What weather variables are important in predicting heat-related mortality? A new application of statistical learning methods

Abstract

Hot weather increases risk of mortality. Previous studies used different sets of weather variables to characterize heat stress, resulting in variation in heat-mortality associations depending on the metric used. We employed a statistical learning method - random forests - to examine which of the various weather variables had the greatest impact on heat-related mortality. We compiled a summertime daily weather and mortality counts dataset from four U.S. cities (Chicago, IL; Detroit, MI; Philadelphia, PA; and Phoenix, AZ) from 1998 to 2006. A variety of weather variables were ranked in predicting deviation from typical daily all-cause and cause-specific death counts. Ranks of weather variables varied with city and health outcome. Apparent temperature appeared to be the most important predictor of heat-related mortality for all-cause mortality. Absolute humidity was, on average, most frequently selected as one of the top variables for all-cause mortality and seven cause-specific mortality categories. Our analysis affirms that apparent temperature is a reasonable variable for activating heat alerts and warnings, which are commonly based on predictions of total mortality in next few days. Additionally, absolute humidity should be included in future heat-health studies. Finally, random forests can be used to guide the choice of weather variables in heat epidemiology studies.

Keywords: Absolute humidity; Heat; Mortality; Random forests; Temperature; Weather.

Figure 1

Importance of weather variables in predicting deviation from typical daily total mortality counts as the response variable for four U.S. cities Note: 1. Importance of weather variables is quantified as the average percent increase in mean squared error; 2. In this analysis, the random forests approach took 20,000 bootstrap samples of summertime (May 1^st to September 30^th) weather and mortality data from one of four cities, and each sample results in a tree. For each bootstrap sample, mean squared error for a variable was calculated by comparing the predictions from the permuted sample of that variable to those from the unpermuted sample of that variable. A higher average percent increase in mean squared error for a variable suggests that it is more important in predicting outcomes.). 3. This figure shows importance scores of the first 30 variables among all 45 variables. 4. TMP, temperature; DPT, dew point; AT, apparent temperature; STP, barometric pressure; AH, absolute humidity; min, minimum; max, maximum; me, mean; lag 1 or 2, one day or two days before deaths occurred.

Figure 2

Importance of weather variables in predicting deviation from typical daily cause-specific and all-cause counts as the response variable in Chicago. Otherwise as Figure 1. Note: CHF, congestive heart failure; COPD, chronic obstructive pulmonary disease; CVD, cardiovascular diseases; MI, myocardial infarction.

Figure 3

Importance of weather variables in predicting deviation from typical daily cause-specific and all-cause counts as the response variable in Detroit. Otherwise as Figure 2.

Figure 4

Importance of weather variables in predicting deviation from typical daily cause-specific and all-cause counts as the response variable in Philadelphia. Otherwise as Figure 1.

Figure 5

Importance of weather variables in predicting deviation from typical daily cause-specific and all-cause counts as the response variable in Phoenix. Otherwise as Figure 1.