Understanding the Human Effects of Climate Change

Climate change has already begun to profoundly alter the relationship betweenhumans and their environment for the vast majority of the world’s population. How-ever, history has demonstrated that human are nothing if not responsive: as theclimate changes, so too will economies, governments, and individuals. This disser-tation examines impacts and responses to climate change with an eye towards un-derstanding how future societies might adapt to substantial climatic changes. Thefirst chapter measures the welfare cost of changes in amenity values due to climatechange by proxying for temperature preferences using contemporaneous changes inmood, as detected from posts on the social media platform Twitter. The secondchapter examines the response of electricity demand to changes in temperature asa means to project patterns of future energy consumption and large-scale capitalinvestments. The third chapter makes a methodological contribution to test threequasi-experimental methods of estimating electricity savings in dynamic pricing pro-grams versus an empirical “gold standard”: the results from this chapter will aidpolicymakers in quantifying the effects these programs on curbing future increasesin electricity generation due to climate change.The first chapter is motivated by a gap in the climate impacts literature: thechange in amenity values resulting from temperature increases may be a substantialunaccounted-for cost of climate change. Without an explicit market for climate, priorwork has relied on cross-sectional variation or survey data to identify this cost. Thispaper presents an alternative method of estimating preferences over nonmarket goodswhich accounts for unobserved cross-sectional and temporal variation and allows forprecise estimates of nonlinear effects. Specifically, I create a rich panel dataset onhedonic state: a geographically and temporally dense collection of updates from thesocial media platform Twitter, scored using a set of both human- and machine-trainedsentiment analysis algorithms. Using this dataset, I find strong evidence of a sharpdeclines in hedonic state above and below 20 ◦ C (68 ◦ F). This finding is robust acrossall measures of hedonic state and to a variety of specifications.The second chapter simulates the effect of climate change on future electricitydemand in the United States. We combine fine-scaled hourly electricity load datawith observations of weather to estimate the response of both average and peakelectricity demand to changes in temperature. Applying these estimates to a set oflocally downscaled climate projections, we project regional end-of-century changesin electricity load. The results document increases in average hourly load across thecountry, with more pronounced changes occurring in the southern United States.Importantly, we find changes in peak demand to be larger than changes in aver-age demand, which has implications for public policy choices around future capitalinvestment.The third chapter compares quasi-experimental designs to experimental designs inthe context of a dynamic pricing setting designed to encourage customers to save en-ergy. Randomized controlled trials (RCTs) are widely viewed as the “gold standard”for evaluating the effectiveness of an intervention. However, because are percievedto be prohibitively expensive and challenging to implement successfully, they arenot broadly executed in policy settings. In particular, analysis of the effect of energypricing has largely been conducted through a two commonly used quasi-experimentalmethodologies: difference-in-differences and propensity score matching. Using a rareset of large-scale randomized field evaluations of electricity pricing, we compare theestimates obtained from these quasi-experimental designs and from a regression dis-continuity design to the true estimates obtained through the experimental method.We demonstrate empirical evidence in favor of four stylized facts that highlight theimportance of understanding selection bias and spillover effects in this context. First,difference-in-differences and propensity-score methods mis-estimate the true effectby up to 5% of mean peak hour usage. Second, propensity score estimates resembledifference-in-difference findings, but standard errors tend to be larger and point esti-mates are more biased for opt-out models. Third, regression discontinuity methodscan be heavily biased relative to the true average treatment effect. Finally, we findstrong evidence that biases are more pronounced in opt-in vs. opt-out designs