## Data Source

The temperature and precipitation data used in this study were derived from the CRU database, which provides a high-resolution (0.5°×0.5° resolution), monthly grid of land-based observations going back to 1901. The CRU database has implemented a degree of homogenization and revealed no substantial discrepancies with other climate databases. It has been widely used throughout the literature, allowing us to compare our results with other research findings (13, 19, 20). We also conducted a robustness check by using the reanalysis data from the ERA5 dataset.

The gross domestic income per capita (2011 PPP) used in this study is obtained from the Global Data Lab (https://globaldatalab.org/shdi/table/lgnic/). For high-income countries and some middle-income countries, subnational GDI per capita was obtained based on data derived from national statistical offices and Eurostat. In the case of most low- and middle-income countries, where official statistical data are lacking or unreliable, subnational GDI per capita data were obtained based on the International Wealth Index (IWI). It is an asset-based wealth index constructed on data from 165 surveys over 2.1 million households in 97 low- and middle-income countries. It measures household wealth on the basis of information from asset ownership, housing quality, and access to public services. Compared to statistical data from various countries and other surveys, the most significant feature of the IWI is its comparability across countries and over time, as it relies on a common set of assets(21). Based on the IWI score, the subnational GDI per capita data were estimated using a regression model between national GDI per capita and national IWI score. It is further scaled to make their population weighted mean for a given year equals the national GNI values derived from the UNDP-HDI database (*SI Appendix*, Table S1). Up to now, the subnational GDI per capita data has been used in sub-national HDI assessments and relevant studies within developing countries (22, 23, 24).

The gross domestic product per capita (2011 PPP) is obtained from Kummu, Taka and Guillaume (25). The database was initially collected by Gennaioli, La Porta, Lopez-de-Silanes and Shleifer (15) based on various government statistical agencies. It includes GDP data from 1569 subnational regions across 110 countries between 1990 and 2010 and covers most countries in Central and South Africa, which is generally omitted by other GDP databases. Kummu, Taka and Guillaume (25) extended the time series of this database from 2010 to 2015 and filled in the missing countries based on national GDP data. Overall, the database extended by Kummu, Taka and Guillaume (25) covers the global subnational GDP data with no missing data areas and converts the data to constant international US dollars from 2005 to 2011, which is consistent with the gross domestic income data used in this study (*SI Appendix*, Table S1).

Two other socioeconomic variables, mean years of schooling, and population, are included in our regression model as control variables since they are also regarded to significantly impact economic growth. These data are also collected from the Global Data Lab (Access data: 22 November 2022).

## Weather variability

Extreme weather events have been found to have a negative impact on human psychology, water supply, and agricultural production that further systematically increase the risk of conflict, violence, or political instability (26–31). However, most recent macroeconomic studies have focused on the level effects of weather conditions on economic growth while neglecting to identify the effects of extreme weather events, which have been found to have pronounced effects on economic activity as well (14, 32). This study therefore uses the Anomaly Standardized Precipitation (ASP) and Anomaly Standardized Temperature (AST) index as measures to identify the effects of weather variability. The ASP and AST are defined by an annual sum of monthly precipitation or temperature anomalies from their climatological means:

$${S}_{r,y}=\sum _{m=1}^{12}\frac{{vec{T}}_{r,m,y}-{\stackrel{-}{vec{T}}}_{r,m}}{{\sigma }_{r,m}}$$

1

Where \({S}_{r,y}\) is the anomaly standardized precipitation/temperature in region \(r\) and year \(y\). \({vec{T}}_{r,m,y}\) is the monthly total precipitation or monthly mean temperature. \({\sigma }_{r,m}\) is the historical standard deviation, for 25 years, of monthly total precipitation or monthly mean temperature in that region. The annual \({S}_{r,y}\) is further standardized by its standard deviation over 25 years to obtain a measure of the relative severity of annual precipitation/temperature variability (see *SI Appendix* for the origins of the method):

$${ASvec{T}}_{r,y}={S}_{r,y}/{\sigma }_{{S}_{r}}$$

2

## Empirical model

We use the fixed-effects regression model to identify the effects of weather conditions on economic growth since this model can strengthen the causal effects identification by controlling for both unobserved time-invariant and time-varying influences, such as the influences from geographical location and institutional differences (10–14). Apart from the effects of weather variability, we also consider the effects of weather and weather change in our regression model. In particular, the effect of weather is measured by \(\beta T+\beta {T}^{2}\). It captures the nonlinear effect of the prevailing weather conditions on transitory and long-run economic growth. The difference between transitory and long-run growth effects is that the transitory effects imply economic growth can eventually reverse itself as the weather conditions return to their prior state. However, the long-run growth effects are not reversed, and a failure to innovate in one period leaves the country permanently further behind (10). (We additionally conducted a long-difference regression to clarify the specific effects revealed by the quadratic function of temperature and precipitation. See *SI Appendix*, Table S8). The effect of weather change is measured by \(\beta Delta T+\beta Delta T\bullet T\) and captures a sudden change in weather conditions on contemporaneous growth rates. If there is no change between the two years, this effect disappears. The interaction term \(\beta Delta T\bullet T\) is to capture the moderating effects of the prevailing weather conditions \(T\) on *∆T*. The regression model is thus defined as follows:

$${g}_{i,t}={\beta }_{1}{Delta vec{T}}_{vec{i},vec{t}}+{\beta }_{2}{Delta vec{T}}_{vec{i},vec{t}}\bullet {vec{T}}_{vec{i},vec{t}}+{\beta }_{3}{vec{T}}_{vec{i},vec{t}}+{\beta }_{4}{vec{T}}_{vec{i},vec{t}}^{2}+{\beta }_{5}{vec{A}vec{S}vec{T}}_{vec{i},vec{t}}+{\beta }_{6}{vec{A}vec{S}vec{T}}_{vec{i},vec{t}}^{2}+$$

$${X}_{i,t}+{\alpha }_{i}+{\delta }_{t}+{\epsilon }_{i,t}$$

3

where \({\text{g}}_{i,t}\) is the GDP per capita or GDP per capita growth rate within year \(t\) in region \(i\). \({vec{T}}_{vec{i},vec{t}}\) is a vector of annual mean temperature levels (*T*, in ℃) and annual mean precipitation values (*P*, in m). Coefficients \({\beta }_{1}\) and \({\beta }_{2}\) capture the effect of weather change on economic growth. Coefficients \({\beta }_{3}\) and \({\beta }_{4}\) represent the effect of weather on economic growth. \({\beta }_{5}\) and \({\beta }_{6}\) capture the effect of weather variability on economic growth. \({\alpha }_{i}\) and \({\delta }_{t}\) are region and year dummies to consider region and year fixed effects, respectively. \({X}_{i,t}\) are control variables, including the mean year of schooling and population. \({\epsilon }_{i,t}\) is the error term.

In addition, a gross of the literature shows that the increase in global temperature results in an observable increase in both the intensity and frequency of anomalous events (33–35). In addition, the effect of the global mean temperature on such events is nonlinear, with a minor increase in mean temperature leading to a substantial escalation in the frequency and intensity of anomalous events (36–38). In this case, the increase in temperature and precipitation (weather effect) is expected to intensify the effects of change in anomalous events (variability effect) on economic growth and vice versa. To assess the interaction effects between the level and variability of weather conditions, we further included the interaction term of them in Eq. (3) as follows:

$${g}_{i,t}={\beta }_{1}{Delta vec{T}}_{vec{i},vec{t}}+{\beta }_{2}{Delta vec{T}}_{vec{i},vec{t}}\bullet {vec{T}}_{vec{i},vec{t}}+{\beta }_{3}{vec{T}}_{vec{i},vec{t}}+{\beta }_{4}{vec{T}}_{vec{i},vec{t}}^{2}+{\beta }_{5}{vec{A}vec{S}vec{T}}_{vec{i},vec{t}}+{\beta }_{6}{vec{A}vec{S}vec{T}}_{vec{i},vec{t}}^{2}+$$

$${\gamma }_{1}{L}_{i,t}\bullet {V}_{i,t}+{\gamma }_{2}{L}_{i,t}^{2}\bullet {V}_{i,t}+{\gamma }_{3}L\bullet {V}_{i,t}^{2}+{\gamma }_{4}{L}_{i,t}^{2}\bullet {V}_{i,t}^{2}+{X}_{i,t}+{\alpha }_{i}+{\delta }_{t}+{\epsilon }_{i,t}$$

4

Where \({L}_{i,t}\) is the annual mean temperature level or annual mean precipitation value within year \(t\) in region \(i\). \({V}_{i,t}\) is the AST or ASP.

Given the limited adaptive capacity and a higher share of agriculture in economic activity, poor regions are regarded to be more vulnerable to climate change than rich regions (24). To assess the heterogeneity of the level and variability effects, as well as the interaction effects of climate conditions on economic growth, we reassess our results separately for subnations with above- and below- median subnational GDI per capita or GDP per capita. The regression model for heterogeneity of level and variability effects reads:

$${g}_{i,t}={\beta }_{1}D\bullet {Delta vec{T}}_{vec{i},vec{t}}+{\beta }_{2}D\bullet {Delta vec{T}}_{vec{i},vec{t}}\bullet {vec{T}}_{vec{i},vec{t}}+{\beta }_{3}D\bullet {vec{T}}_{vec{i},vec{t}}+{\beta }_{4}D\bullet {vec{T}}_{vec{i},vec{t}}^{2}+{\beta }_{5}D\bullet {vec{A}vec{S}vec{T}}_{vec{i},vec{t}}+{\beta }_{6}D\bullet {vec{A}vec{S}vec{T}}_{vec{i},vec{t}}^{2}+$$

$${X}_{i,t}+{\alpha }_{i}+{\delta }_{t}+{\epsilon }_{i,t}$$

5

The regression model for heterogeneity of interaction effects reads:

$${g}_{i,t}={\beta }_{1}{Delta vec{T}}_{vec{i},vec{t}}+{\beta }_{2}{Delta vec{T}}_{vec{i},vec{t}}\bullet {vec{T}}_{vec{i},vec{t}}+{\beta }_{3}{vec{T}}_{vec{i},vec{t}}+{\beta }_{4}{vec{T}}_{vec{i},vec{t}}^{2}+{\beta }_{5}{vec{A}vec{S}vec{T}}_{vec{i},vec{t}}+{\beta }_{6}{vec{A}vec{S}vec{T}}_{vec{i},vec{t}}^{2}+$$

$${\gamma }_{1}D\bullet {L}_{i,t}\bullet {V}_{i,t}+{\gamma }_{2}D\bullet {L}_{i,t}^{2}\bullet {V}_{i,t}+{\gamma }_{3}D\bullet L\bullet {V}_{i,t}^{2}+{\gamma }_{4}{D\bullet L}_{i,t}^{2}\bullet {V}_{i,t}^{2}+{X}_{i,t}+{\alpha }_{i}+{\delta }_{t}+{\epsilon }_{i,t}$$

6

Where \(D\) is a binary variable that equals 1 when GDI per capita or GDP per capita is below the median and 0 otherwise.

Considering that the regions’ adaptation to climate change may not change rapidly over a short period, we took the five-year average of GDI per capita and GDP per capita for each region and determined their median values separately. The regions with averaged GDI per capita or GDP per capita above the median are considered to be rich, while those below the median are considered to be poor.