Temperature and Temperament: Measuring Human Emotion with Twitter Data

A growing literature in environmental economics documents the causal impacts of climate over a range of economic outcomes. Notably, while there is considerable evidence that warmer temperatures increase conflict, lower income, and lower productivity, the mechanisms through which these impacts operate remain uncertain. Since human decision making is an essential contributor to many of these outcomes, a plausible candidate mechanism is the behavioral changes induced by changes in temperature. Prior research on the behavioral effects of temperature suggests that changes in mood in particular may factor into many of the observed outcomes. However, due to the scarcity of available data, evidence of the effect of temperature on emotion remains thin and is largely confined to laboratory settings.

To fill this gap, I build a rich, novel dataset on human emotion: a geographically and temporally dense corpus of Twitter status updates with over 700 million observations, scored using a set of both human- and machine-trained sentiment analysis algorithms. These data allow me to estimate a consistent non-linear relationship between temperature and mood: mood is unaffected by temperature below 60 degrees F, and slopes downward above 60 degrees F. The shape of this response function is similar to those observed for conflict, income, and productivity. I additionally document increased aggression and spelling errors in higher temperatures as evidence of other dimensions of behavioral change in high temperatures. The results have implications for both our understanding of the effect of environment on human behavior and suggest a channel for many of the documented effects of climate change.

Speaker: Patrick Baylis, UC Berkeley