While exposure to high ambient temperature (i.e., heat) has long been recognized as a threat to public health, the burden of illness and death attributable to heat in the US remains high. In an effort to reduce heat- related mortality and morbidity, the US National Weather Service (NWS) issues heat alerts in advance of forecasted extreme heat events to communicate these risks to the public and government officials. However, it is largely unknown: (1) what are the optimal metrics of heat stress to inform when to issue heat alerts, (2) how effective are heat alerts in protecting the public?s health and (3) what factors make heat alerts comparatively more or less effective in some places or in some people versus others. In the absence of such information, we will fail to maximize the public health benefits of heat alerts. The goals of this proposal are to identify the optimal health-based and location-specific metrics for issuing heat alerts, to estimate the causal benefits of heat alerts, and to identify characteristics of individuals or communities associated with the greatest reductions in morbidity or mortality following heat alerts. Specifically, using national claims data on deaths and hospital admissions among the large, geographically diverse population of >60 million US Medicare beneficiaries age ?65 years enrolled between 2001 and 2015, and on emergency department visits among >130 million participants of all ages from one of the nation?s largest health insurers, we propose to: Use novel machine learning methods to identify the heat metric(s) (e.g., heat index, ambient temperature, spatial synoptic classification, wet bulb globe temperature, absolute humidity) that best predict excess heat-related deaths, emergency hospitalizations, and emergency departments visits in each location, (Aim 2) estimate the causal effects of NWS heat alerts on rates of mortality, hospitalizations, and emergency department visits across the country and within groups stratified by health outcome, sex, and age group, and (Aim 3) assess how the benefits of heat alerts vary across characteristics of communities. Key innovations of this proposal include a very large sample size, geographic diversity encompassing the entire US, the assessment across multiple health endpoints and age groups, and the use of sophisticated methods in statistical learning and causal inference. Collectively, the findings from this proposal will provide meteorologists, public health and emergency management officials, and local policy-makers with critical information to better protect public health during extreme heat events and guide more targeted future research on strategies to mitigate the adverse health effects of heat.