Correcting for spatial and temporal dependence in estimating the economic effect of extreme heat

Callahan and Mankin argue that extreme heat, measured by maximum 5-day average temperature, has a precise and economically meaningful effect on subnational gross domestic product per-capita growth. We reexamine this claim and find that the reported precision does not survive methods that respect the data's spatial and temporal dependence. Country-preserving permutations and postselection correction substantially widen marginal-effect intervals. The estimates are also highly sensitive to influential countries and years, with exclusions sometimes collapsing the effect or reversing its sign. A hierarchical Bayesian model with region-within-country structure and AR(1) year effects yields much larger uncertainty, shrinks the estimated effect by 67 to 90%, and produces credible intervals spanning zero. Last, rolling out-of-sample tests show that adding extreme-climate variables does not improve prediction over models without them. We conclude that the claimed precision is not sustained once dependence, model selection, and influential observations are handled appropriately.