Fatal Errors: The Mortality Value of Accurate Weather Forecasts

Routine weather forecasts are the product of a sophisticated scientific and policy effort requiring global cooperation (Benjamin et al., 2018). Despite their ubiquity and the effort that goes into their creation, surprisingly little is known about their value to the public. Understanding this value is important for at least two reasons. First, weather forecasts are a far-reaching information intervention that is particularly well-suited to offering empirical insights into how individuals use information. Yet researchers have only rarely used forecasts to study informational questions (e.g., Roll, 1984). Second, forecasts are a costly but potentially important policy tool. Globally, public funding for forecasting agencies exceeds $15 billion per year (Rogers and Tsirkunov, 2013), and decisions about future investments in forecasts require information about the value of making forecasts more accurate.

Methods
We combine new theory and comprehensive data to provide the first revealed preference evidence of behavioral responses to forecasts and the value of improved forecasts. In particular, we examine how much more accurate forecasts help people avoid mortality from temperature. We do so by formally demonstrating that whether more accurate forecasts reduce mortality depends on the relationship between mortality risk and forecast errors. If mortality risk is convex in forecast errors, then improvements in forecast accuracy will reduce expected mortality. The convexity of mortality risk depends on how agents’ chosen actions depend on whether they receive one forecast or another. Namely, risk is convex if actions are “appropriate” in the sense of seeking to suit the weather rather than being “protective” in the sense of monotonically reducing mortality as more action is taken. Because either model of adaptation is plausible, it is an empirical question whether or not more accurate forecasts reduce mortality.

Data
To estimate the effect of weather forecasts on mortality, we combine the universe of deaths reported by the Centers for Disease Control and Prevention (CDC) with daily weather and forecasts issued by the National Weather Service (NWS). We study the continental U.S. from 2005 through 2017 and focus on day-ahead forecasts of temperature. Our regression framework accounts for potential location-specific and time-varying confounders as well as for the potential direct effects of temperature and other aspects of weather on mortality.

Results
Across the full sample, we find that mortality risk is indeed convex in forecast errors. Reducing the standard deviation of forecast errors by 50% would save 2,200 lives per year. We find that mortality risk is especially convex in forecast errors on days with extreme heat. Climate change will make such days more common over the coming century. As a result, climate change increases the mortality benefit of improved forecasts to 2,400 lives saved annually by 2100. Short- run weather forecasts thereby facilitate adaptation to climate change.
We also show how to estimate agents’ ex ante willingness to pay for more accurate forecasts, accounting for their costs of acting on forecasts. In particular, we formally derive the change in an agent’s value function from more accurate forecasts. We find that the net value of making forecasts 50% more accurate is $2.1 billion per year, or roughly 10% of the monetized mortality benefit. Climate change increases that value to $2.9 billion per year by 2100, so that the present value of improving forecast accuracy by 50% over the remainder of the century is $112 billion.

Forecasts can affect mortality only if people take actions based on them. We also report direct evidence that people do indeed act on forecasts: electricity use responds to forecasts, time use responds to forecasts, estimated responses to forecasts vary cross-sectionally in intuitive ways, and, in a new survey conducted for this paper, college students state that they modify behavior in response to forecasts. This evidence and the robustness of our results to model specification suggest that our estimated mortality benefits do indeed capture actual use of forecasts.

Our results are the first revealed preference estimates of the benefits of routine weather forecasts. Recent theoretical work emphasizes that short-run forecasts such as those studied here can be especially valuable for planning purposes (Millner and Heyen, 2021). Previous valuations of routine weather forecasts calibrated models of particular decision problems (e.g., Lave, 1963, Wilks, 1997, Richardson, 2000), tallied up the value of sectors judged to be sensitive to weather (e.g., National Research Council, 1998), or surveyed potential users (e.g., Haas and Rinkle, 1979, Stewart, 1997, Stratus Consulting Inc., 2002, Lazo et al., 2009). Many authors in the forecasting literature have recognized that it would be ideal to find a market in which people reveal their value for forecasts with real bets but lament that such markets do not exist for publicly provided forecasts (e.g., Freebairn and Zillman, 2002, Letson et al., 2007, Morss et al., 2008, Katz and Lazo, 2011). We here infer that agents use and value forecasts by exploring how forecasts affect observed mortality.

Previous work has found that both hot and cold temperatures are associated with excess mortality but modeled weather as an unanticipated shock. Accurately forecasted weather shocks can have very different implications from inaccurately forecasted weather shocks if people act on their information about coming weather. It is important to disentangle these effects when assessing policy responses to extreme temperatures and also when extrapolating to the effects of climate change. Our study shows that one of the most pervasive, prominent, and complex informational interventions undertaken by governments does generate substantial value for agents in the economy.