dev. = 3.43% Between = 0.64% Within = 3.41% - ∆ Expenditure-to-GDP (p.p.) N=3,849 Mean = 0.079% St. Dev. = 3.23% Between = 0.87% Within = 3.19% - ∆ Balance-to-GDP (p.p.) N=3,849 Mean = 0.027% St. Dev. = 4.01% Between = 0.41% Within = 4.01% - ∆ Debt-to-GDP (p.p.) N=3,849 Mean = -0.013% St. Dev. = 11.4% Between = 2.95% Within = 11.1% - ∆ Revenue (p.c.) N=3,849 Mean = 3.968% St. Dev. = 14.5% Between = 2.8% Within = 14.2% - ∆ Expenditure (p.c.) N=3,849 Mean = 3.864% St. Dev. = 11.5% Between = 2.7% Within = 11.3% - ∆ Debt (p.c.) N=3,849 Mean = 4.136% St. Dev. = 18.6% Between = 5.1% Within = 18.1% - Notes: GDP per capita measured as difference of log GDP per capita. Fiscal variables measured as percentage of GDP; first differences measured in percentage points. ### Correlation analysis (Tables A.4, A.5) - Summary statistics of first differences of selected climate variables (Table A.4): - Harsh Drought Prevalence (W): Mean = 0.005 St. Dev. = 0.010 Between = 0.173 - Max T C above 35 (W): Mean = 0.001 St. Dev. = 0.001 Between = 0.020 - Mean T C in [9;12): Mean = 0.000 St. Dev. = 0.001 Between = 0.020 - Cold Spell Duration: Mean = -0.315 St. Dev. = 0.426 Between = 14.14 - Day T C Maximum: Mean = 0.044 St. Dev. = 0.050 Between = 1.282 - Wetness Intensity: Mean = 0.000 St. Dev. = 0.003 Between = 0.067 - Longest Dry Spell (.80) (W): Mean = -0.003 St. Dev. = 0.021 Between = 0.359 - Cont’d Heavy PPT: Mean = 0.029 St. Dev. = 1.013 Between = 19.52 - Correlation matrix with GDP growth and selected variables (Table A.5), computed using first differences: - Lag(1) of GDP Growth with GDP Growth: 0.394 - Lag(2) of GDP Growth with Lag(1): 0.227 and with GDP Growth: 0.378 - Harsh Drought Prevalence (W) correlations: with GDP Growth = -0.054; with Lag(1) = 0.021; with Lag(2) = 0.022 - Max T C above 35 (W) correlations: with GDP Growth = -0.046; with Lag(1) = 0.013; with Harsh Drought Prevalence = 0.183 - Mean T C in [9;12) correlations: with GDP Growth = 0.040; with Harsh Drought Prevalence = -0.033; with Max T above 35 = -0.032 - Mean Temperature (T) correlations: with GDP Growth = -0.006; with Harsh Drought Prevalence = 0.159; with Max T above 35 = 0.356 - Total Precipitation (P) correlation with GDP Growth = 0.009. - Interpretation: Raw correlations between GDP growth and first differences of weather variables generally small. Correlation negative for Max T above 35 and Harsh Drought Prevalence, positive for Mean T in [9;12). Average annual temperature and annual precipitation show very low correlation with GDP growth compared to selected weather variables. ### LASSO selection results and GDP effects (Tables A.7–A.10, A.14) - Baseline LASSO selection (Table A.7; N=6,550; R-squared = 0.272; Within R-squared = 0.110). Top selected variables and OLS coefficient estimates (country fixed effects; climate variables first-differenced and standardized; (W) = population-weighted; standard errors clustered by country): - Rank 1: Lag-1 GDP p.c. Growth Est. 0.285*** (0.0408) - Rank 2: Lag-2 GDP p.c. Growth Est. 0.0407* (0.0218) - Rank 3: Harsh Drought Prevalence (W) Est. -0.202*** (0.0529) - Rank 4: Max T C above 35 (W) Est. -0.173** (0.0730) - Rank 5: Mean T C in [9; 12) Est. 0.179*** (0.0424) - Additional selected variables up to rank 30 reported with estimates (see Table A.7). - Robustness/specification variants: - Without year effects (Table A.8): R-squared = 0.261; Within R-squared = 0.158. Top variables include Lag-1 GDP p.c. Growth 0.293*** and World GDP p.c. Growth 0.777***. Harsh Drought Prevalence (W) Est. -0.262***. - With quadratic trends (Table A.9): R-squared = 0.355; Within R-squared = 0.0447. Top variables include Lag-1 GDP p.c. Growth 0.173*** and Harsh Drought Prevalence (W) Est. -0.178***. - Balanced sample (Table A.10; Observations = 4,860; R-squared = 0.230; Within R-squared = 0.0858). Harsh Drought Prevalence (W) Est. -0.260***; Max T above 35 (W) Est. -0.242**. - Full baseline OLS estimates for key climate variables (Table A.14; Observations = 6,550): - Lag-1 GDP p.c. Growth coefficients across columns (A–G): 0.280***, 0.279***, 0.279***, 0.279***, 0.288***, 0.171***, 0.239*** (standard errors reported in table). - Lag-2 GDP p.c. Growth coefficients: 0.0410*, 0.0412*, 0.0407*, 0.0397*, 0.0479**, -0.03420, 0.0657***. - First difference in Harsh Drought Prevalence (W): -0.211***, -0.251***, -0.231***, -0.191***, -0.257*** (standard errors reported). - Max T C above 35 (W): -0.201***, -0.244***, -0.175**, -0.215***, -0.254***. - Mean T C in [9; 12): 0.155***, 0.174***, 0.153***, 0.142***. - Additional controls and specification notes: country fixed effects included; climate variables standardized; (W) indicates population-weighted variables. ### Comparisons with literature (Table A.15) - Re-estimation aligning with Burke et al. (2015) and Kahn et al. (2021) style specifications: - Burke-style columns show Lag-1 GDP p.c. Growth ~0.117* and temperature quadratic terms (Average Annual Temperature 0.0127***; squared -0.000479***). Harsh Drought Prevalence (W) and Max T above 35 (W) included in augmented specifications with coefficients -0.327*** and -0.229** respectively in augmented models. - Kahn-style columns show Lag-1 GDP p.c. Growth ~0.300*** and similar patterns when adding Harsh Drought Prevalence and Max T above 35. ### Macro-fiscal effects and robustness (Tables A.11–A.13, A.16, A.17) - LASSO-selected climate variables for government expenditure (Table A.11; Observations = 3,849; R-squared = 0.124; Within R-squared = 0.0693): - Lag-1 Expenditure Est. -0.249*** (0.0499) - Lag-2 Expenditure Est. -0.0618** (0.0297) - Lag-1 Harsh Drought Prevalence (W) Est. -0.0720* (0.0418) - Harsh Drought Prevalence (W) Est. 0.127** (0.0523) - Mean T in [-3; 0) (W) Est. 0.114* (0.0578) - LASSO-selected climate variables for government revenue (Table A.12; Observations = 3,849; R-squared = 0.117; Within R-squared = 0.0747): - Lag-1 Revenue Est. -0.254*** (0.0491) - Lag-2 Revenue Est. -0.132*** (0.0172) - Longest Dry Spell (.80) (W) Est. -0.0879 (0.104) - Cont’d Dry Days Est. -0.115 (0.0841) - LASSO-selected climate variables for government debt (Table A.13; Observations = 3,849; R-squared = 0.138; Within R-squared = 0.0285): - Lag-1 Debt Est. 0.0735** (0.0293) - Lag-2 Debt Est. 0.136** (0.0681) - Wetness Intensity Est. 0.370** (0.182) - Macro-fiscal baseline robustness (Table A.16; Observations = 3,849 across columns): - Lag(1) of Fiscal Variable significant in many specifications: e.g., ∆ ln GDP POP (p.c.) Lag(1) = 0.169*** (0.0561). - First difference in Harsh Drought Prevalence (W) effects: - ∆ ln GDP POP (p.c.): -0.149** (0.0663) - ∆ Revenue GDP (p.p.): 0.0861* (0.0509) - ∆ Expenditure GDP (p.p.): 0.121** (0.0499) - ∆ Balance GDP (p.p.): -0.0357 (0.0530) - ∆ Debt GDP (p.p.): -0.0821 (0.114) - ∆ ln Revenue p.c.: 0.266 (0.207) - ∆ ln Expenditure p.c.: 0.301* (0.179) - ∆ ln Debt p.c.: -0.204 (0.229) - Max T above 35 (W) and Mean T in [9;12) show varying significance across fiscal outcomes; Very Wet Conditions and Longest Dry Spell (.95) (W) also show effects in some columns. - Controls include first two lags of dependent variable, year and country fixed effects. Climate variables standardized with standard deviations equal to 100 for interpretation. - Role of fiscal space (Table A.17; Observations = 3,849): - Lag(1) of Fiscal Variable coefficients: 0.172*** (∆ ln GDP POP), -0.257*** (∆ Revenue GDP), -0.252*** (∆ Expenditure GDP), 0.139*** (∆ Debt GDP). - Interaction specifications include “Mean-deviation of debt (cubic)” and interactions of climate changes with lag(1) debt deviation to assess role of fiscal space in moderating climate impacts. - Example interaction: ” interacted with lag(1) debt deviation for Lag-1 Cont’d Heavy PPT = -0.0438** (0.0213) and the interaction coefficient = 0.0570* (0.0325) for ∆ Revenue GDP. ### Additional notes on model performance and figures - Within R-squared improvements and variable selection outcomes illustrated in Figures A.2–A.4, A.6–A.7 (LASSO paths, information criteria, and within R-squared reported relative to baseline specifications in notes). - Persistence of temperature shock on GDP per capita shown in Figure A.5: impulse response of per capita output in levels to one standard deviation shock of absolute deviation in annual average temperature relative to past 30-year average; horizon 0 is year of shock; shaded area 90 percent confidence intervals. - Balanced-sample and trend-robust specifications reported (Tables A.9–A.10, Figure A.4), showing consistent selection of Harsh Drought Prevalence and Max T above 35 as influential climate variables for GDP outcomes. *Estimating Macro-Fiscal Effects of Climate Shocks From Billions of Geospatial Weather Observations — Working Paper No. WP/2022/156* --- _Source: https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022156-print-pdf.pdf_