Climate change may alter human physical activity patterns

Regular physical activity supports healthy human functioning1–3. Might climate change—by modifying the environmental determinants of human physical activity—alter exercise rates in the future4? Here we conduct an empirical investigation of the relationship between meteorological conditions, physical activity and future climate change. Using data on reported participation in recreational physical activity from over 1.9 million US survey respondents between 2002 and 2012, coupled with daily meteorological data, we show that both cold and acutely hot temperatures, as well as precipitation days, reduce physical activity. We combine our historical estimates with output from 21 climate models and project the possible physical activity effects of future climatic changes by 2050 and 2099. Our projection indicates that warming over the course of this century may increase net recreational physical activity in the United States. Activity may increase most during the winter in northern states and decline most during the summer in southern states. Obradovich and Fowler use data on participation in physical activity from 1.9 million US residents from 2002–2012, coupled with daily temperature data, to show that unmitigated climate change is likely to alter future patterns of physical activity.

activities? Second, do the effects of temperature on physical activity vary by demographic factors like weight and age? Third, how might climate change alter the distribution of physical activity throughout the months of the year in the future? Finally, how might the future effects of climatic changes on physical activity vary geographically?
To investigate whether outside weather conditions alter the propensity to engage in physical activity, we constructed a dataset of individuals' reported monthly participation in recreational physical activity linked with historical meteorological data. Our individual response data come from the Center for Disease Control and Prevention's Behavioral Risk Factor Surveillance Survey (BRFSS) pooled over the period 2002-2012. Randomly selected respondents answered the following question: "During the past month, other than your regular job, did you participate in any physical activities or exercises such as running, calisthenics, golf, gardening, or walking for exercise?" This question includes the most common leisure-time physical activities engaged in by adult residents of the United States 13 . Questions from the BRFSS have been assessed for validity 14 and reliability 15 and are largely consistent with other health-related activity measures, including on dimensions of physical activity 16 . Furthermore, this specific question is used in widely cited public health studies relating to physical activity 17,18 .
We combine individual responses to this question-marked by interview day and geolocated to the city level-with station-level daily temperature and precipitation data from the National Centers for Environmental Information's Global Historical Climatology Network -Daily (GHCN-D) 19 as well as humidity and cloud cover data from the National Centers for Environmental Prediction (NCEP) Reanalysis II project 20 (see Supplementary Information). Of note, our analysis is also robust to the use of gridded daily weather data from the PRISM Climate Group 21 (see Supplementary Information). Our theoretical relationship of interest is the effect of meteorological conditions on the probability of being physically active. We empirically model this relationship as: In this pooled cross-sectional linear probability model fitted using least squares, i represents individuals, j cities, s seasons and t calendar days. Our dependent variable Y ijst is binary and indicates whether respondents were physically active over the thirty days before their interview day. Our main independent variable, temp jst , represents the thirty-day average of daily maximum temperatures over the same thirty-day window as the physical activity reported by each respondent. Our relationship of interest is represented by f(temp jst ), which provides separate indicator variables for each 1 °C monthly average maximum temperature bin, enabling flexible estimation of a non-linear relationship between temperature and physical activity 22,23 (of note, our results are robust to using bins of one, two and five degrees Celsius; see Supplementary Information). We omit the 28-29 °C indicator variable, and therefore interpret our estimates as the change in probability of being physically active associated with a particular temperature range relative to the 28-29 °C baseline.
In addition, the Zη term in Equation 1 represents an additional set of meteorological variables that include the number of precipitation days during the thirty-day window, average temperature range, average cloud cover and average relative humidity. We include these other meteorological variables as their exclusion might bias our estimates of the effect of included meteorological variables 24,25 (although the magnitude of the estimates of f(temp jst ) are mostly unaffected by the exclusion of these variables, see Supplementary  Information).
Unobserved characteristics may influence participation in physical activity. For example, people may exercise more in cities with better infrastructure or on days when they are more likely to have leisure time. To be sure that geographic and temporal factors such as these do not interfere with our estimates, we include γ t and ν js in equation (1). These terms represent calendar date and city-byseason indicator variables that account for unobserved characteristics constant across cities and days as well as seasonal factors that might vary differentially by city 26 . Of note, our results are robust to varying the specification of these controls (see Supplementary  Information). Our empirical identifying assumption, consistent with the literature 25,[27][28][29] , is that meteorological variables are as good as random after conditioning on these fixed effects. The estimated model coefficients can therefore be interpreted as the effect of meteorological conditions on reports of participation in recreational physical activity 24,30,31 .
Because our estimation procedure uses exogenous city-level variation in temperature to predict individual-level outcomes, we account for within-city and within-day correlation 32 by using heteroskedasticity-robust standard errors clustered on both city and day 33,34 . Finally, we omit non-climatic control variables from equation (1), because of their potential to generate bias-a phenomenon known as a 'bad control' 25,31 -in our parameter of interest (nonetheless, results are robust to the inclusion of common demographic covariates; see Supplementary Information). Figure 1a, which presents the estimates of f(temp jst ) from equation (1), shows that the probability of participation in physical activity increases up to 28-29 °C and decreases past 36 °C, although effects at higher temperatures are estimated with greater error. Average maximum temperatures around 0 °C produce a reduction of approximately seven percentage points in the probability of being physically active compared to the 28-29 °C baseline (coefficient = − 7.249, P < 0.001 , n = 1,941,429)). Average maximum temperatures above 40 °C reduce the probability of physical activity, although the effect is only about one-half as large as the effect of freezing temperatures and is only significant at the P = 0.10 level (coefficient = − 2.815, P = 0.079, n = 1,941,429).
To indicate the magnitude of our estimated relationship, a + 2 °C shift from an average monthly temperature of 26-27 °C to an average of 28-29 °C, if extrapolated across the current population of the United States, would produce over six million additional personmonths (that is, an individual engaged in any physical activity over the course of a calendar month.) of physical activity annually.
However, temperature alone may not fully capture the effect of heat stress on participation in physical activity [35][36][37][38] . Heat stress indices, composites between temperature and relative humidity, might indicate that participation in physical activity declines in instances of both high temperature and high humidity 39 . In order to investigate whether heat stress metrics provide substantially different results compared to temperature alone, we examine the results of estimating: Where we use the same set of control variables as in equation (1) The results of estimating this equation (Fig. 1b), closely mirror the results of colder maximum temperatures (Fig. 1a) and we observe no significant decline in physical activity participation at high levels of combined heat and humidity (see Supplementary Information for the tables associated with this regression). These results, coupled with the insignificance of the marginal effects of relative humidity (Fig. 1d) indicate that the effect we observe is primarily driven by ambient temperatures (for results using alternative heat stress indices 37 , see Supplementary Information). Furthermore, we observe a small, mostly linear effect of added precipitation days on physical activity participation (Fig. 1c), whereby more than 20 days with measurable precipitation in a month lead to a reduction of approximately one percentage point in the probability of participation in physical activity (coefficient = − 1.144, P = 0.009, n = 1,941,429). We observe no significant relationship between average cloud cover and participation in physical activity. The above estimates represent an average effect of temperatures on physical activity over the course of a full year across all respondents in our sample. However, because individuals vary in their sensitivity to heat and cold 42 , we might expect to observe heterogeneous responses to changes in temperature. For example, individuals with higher body mass index (BMI) may experience greater physical stress associated with hot temperatures. In addition, older individuals, owing to a less robust thermoregulation ability 43 , may similarly experience more acute reductions in physical activity due to extreme heat. This leads us to our second question: do the effects of temperature on physical activity vary along notable demographic factors like weight and age?
In order to examine whether heavier respondents are more sensitive to temperature, we stratify our sample by BMI and estimate Equation 1 for each sub-sample 28 . Figure 2a shows that the negative effect of temperatures greater than 40 °C on the probability of physical activity is greatest for obese (BMI ≥ 30) respondents (coefficient = − 6.567, P = 0.002, n = 466,754). This is over seven times the effect observed among normal weight (BMI < 25) adults (coefficient = − 0.866, P = 0.549, n = 711,662) (see Supplementary Information).
We repeated this procedure to examine whether older respondents are more sensitive to temperature. Figure 2b shows that the negative effect of temperatures greater than 40 °C on the probability of physical activity is greatest for those 65 years of age or older (coefficient = − 7.026, P < 0.001, n = 517,700). This is over four times the effect observed among those under 40 years of age (coefficient = − 1.574, P = 0.569, n = 456,383). Thus, our results suggest that the physical activity rates of both obese and elderly individuals may be most susceptible to higher ambient temperatures.
Our historical data indicate that past temperatures have altered historical physical activity patterns in meaningful ways. Furthermore, climate change is likely to produce positive shifts in monthly temperature distributions in the future 44 (see Fig. 3b). As can be seen in Fig. 3a, most historical temperatures fall below 28-29 °C, the temperature range associated with peak physical activity in our sample. Positive shifts in temperatures below 28-29 °C might increase physical activity, whereas shifts that amplify the incidence of markedly hot temperatures may reduce physical activity. Combining these insights leads us to our third question: how might climate change alter the distribution of physical activity in the United States throughout the months of the year in the future?
To examine this question, we calculated the projected average monthly maximum temperatures for 2050 and 2099 from the NASA Earth Exchange's (NEX) bias-corrected, statistically downscaled daily maximum temperature projections 45 , which were drawn from 21 of the CMIP-5 ensemble models 46 run on the RCP8.5 high emissions or 'business as usual' scenario 47 . We couple these predicted temperatures with our historical estimate of the relationship between average maximum temperatures and participation in physical activity-using a spline regression model that closely matches the results from equation (1)-to calculate a forecast of possible physical activity alterations due to climate change for each month of the year for each city across each downscaled climate model (see Supplementary Information).

Δ
=ˆ and for the effect from 2010 to 2099 (Δ Y m2099 ) as: Where m indicates the month of year, k the 21 specific climate models, j the city, and t the day of year. In addition, temp kjmt is our measure of the thirty-day average maximum temperatures, as calculated in equation (1) and f () represents the fitted spline function from our main forecast model (see Supplementary Information).
Notably, this estimation procedure allows us to incorporate uncertainty regarding the underlying climatic forecasts into our physical activity predictions 27 . It also allows us to account for estimation uncertainty in our model of physical activity. Figure 3 shows the monthly forecast results for 2050 (Fig. 3c) and 2099 (Fig. 3d). The bar for each month represents an average prediction across each of the 21 climate models, across each of the cities in our analysis and across the days in that month. The error lines on these bars represent the full range of the monthly estimates across the 21 downscaled climate models and incorporate the 95% confidence interval of the estimated historical relationship between temperature and physical activity. As can be seen in these figures, the probable temperature changes produced by climate change may increase physical activity the most in traditionally cooler months of the year. In the months of June, July, and August, climate change by 2099 may reduce physical activity on average. Taking a yearly Additionally, temperature alterations associated with climate change-and therefore the potential effects of climate change on physical activity-are likely to vary spatially across the United States. To investigate the geographic distribution of potential modifications in physical activity due to climate change, we take the ensemble average of the 21 NEX downscaled climate models for each of 2010, 2050 and 2099. We then take the monthly average of maximum temperatures for each approximately 25 km × 25 km grid cell in the continental United States in each year. For the 2050 forecast, we assign to each grid cell the predicted net monthly difference in physical activity between 2010 and 2050. For the 2099 forecast we assign to each grid cell the predicted net monthly difference in physical activity between 2010 and 2099. Figure 4 shows that most areas of the United States may see net increases in physical activity as a result of climate change this century, whereas the southernmost areas may experience some net decreases in physical activity (see Supplementary Information).
Finally, smoothing geographic forecasts across the full year (as in Fig. 4) masks the temporal heterogeneity associated with our physical activity forecasts. Figure 5 shows the geographic forecasts for each month in the years 2050 and 2099. Future winter months, especially in northern areas of the United States may see the greatest increases in physical activity rates, whereas future summer months, particularly in the southern areas of the United States may see net decreases in physical activity rates.
Historical data demonstrate a robust link between temperature and human physical activity. The effects of cold historical temperatures on reduced participation in physical activity are highly statistically significant and substantially large in magnitude. Moreover, in both our city-level and geographic forecasts, we predict that much of the United States will experience increased physical activity due to future climatic changes. These increases occur primarily during the cooler months of the year, with summer months-especially in the southern parts of the US-probably showing potential declines in future physical activity.

NATURE HUMAN BEHAVIOUR
There are several considerations important to the interpretation of these results. First, although we have data on millions of individuals' reported monthly participation in physical activity, optimal data would also contain measurements of each individual's participation in daily physical activity and the intensity of such activity. Second, because respondents are geolocated to the city-level, measurement error may exist between the temperatures observed at weather stations and the temperatures respondents actually experienced, possibly attenuating the magnitude of our estimates 48 . Third, our analysis is conducted on a randomly sampled, pooled cross-section of respondents. An ideal source of data would track the same individuals over time to enable controlling for individual-specific characteristics. Fourth, our data are restricted to observations from one country with a temperate climate. It is critical to repeat this analysis where possible in countries with warmer average climates 49 and lower prevalence of air conditioning 50 , as they may see net reductions in physical activity due to climate change. Fifth, in this analysis we focus on recreational physical activities. However, because occupational physical activity may be less discretionary, the effects of climate change on occupational physical activity deserve further scrutiny 38,51 . Sixth, because the hottest historical regions in our data also tend to have lower humidity, our analysis may understate the potential for amplified future levels of heat stress to reduce physical activity. Future studies should investigate this question. Finally, it is possible that humans may adapt technologically and physiologically to warmer climates with behaviours not seen in the historical data 42,52,53 .
Ultimately, most of social impacts of climate change are likely to be negative 30 . Climate change may reduce economic output 24 , amplify rates of conflict 31 , produce psychological distress 54 , increase exposure to the social effects of drought 55 and increase heat-related mortality and morbidity [56][57][58][59][60][61] , among other illnesses. However, climatic changes are unlikely to be uniformly costly to society, and it is important to investigate both costs and benefits. Here we uncover a possible beneficial effect of climate change for the United States. If observed temperature-activity relationships from the recent past persist, further climate change may increase nationwide net participation in recreational physical activity during many months of the year, in turn magnifying many of the physiological and psychological benefits of exercise. However, adaptations to changing temperature distributions or interactions with potential pernicious effects of climate change 30 -like increased stress and anxiety 54 -may counteract these effects. The more we know about the full range of potential climate impacts, the better we will be able to prepare for what is likely to be humanity's greatest challenge in the 21st century.