Convex temperature damage to global food 1 production, concentrated in drought years and 2 summer-dry economies
Data and reproduction package for "Convex temperature damage to global food production, concentrated in drought years and summer-dry economies" (Leirvik, Nature Food). This deposit contains the processed analytical dataset, all intermediate data files, and the analysis scripts needed to reproduce the results, tables and figures in the manuscript. The study analyses a balanced country-year panel of 141 countries over 1961–2023, built from CRU TS v4.08 climate data, the World Bank Food Production Index and ESG indicators, and FAOSTAT crop yields, together with CMIP6 SSP scenario projections. It estimates a convex temperature damage function for national food production growth, tests where damage falls across the growth distribution and across countries, and evaluates whether a compound heat–drought channel is detectable at annual national resolution. Contents. The archive root holds the merged analytical dataset (df_bf.xlsx, 9,024 × 98), the CRU-derived climate panel, the World Bank and FAOSTAT inputs, the country-specific bridging regressions used for the SSP projections, the projection objects, and the causal-machine-learning outputs (causal forest, X-learner and Bayesian causal forest country-level treatment effects). code/ contains a script that rebuilds the cached analysis workspace from the deposited data, plus the scripts generating the event-study coefficients, the projection gap tests and their decomposition, the level-versus-growth tests, the specification sensitivity analysis, the bootstrap standard errors and the crop-level marginal effects. derived/ retains the estimates those scripts write, so published numbers can be checked without re-running anything. CODEBOOK.md documents every variable, including the standardisation conventions, which differ between the panel regressions and the causal forest. Reproducibility. R ≥ 4.5 (results produced under 4.5.1); key packages fixest, grf, quantreg, DIDmultiplegtDYN. Random seeds are fixed and recorded: the causal forest uses seed 42, the heterogeneity-robust event study and all bootstraps use seed 20260725. Every other estimate is deterministic. README.md gives the directory layout the scripts expect and the order in which to run them. Provenance. CRU TS v4.08 (University of East Anglia CRU/NCAS, Open Government Licence); World Bank Open Data (CC BY 4.0); FAOSTAT via Our World in Data (CC BY 4.0); CMIP6 via the KNMI Climate Explorer. Derived files and scripts are released under CC BY 4.0.