Crop-damaging temperatures increase suicide rates in India

Suicide and Agricultural Data.

Annual suicide data are reported by the Indian NCRB at the state level beginning in 1967 for 27 of India’s 29 states and 5 of its 7 union territories. Suicide records are in NCRB’s “Accidental Deaths and Suicides in India” report and include the total number of state suicides per year. I calculate suicide rates as the number of total suicides per 100,000 people, with population values linearly interpolated between Indian censuses. I use agricultural data from ref. 25. These are district-level annual yield records for major crops (rice, wheat, sugar, sorghum, millet, and maize) between 1956 and 2000, compiled from Indian Ministry of Agriculture reports and other official sources. These data cover 271 districts in 13 major agricultural states, and provide log annual yield values of a production-weighted index across all crops measured in constant Indian rupees, where prices are fixed at their 1960–1965 averages. Details on these data and summary statistics are provided in SI Appendix.

Climate Data.

Climate data are generally available at higher spatial and temporal resolution than social outcome data. Although suicides and yields are only measured annually, if the relationship between these outcomes and temperature is nonlinear, daily climate data are required, as annual averages obscure such nonlinearities (26). For daily temperature data, I use the National Center for Environmental Prediction gridded daily reanalysis product, which provides observations in a grid of ∼1° × 1°. These data include daily mean temperature for each grid over my sample period. To convert daily temperature into annual observations without losing intraannual variability in daily weather, I use the agronomic concept of degree days (details in SI Appendix). I aggregate grid-level degree day values to state-level observations using an area-weighted average (see SI Appendix, Table S5 for robustness checks using weights based on population and area planted with crops). When these state-level degree day values are summed over days within a year, regressing an annual outcome on cumulative degree days imposes a piecewise linear relationship in daily temperature, in which the outcome response has zero slope for all temperatures less than T∗. Although a body of literature identifies biologically determined cutoffs T∗ for yields of a variety of major crops, there is no empirical support to draw on in selecting T∗ for suicides. Thus, although I use T∗=20°C throughout this study, I show robustness for a range of plausible cutoffs based on the distribution of my temperature data, and, in Fig. 1, I estimate a flexible piecewise linear function using four different degree day cutoffs to impose minimal structure on the response function.

Because reanalysis models are less reliable for precipitation data, and because nonlinearities in precipitation that can’t be captured with a polynomial appear to be less consistently important both in the violent crime literature (27) and in the agriculture literature (19), I use the University of Delaware monthly cumulative precipitation data to complement daily temperature observations (28). These data are gridded at a 0.5° × 0.5° resolution, with observations of total monthly rainfall spatially interpolated between weather stations. I again aggregate grids up to states using area-based weights, after calculating polynomial values at the grid level first.

Regression Estimation.

To identify the impact of temperature and precipitation on suicide rates, I estimate a multivariate panel regression using ordinary least squares, in which the identifying assumption is the exogeneity of within-state, annual variation in cumulative degree days and precipitation. My primary estimation approach uses a flexible piecewise linear specification with respect to temperature and a cubic polynomial function of precipitation. To isolate the impact of economically meaningful climate variation, I separately identify the temperature and precipitation response functions by agricultural seasons (see SI Appendix for details). My empirical model takes the general formsuicide_rateit=∑s=12∑k=1κβks∑d∈sDDidtk+∑s=12gs(∑m∈sPimt)+δi+ηt+τit+εit,[1]where suicide_rateit is the number of suicides per 100,000 people in state i in year t, s indicates season (growing and nongrowing), and k=1,…,κ indicates a set of degree day cutoffs that constrain the piecewise linear response. In my most flexible model, I let κ=7 with degree day intervals of 5 °C, and, in my simplest model, I let κ=2 and estimate a standard degree day model with just one kink point and two piecewise linear segments. DDidtk is the degree days in bin k (e.g., degree days between 10 °C and 20 °C) on day d in year t in state i, and Pimt is cumulative precipitation during month m in year t in state i. I estimate g(⋅) as a cubic polynomial. State fixed effects δi account for time-invariant unobservables at the state level, year fixed effects ηt account for India-wide time-varying unobservables, and state-specific time trends τit control for geographically differentiated trends in suicide driven by time-varying unobservables. Robustness to different temporal adjustments is shown in SI Appendix, Table S8.

Eq. 1 identifies β^ks, the season-specific estimated change in the annual suicide rate caused by 1 d in bin k becoming 1 °C warmer. This annual response to a daily forcing variable is described in detail in ref. 26. The polynomial response function for precipitation generates marginal effects of one additional millimeter of rainfall, again estimated seasonally. Due to likely correlation between errors within states, I cluster standard errors at the state level. This strategy assumes that spatial correlation across states in any time period is zero, but flexibly accounts for within-state, across-time correlation. I estimate a nearly identical specification as shown in Eq. 1 for agricultural yields. However, with district-level data, I include district fixed effects and state-specific time trends, and I cluster standard errors at the district level.

Adaptation.

Fig. 3 shows results from four sets of tests for evidence of adaptation. The exact specifications for all regression models are shown in SI Appendix. All models use a variant of Eq. 1 in which κ=2, the degree day cutoff is set to 20 °C, and state-specific linear trends are included. In Fig. 3 A and B, I estimate Eq. 1, but add an interaction term between degree days in the growing season and an indicator for the tercile of average growing season degree days that state i falls into (Fig. 3A) or an indicator for the tercile of average gross domestic product (GDP) per capita that state i falls into (Fig. 3B). These distributions are defined over all states and all years in the sample. In Fig. 3C, I split the 47 y in my sample into three temporal subsamples, and estimate the coefficient on an interaction between growing season degree days and an indicator for each of these three subsamples. In Fig. 3D, I estimate a “panel of long differences” empirical model in addition to the standard panel regression in Eq. 1 (29). To do so, I collapse my data to four observations for each state, where each observation measures the 10-y change in suicide rates and climate variables for each decade, and where these changes are “smoothed” by taking 5-y averages at the end points. I then estimate the effect of changes in average degree days and precipitation on changes in average suicide rates.

Attribution of Climate Trends.

To compute estimates of the effect of warming temperature trends since 1980, I follow the approach outlined in refs. 18 and 30. I first estimate a state-specific linear trend in growing season degree days above 20 °C for the years 1980–2013. I then generate a detrended degree days residual that is normalized to temperature in 1980 and predict suicide rates using actual and detrended growing season degree days. In so doing, I use the coefficient estimates from the model in Table 1 which includes both state trends and year fixed effects (column 3). The elevated risk of suicide attributable to the trend, relative to the detrended counterfactual, is the difference between these two predictions. Multiplying by the population in each state and each year recovers the total additional number of suicides. Fig. 4B displays these additional deaths in each year; integrating over states and years gives the cumulative effect of temperature trends for all of India over the entire period since 1980 (see SI Appendix for details).