Study population and setting

We assessed heat exposure in the Community of Mines Study, an observational study with enrollment between 2014 and 2017 in San Diego County, California. San Diego County has a population of about 3.3 million per the 2020 census [28]. The urban area of San Diego County has an arid Mediterranean climate with average daytime high temperatures in the mid 70 degrees Fahrenheit in the summer and in the mid 60 degrees Fahrenheit in the winter [29]. Annual precipitation totals less than 12 inches, with most rainfall occurring during the cooler months. The county includes mountains and sparsely populated areas in its eastern region, which receives snowfall [28].

The Community of Mines protocol is available elsewhere [30], and aspects of the study have been described in other research [20, 31]. Briefly, 602 adults aged 35-80 years old (mean age=59 years) who had lived for at least 6 months in a selected census block group completed the study. The study population is comprised of 56% women and is ethnically (42% Hispanic/Latino) and socioeconomically diverse (21% income $30,000 or less; 60% $55,000 or more) [20, 31]. Study ethics approval was obtained from UCSD IRB protocol #140510. Signed informed consent was obtained from all participants who enrolled in the study.

Participant GPS measures

Participants were instructed to wear Qstarz GPS devices (Qstarz International Co. Ltd, Taipei, Taiwan) during waking hours to measure their movement [30, 31]. The GPS observations, which we also call pings, each had a latitude value, a longitude value, and a time stamp. GPS data were processed, cleaned, and aggregated to the minute level, as detailed in the appendix of Jankowska et al. [31].

Of the 602 participants, 599 participants wore devices for at least one valid day, where a valid day is defined as having at least 10 h of wear time [31]. We assume pings that occurred one minute apart occurred while the GPS was being worn. Participants wore GPS devices on average for 13.8 days and for 13.3 h per day (median=13.8 h per day) at various two-week intervals between 2014-10-17 and 2017-10-06. The average GPS follow-up of about 14 days corresponds to the minimum number of days recommended to measure activity spaces [14]. Among those pings that did not occur one minute apart, the per-person mean duration between pings was 10.7 h (median=10.2 h), and there were on average 14.1 such pings per participant. We assume these pings with an interval longer than one minute represent non-wear time between going to bed and waking the next day.

We created activity paths [16] with each participant’s GPS data, ordering each participant’s GPS pings in time. We excluded points with as-the-crow-flies speeds exceeding 100 miles per hour (0.04% of person-time recorded) from that point to the next point in time. Upon inspecting these segments visually, they appeared to either represent flights, or implausible or erroneous GPS bouncing. We also excluded points outside San Diego County (1.1% of total person-time).

To descriptively characterize the activity space of participants, we measured their average distance traveled per day, where distance traveled is the cumulative distance between their GPS pings for that day. We also calculated their average daily time spent at home (home defined below) within a 200 m buffer of their home, including non-wear time.

Heat indicators and data sources

We used two gridded measures of heat: land surface temperature (LST) and air temperature (Table 1). LST is defined as “how hot the surface of the Earth would feel to the touch in a particular location” [32]. We downloaded LST data from Landsat 8-9 Collection 2 Level 2 Band 10 ST [33] from the U.S. Geological Survey’s EarthExplorer data portal (https://earthexplorer.usgs.gov; accessed April 29th, 2024). The data are available every 15 days at a spatial resolution of 30 meters. We examined 69 images over San Diego County (Path 40, Row 37) between 2014-09-30 and 2017-10-08. We removed images with more than 3% cloud cover, resulting in a total of 35 images. Specific image identifiers and dates appear in eTable 1.

Table 1 Description of heat indicators.

The important advantage of Landsat-measured LST for this research is its high spatial resolution (30 m), which is fine enough to measure a wide street [34]. LST has limitations, however, as a direct measure of human heat exposure [10]. We thus also gathered gridded data on air temperature from temperature from gridMET [11], which may more closely reflect human thermal comfort than LST [5, 10]. The gridMET air-temperature data are also available at a higher temporal resolution than LST (daily) and allow consideration of intra-diurnal variation in heat, as the data include both daily maximum and minimum air temperature. The gridMET air-temperature has a coarser spatial resolution than the Landsat LST data, however. Its spatial resolution of 4 km corresponds roughly to a neighborhood-scale air-temperature measure [35]. We gathered 1054 days of gridMET data for both maximum and minimum air temperature in San Diego County.

The distributions of LST, maximum air temperature, and minimum air temperature over time within a 200 m buffer of the area traveled by all of the study participants (regardless of time elapsed therein) appear in Table 1 and Fig. 1. The average LST over this area during the study period ranged from a low of 286 K to a high of 319 K, with an average within-day spatial variation (standard deviation) of 8.36 K. Maximum daily air temperature ranged over time from an average (over pixels on that day) of 282 K to 314 K, and the corresponding average minimum daily air temperature ranged from 273 K to 295 K. Within-day spatial variability was lower for air temperature (average standard deviation of 3.75 K and 3.02 K, respectively for maximum and minimum) than for LST (8.36 K).

Fig. 1: Distribution of the two heat indicators over the area traveled by all study participants over the study period.

The points in panel A depict the median land surface temperature over the area traveled by study participants on each day with imagery. Lines between points in panel A are imputed values to facilitate visualization. The dark lines in Panel B depict the daily median maximum air temperature and daily median minimum air temperature over the area traveled by study participants. The shaded regions range from the 5th through 95th percentiles of each measure over the area traveled by study participants on that day. Max maximum, min minimum.

Assessment of heat exposureMobility-based

We calculated two mobility-based measures of heat exposure for both LST and air temperature: a time-weighted mean (denoted as \(}_-}\) to facilitate notation below) and the highest average value of 10 min intervals of their activity paths (\(}_-}\)). We calculated the time-weighted mean as a measure of cumulative exposure and the maximum over 10 min intervals as a measure of the highest acute exposure.

Mean

To calculate mean mobility-based exposure to LST, we began by extracting LST values at the point location of each GPS ping for each participant, using both the image that most closely preceded the date of the GPS ping (the index date) and the image that most closely followed that day. For example, if a participant recorded activity on March 15th, 2015, we extracted the LST value from March 9th, 2015, and March 25th, 2015 for each ping recorded on March 15th, 2015. Then, for each ping on each index date, we calculated the weighted average of its two LST values, weighting the two values by the inverse of the time between the index date and the image date so that the LST of the image nearer in time would receive a stronger weight in the average. Specifically, if \(}_}\) is the weighted average between these two days, then \(}_}=\frac}_}* }_}+}_}* }_}}}_}+}_}}\), where \(}_}=\frac}-}\) and \(}_}=\frac}-}}\). For each participant, we then calculated their weighted average mobility-based LST exposure, weighting the value of each point’s value by its elapsed time until the next one, including overnight time.

To calculate mean mobility-based exposure to air temperature, we broadly followed the same method with some differences because air-temperature is available at a daily temporal resolution. To roughly estimate minute-level air-temperature exposure within day, we averaged the maximum and minimum air temperature values at the location of each GPS ping based on the ping’s time of day, assuming the maximum air temperature occurred at 3 pm and that the minimum air temperature occurred at that day’s sunrise. We calculated sunrise time for all days during the study period for San Diego, California using NOAA’s calculator [36]. Specifically, again taking an inverse-distance-in-time-weighting approach, if \(}_}\) denotes the weighted average between the maximum and minimum air temperatures at a time of day (the index time) on the index date, then \(}_}=\frac}_* }_+}_* }_}}_+}_}\), where \(}_=\frac}} \; }}\; }}\; }}3}}+}}\; }}\; }}\; }}\; }}}}}\; }}\; }}\; }}3}}}}\), and \(}_=\frac}}\; }}\; or\; }}3}}+}}\; }}\; }}\; }}\; }}}}}\; }}\; }}\; }}\; }}}}\). Time until or after 3 pm is the elapsed time until 3 pm if the index time is before 3 pm or the elapsed time after 3 pm if after. The time since or until sunrise is the elapsed since sunrise if the index is before 3 pm or until sunrise if after 3 pm.

We then calculated each participant’s mobility-based mean air-temperature exposure as the weighted average of these estimated air-temperature values over their duration of GPS follow-up, weighting each GPS ping’s air temperature by the elapsed time until the next GPS ping, again allowing that the elapsed time until the next GPS ping may include non-wear time between going to bed and waking the next day.

To assess the impact of our decision to include non-wear time in the time-weighted exposure measures, we also calculated the weighted mean LST and air temperature excluding non-wear time.

Maximum

To assess acute exposure to the heat indicators, we grouped participant’s GPS points into ordered sequences of 10 min intervals and took the mean exposure value of each 10 min interval over their constituent point-level values. We grouped points into 10 min intervals because heat stroke can develop in 10 min [24]. We then found the maximum (denoted as \(}_-}\)) of each participant’s 10 min-interval averages.

Residence-based Mean

To assess residence-based exposure to these heat indicators (\(}_-}\)), we first defined participants’ home location. To define their home location, we drew a grid of hexagons over San Diego County, each with an area of 1000 m2 (each side ~ 62 m). We defined home as the centroid of the hexagon where the participant spent the most elapsed time, where elapsed time is defined by the difference between sequential GPS pings, including the difference between the last ping of one day and the first of the next. We inferred home location as the location where the participant spent the most time because we observed that many study participants were rarely at the home address that they provided in the survey. Inferring home location based on elapsed time is consistent with research using mobile-phone-based location data that has inferred home location as the location where their device spent the most time overnight [37,38,39]. For 587 (98%) of the 599 participants, the hexagon where the participant spent the most elapsed time overnight (8 p.m.—6 a.m.) was either the same as (n = 578) or shared an adjacent edge with (n = 9) the hexagon where they spent the most total elapsed time. We chose to use the hexagon where the most total elapsed time was spent rather than that where the most time was spent at night because for some of the 2% (n = 12) of participants for which the two locations differed, the hexagon where the most time was spent at night was either a place of employment (suggesting night-shift work) or a campground (suggesting the study period may have coincided with a vacation).

To calculate the mean residence-based exposure to each indicator for each participant, we followed the same approach as for mobility-based means, but instead of using the actual location of each GPS ping, we extracted the heat-indicator value corresponding to the participant’s home location at the time of each GPS ping. We then summarized the information the same way as above.

Maximum

The approach for calculating maximum residence-based exposure (\(}_-}\)) for LST differed from that of air temperature because of the differences in the temporal resolution of these indicators. Each participant’s maximum residence-based LST exposure is the maximum of their day-level residence-based LST exposure during their days of GPS follow-up. In contrast, the maximum residence-based exposure to air temperature is the maximum value of each participant’s set of 10 min intervals, where the location used to calculate the interval-specific means is always their home location.

Analysis

We compared mobility-based means with residence-based means and mobility-based maximums with residence-based maximums using differences: \(}_}=}_-}-}_-}\); \(}_}=}_-}-}_-}\). We also compared means using Pearson correlations.

To explore variation in the comparison measures by mobility patterns and socio-demographic characteristics, we stratified results by tertiles of average daily distance traveled (5.63–41.5 km; 41.5–65 km; 65–368 km), age (35–50 years old; 51–65 years old; 66–80 years old), sex (female; male), income (<$30 k; $30 k–$55 k; $55k + ), and race and ethnicity (white; Latino; Asian; Black; Native American or Pacific Islander).

We plotted histograms to visually assess the distribution of the difference measures.