Main

A growing body of work, synthesized most recently in the 2014 National Climate Assessment, documents how global climate change is altering US weather systems1. Average surface temperatures have increased, and recent decades have seen a rise in the frequency of extreme heat events and a decrease in cold spells2,3,4,5. Patterns in these trends vary across regions within the United States as well as across seasons. Thus far, temperatures have risen more in winters than in summers6,7, producing in most regions a climate that is more temperate than it had been previously. On the basis of findings in the economics literature indicating that Americans prefer to live in temperate climates8,9,10,11,12,13,14, we hypothesized that Americans have experienced an improvement in their local weather over recent decades, which may help explain low levels of concern about the problem among the mass public.

Work assessing the impact of climate change on weather typically uses spatially averaged changes in individual weather indicators, such as temperature or precipitation. Our approach is to examine change in the weather on the basis of population-weighted exposure and to evaluate this change in light of the trade-offs people have been shown to make among different weather conditions. Our focus on weather preferences requires analyses that account for seasonal variation in warming patterns. Moreover, recognizing that Americans have limited understanding of the likely effects of climate change, our analysis includes shifts in weather indicators demonstrated to be meaningful to the mass public in addition to those theorized by climate scientists to be most associated with global warming.

We investigated our hypothesis using 40 years of daily weather data reported in the Global Surface Summary of the Day (GSOD) and the US Historical Climatology Network (USHCN) data sets21,22 (Supplementary Table 1 and Extended Data Fig. 1). After assigning weather stations to counties based on station distance from county population centroids, we calculated annual means of weather indicators weighted by 2010 county population (Extended Data Table 1). We used constant population weights—rather than adjusting them for shifts over time in county population—to isolate the impact of weather trends from any changes in aggregate exposure attributable to population growth or migration23.

As objective measures of the average American’s weather preferences, we used estimates of revealed preferences produced by a sizeable literature analysing the effects of local climate on population growth8,9,10,11,12,13,14. These studies estimate change over time in population or net migration at the city, county or metropolitan-area level as a function of normal weather and a set of control variables. We used each study’s coefficients on weather-related variables to assign regression scores to our county-level weather data. With these values, we calculated an annual weather preference index (WPI) score for each county, an indicator of the average American’s revealed preference for different types of weather conditions. The WPI’s unit of measure is the ceteris paribus expected rate of population change associated with that year’s weather. In using these studies for our purposes, we assume that, all things being equal, Americans move to, and continue to reside in, places with local climates that they prefer. We further assume that aggregate tastes for weather were stable over the period between 1974 and 2013, and that controls in published models of weather-related population growth remove any confounding between weather tastes and the intensity of those tastes or ease of residential mobility. Our focus is only on preferences for long-term weather conditions, not on less predictable public responses to extreme weather events.

We report here on the WPI developed from estimates in a widely cited study that includes the most comprehensive models published in this literature14. Results were consistent using WPIs derived from parameter estimates in every other published study of this type9,11,12,13 (Extended Data Tables 2 and 3). In the model used here, county population growth from 1970 to 2000 is a function of five long-term normal weather indicators: January average daily maximum temperature; July daily heat index (derived from July average daily maximum temperature and July average daily mean relative humidity); July average daily mean relative humidity (entered into the model separately); annual precipitation; and the number of days on which precipitation occurs annually.

Calculations of annual trends for these five indicators (Fig. 1) confirmed that Americans experienced pronounced winter warming between 1974 and 2013: a regression of daily maximum January temperatures on year estimated a population-weighted average increase of 0.58 °C per decade. By contrast, summer warming trends were much less substantial: daily maximum July temperatures rose by only 0.07 °C per decade. Summer relative humidity rose in the mid-1990s but since then has declined in an abrupt pattern noted elsewhere24. As a result, summer heat-index values—calculated from temperatures and humidity—have remained relatively stable, with July heat index values rising only by the equivalent of 0.07 °C per decade. Annual precipitation rose slightly (increasing by 5.64 mm per decade) and occurred a bit more frequently (0.31 additional days per decade). In a complementary analysis, we estimated within-county change over the four-decade period with county-by-county regressions of each weather indicator on year (Table 1). These estimates reflected aggregate trends: since 1974, winter temperatures have become warmer for virtually all Americans; within-county change in the other indicators has been much less consistent.

Figure 1: Weather as experienced by the US population, 1974–2013.
figure 1

a–f, Annual means of county-level January mean daily maximum temperature (a), July mean daily maximum temperature (b), July mean daily relative humidity (c), annual precipitation amount (d), annual precipitation days (e) and WPI score (f), weighted by 2010 county population. Solid lines are locally weighted scatterplot smoothing (LOWESS) smoothers at 0.8 bandwidth; dashed lines display linear fits with shaded 95% confidence intervals.

PowerPoint slide

Source data

Full size image
Table 1 Change in county weather indicators per decade, 1974–2013
Full size table

We measured Americans’ preferences towards the weather they have been experiencing by using coefficient estimates from the population growth models to calculate WPI scores for each county in each year between 1974 and 2013. Details on our calculations appear in Methods. All published population growth models reveal the same preference structure with respect to seasonal temperatures: Americans appreciate warm winters and dislike hot, humid summers. The most comprehensive WPI model indicates that these preferences become more pronounced at extreme weather values14. Preferences about precipitation are mixed and have modest substantive impact, but Americans typically prefer less precipitation spread out over more days. Face validity of the WPI measure was confirmed by the geographical distribution of county mean scores (Fig. 2a).

Figure 2: WPI score by county, 1974–2013.
figure 2

a, Mean WPI, derived using calculations from equation (1) in Methods with 40-year mean values of weather indicators. b, Population growth rate equivalent change in WPI by decade, derived from county-by-county regressions of annual WPI on year as summarized in Table 1.

PowerPoint slide

Source data

Full size image

Analyses of the changes in these scores demonstrate that over the past 40 years, US weather has become more preferable for most Americans. The trend in the population-weighted annual means of county WPI scores (Fig. 1f) was in a steadily upward direction over the entire period at a population growth rate equivalent of 0.04 per decade. Estimated trends derived from county-by-county regressions of WPI on year indicated that improved weather is prevalent across most of the country (Fig. 2b), including areas covered by more weather stations, for which measurement is less susceptible to error (Extended Data Fig. 2). Currently, 80% of Americans live in counties where weather has improved over the past four decades (Table 1). Sensitivity analyses using a variety of measurement choices and WPI models produced estimates of the population exposed to improved weather ranging from 72% to 96% (Extended Data Tables 2, 3, 4, 5).

However, climate models project that these seasonal trends will eventually reverse, and that future US warming will be more severe in summer than in winter25. To estimate Americans’ preferences regarding these weather changes, we used projections from models in phase 5 of the Coupled Model Intercomparison Project (CMIP5) that have been downscaled and aggregated to the county level in the US Geological Survey’s National Climate Change Viewer26,27. We calculated the projected change in WPI scores from the observed 1974–2013 mean values to the year 2099 under two Representative Concentration Pathway (RCP) scenarios, fixing all non-temperature weather indicators at their means for the final 10 years of our study period. In the first scenario, greenhouse gas emissions diminish to the extent that total radiative forcing is stabilized shortly after the year 2100 (RCP4.5). In the second, emissions instead proceed at an unabated rate (RCP8.5)28. Under both scenarios, mean county WPI scores weighted by 2010 county population are projected to decline relative to the observed historical trend—and dramatically so under RCP8.5 (Fig. 3a), in which we estimate that 88% of the US public will experience less pleasant weather at the end of the century than in the recent past (Extended Data Table 6). The projected decline in weather pleasantness persists after adjusting estimated future WPI upward to reflect discrepancies between the observed temperature record and climate model simulations of historical temperatures7 (Fig. 3b).

Figure 3: Historical and projected trends in pleasantness of US weather.
figure 3

a, b, Black line is the smoothed historical trend (and grey area a 95% confidence interval around a linear fit) in WPI scores experienced by the US population, 1974–2013. Coloured areas depict projected scores under RCP4.5 (blue) and RCP8.5 (red). Dashed lines show the median forecast from 30 CMIP5 models, and shaded areas indicate the range from the 10th to 90th percentile values. a, Projections calculated directly from CMIP5 model output. b, Projections adjusted to reflect discrepancies between observed historical data and modelled output for the historical period. Analysis includes the contiguous US counties.

PowerPoint slide

Source data

Full size image

Our analysis is necessarily limited to evaluating population exposure to weather conditions in the United States, where ample data availability has made possible the research on long-term weather and population mobility that provides our measure of the average American’s preferences about weather. Although we cannot assume that the same preference structure holds outside of the United States, the temperate zone of the Northern Hemisphere is where people are most likely to share a similar taste for weather. Surface-averaged projections from CMIP5 models indicate that seasonal changes will vary in this zone29. Europe is similar to the United States in that summers are expected to warm as much or more than winters up to the year 2100; by contrast, in Canada and Russia—and to a lesser extent, China—weather is projected to become more temperate (Extended Data Table 7).

Our results have implications for how the complex phenomenon of climate change should be explained to the public. The scientific community’s emphasis on mean temperatures averaged over space and across seasons does not correspond with how people actually experience the weather30, which varies geographically and is interpreted through preference structures that value warmer temperatures differently at different times of year. In the United States, people’s experiences with daily weather since the time that they first heard about climate change have thus far mostly been positive. For the large portion of the public who form their beliefs and concern about climate change in part on the basis of their direct experience with weather15,16,17,18,19,20, the changes to which they have been exposed to date cannot be relied upon to provide the motivation needed to overcome apathy in responding to global warming. Our projections indicate that this may change in the decades to come as trends shift and weather becomes less pleasant. Both now and in the future, the public’s evaluations of the weather changes brought about by climate change should be taken into account by scientists and others seeking to explain the problem, mobilize concern and catalyse policy response.

Methods

Data sources

Our historical weather data came from two sources: the Global Surface Summary of the Day (GSOD) data set, version 7 (https://data.noaa.gov/dataset/global-surface-summary-of-the-day-gsod; accessed 22 July 2014)21, and the US Historical Climatology Network (USHCN), version 2.5 (http://www.ncdc.noaa.gov/oa/climate/research/ushcn/; accessed 22 September 2015)22, both maintained by the National Centers for Environmental Information. Block group-level and county-level US Census data, including geographical boundary data, came from the Minnesota Population Center’s National Historical Geographic Information System, version 2.0 (https://data2.nhgis.org/main; accessed 30 July 2014)31. We obtained county-level monthly temperature projections from the National Climate Change Viewer (NCCV) (http://www.usgs.gov/climate_landuse/clu_rd/nccv.asp; accessed through direct communication with J. Alder on 20 July 2015 and 3 December 2015)26,27, a US Geological Survey product that takes downscaled climate scenarios prepared by NASA (the National Aeronautics and Space Administration) and averages the 800-m gridded temperature data to the county level. Our international projections data came from the Royal Netherlands Meteorological Institute’s Climate Change Atlas (http://climexp.knmi.nl/plot_atlas_form.py; accessed 11 September 2015)32.

Measurement approaches

We limited our analysis to data from weather stations in the contiguous United States that operated continuously between 1974 and 2013. This 40-year period is long enough to minimize sensitivity to natural variability in weather data, and it begins at a point in time when the number of weather stations included in standard data sets and the completeness of the data they reported both increased. The timespan covers the entire history of Americans’ exposure to the idea of climate change, allowing us to track how weather has shifted during the time when the public might have perceived such shifts as attributable to climate change.

Temperature and humidity measurement

Daily weather data on temperature and humidity came from the GSOD data set, produced by the National Centers for Environmental Information from hourly weather station observations contained in the Integrated Surface Hourly data set21. Of the various land-based weather station data sets that offer daily summary data, GSOD is the only one that includes weather records necessary to measure a location’s relative humidity, which the urban economics literature has shown to be an important climate amenity driving regional population growth. Temperature and humidity records in our data set came from the GSOD’s daily station records of mean and maximum temperatures and mean dew point temperature (from which we calculated daily relative humidity and, in turn, daily heat index values). We included in the study only those GSOD stations reporting valid data on each of these weather indicators for at least 50% of the days in each of the 480 months of our study period, reducing the total number of stations in the analysis from 672 to n = 324 (Supplementary Table 1). Raising the threshold for valid data from 50% to 75% produced similar results (Extended Data Table 4a). Our final data set included a small number (n = 34) of stations that were relocated to nearby sites at some point between 1974 and 2013. In each case, the site location changed no more than 10 m in elevation and 0.1 decimal degree in latitude or longitude, and data reported by the relocated stations covered the entire study period with no more than a 15-day gap. Running our main analysis after omitting these relocated stations produced similar results (Extended Data Table 4b).

Precipitation measurement

Synoptic reporting of weather conditions in the GSOD data set introduces error in the calculation of daily precipitation indicators, so our main analyses employ daily precipitation records from weather stations in the USHCN, a designated subset of the National Oceanic and Atmospheric Association’s Cooperative Observer Program Network22. Sites are chosen for inclusion in the USHCN according to their spatial coverage, record length, data completeness, and historical stability. USHCN records are subject to rigorous quality control checks and have been demonstrated to be less error-prone than the GSOD33. We included in the study only those USHCN stations for which at least 90% of daily precipitation data were available in no fewer than 95% of the 480 months of our study period, reducing the total number of stations in the analysis from 1,218 to n = 601. This reduced the share of days in the USHCN data set with missing precipitation data to 1.2%. Because any time-dependent missing daily precipitation data could potentially affect our measurements of annual total precipitation and precipitation days, we used a procedure for simulating the occurrence of precipitation on missing data days that has been employed in leading research on over-time precipitation trends34,35. All simulations were conducted at the station-by-month level. For any day with missing precipitation data, we first employed a random-number generator to simulate whether precipitation occurred by using the observed frequency of precipitation within the station-month over the 40-year period of our analysis. We fitted a separate gamma distribution—which has been shown to realistically represent precipitation processes—to each station’s daily precipitation by month (for a total of 601 × 12 = 7,212 distributions), using only months for which the station had complete data. A random draw from the fitted distribution was then used to simulate missing daily precipitation for any day in the station-month on which precipitation was simulated to occur. As a robustness check, we carried out the same analysis using GSOD precipitation data; results were similar (Extended Data Table 4c).

Linking weather to population

To estimate population exposure to weather conditions, we used a method employed by health geographers that weights weather station observations based on their distance from population centroids of US counties36,37. Unlike other geographic units we might use to measure population exposure, counties are the smallest unit of geography for which boundaries remained almost entirely unchanged during the 40-year period. We located the population-weighted centroid for each county using block group population and boundary data from the 1990 Census, which was conducted approximately at the midpoint of our study period. We then assigned weights to GSOD and USHCN weather stations located within 160 km of a county’s population centroid based on the inverse of the station’s squared distance from the centroid. Counties with no weather station within 160 km from their centroids (n = 66 of the 3,103 counties in the contiguous United States) were dropped from the analysis; the counties remaining accounted for 98% of the 2010 contiguous US population. The median number of GSOD weather stations assigned to counties was 7; the median number of USHCN stations was 13 (Supplementary Table 1). Extended Data Figure 1 shows a map of the weather stations and counties in our data set.

Our findings are robust to other methods of matching weather conditions to the population. The results were similar when including only stations located within 80 km of population centroids (Extended Data Table 4d). In a separate analysis, we created a Voronoi polygon around each GSOD weather station with valid temperature, humidity and precipitation data (Extended Data Fig. 3) and then assigned population to the polygons by 1990 Census block groups. Using this method—which relies only on a single weather station’s data for each block group and does not include USHCN data—we find results similar to those in our main analyses (Extended Data Table 5).

We used daily data to calculate monthly averages by weather station for each of the weather indicators in our data set, yielding data at the station × year × month level, and then calculated annual values of January average daily maximum temperature, amount of precipitation, and number of days on which precipitation occurred. We used standard formulas to calculate July average daily mean relative humidity38 and July daily heat index39.

Because we were interested in Americans’ experience with the weather rather than distinguishing between short-term natural variability and long-term climate trends, we did not adjust the data to remove urban heat island effects. We also did not adjust for changes in instrumentation or observation routine. Research on the effects of these changes on temperature measurements suggests that the effects are modest and should bias results against our findings. The transition from afternoon to morning temperature observations and the adoption of electronic instruments both had the effect of recording lower maximum temperatures4,40,41. These effects do not seem to vary between winter and summer41. Because warming that has occurred over the last 40 years has been more pronounced and widespread in winter than in summer, instrument changes would result in understating the amount of January warming that has occurred and the corresponding increase in WPI. The effects of instrument changes on dewpoint temperature, and thus relative humidity measurements, are less systematic over time, but they have been detected at only a small proportion of stations24. Considering the modest role that relative humidity plays in our preference model and the limited evidence that instrumentation substantially affects measurements, we determined that data adjustments were not necessary.

We assigned weather station data to counties using the inverse-distance weights described earlier. Because our interest was in population exposure, we weighted our annual indicators by 2010 county population42. We used constant population weights—rather than adjusting them for shifts over time in county population—to isolate the impact of weather trends from any changes in aggregate exposure that are attributable to population migration or growth. As shown in Table 1, using population weights from 1970 rather than 2010 or unweighted values produced similar results. Summary statistics for the county-level weather indicators over our 40-year study period are reported in Extended Data Table 1; mean values of WPI by county are shown in Fig. 2a.

WPI measures

To produce the results reported in the paper, we used a WPI derived from a population growth model reported in a widely cited study14 (reported in ref. 14 table 3, model 6). The model includes both linear and quadratic weather terms to flexibly assess preferences about weather. In the model used here, county population growth from 1970 to 2000 is a function of five long-term normal weather indicators: January average daily maximum temperature (JAN_MAX); July daily heat index (JULY_HI); July average daily mean relative humidity (JULY_RH); annual precipitation (PRECIP_IN); and the number of days on which precipitation occurs annually (PRECIP_DAYS). Control variables include county geographical, coastline and topological features, baseline population density and total population, and baseline shares of county population employed in different sectors, including those tied closely to weather such as agriculture and transportation. Taking the reported coefficients estimating the partial relationships between the weather indicators included in the model and population growth, we calculated a WPI score for each county j in each year t using

All weather indicators are centred at their means, and thus the linear term coefficients can be interpreted as the effect of a one-unit shift in the indicator on WPI at the indicator’s mean value. As shown in equation (1), the analyses in this study (and all published studies from which we derive WPIs) were conducted using US conventional (imperial) units of measure. To comport with these studies, we employed imperial units in our calculations of all WPIs and then transformed results into SI units to report temperature and precipitation trends.

We checked the robustness of our finding by calculating alternative WPIs based on five other published analyses9,11,12,13 estimating the effect of climate amenities on local population growth. These studies employ simpler treatments of climate amenities, in some cases including only two or three indicators related to temperature, precipitation or humidity. All treat summer and winter temperatures separately. For each study, we developed a WPI based on reported coefficients on the study’s weather-related variables (Extended Data Tables 2 and 3). We then used our county-level weather data to calculate annual WPI scores for all US counties from each of these WPI formulas. We were able to measure all weather-related variables at the county level over the entire 40 years except for sunshine hours, a variable that appears in two of the models12,13. Our estimates therefore assume no long-term change in the amount of sunshine experienced by individual counties.

Future warming projections

County-level temperature estimates for the RCP4.5 and RCP8.5 emissions scenarios came from the NCCV, which uses NASA Earth Exchange Downscaled Climate Projections (NEX-DCP30) data to project future changes in climate and water balance for states, counties and hydrologic units26,27. The NEX-DCP30 data set statistically downscales projections from 33 models included in the 5th Climate Model Intercomparison Program (CMIP5) to an 800-m grid. The NCCV includes 30 of the 33 models that cover both emissions scenarios and creates area-weighted averages at the county level. Consistent with the NCCV’s presentation of these data, we have examined projections over three time periods—2025–2049, 2050–2074, and 2075–2099—under the two emissions scenarios, comparing mean WPI values within each time period to the observed 1974–2013 means for every county (Extended Data Table 6). Because of data availability, we estimated changes in WPI based only on changes in summer and winter temperatures, weighting counties by their 2010 populations and fixing other weather indicators at their means for the final 10 years of our study period. Consistent with studies of CMIP5 model performance, we found discrepancies between the observed temperature record and hindcasts yielded by climate models7,43, with the effect that simulated temperature data for the 40-year historical period of our study produce average WPI scores that are lower than those calculated using observed data. Recognizing that discrepancy between modelled and observed temperatures may persist into the future, we performed an additional analysis in which we regression-adjusted projected future WPI scores under all time frames and scenarios to account for the discrepancy. The adjustment was performed by regressing observed annual WPI on simulated WPI derived from the CMIP5 models for the 1974–2013 period, yielding

We adjusted the projections in Fig. 3a using predictions from this model and display the adjusted projections in Fig. 3b.

Northern Hemisphere temperate zone projections

We obtained these projections using the Royal Netherlands Meteorological Institute’s Climate Change Atlas, which provides CMIP5 climate model output for a variety of countries, seasons, time periods and scenarios through its web-based interface32. For each country, we obtained the mean surface-averaged projections of change in maximum winter and summer near-surface temperatures in the 2075–2099 period under RCP4.5 and RCP8.5 with respect to the reported mean of those observed in the 1974–2013 period (Extended Data Table 7).