How will climate change shape climate opinion?

Of the articles reviewed, 32 (44%) examine the association between subjective experience with abnormal weather conditions or extreme events and climate opinion. Studies that examine subjective experience ask participants to self-report whether they have personally experienced a specific weather-related phenomenon, trend, or extreme event [13–15, 24, 32, 36, 38, 46–50], whether they have experienced its effects [21, 25, 51–54], or if they have generally experienced unusual weather patterns [55, 56]. Alternatively, some studies ask participants if they have personally experienced the effects of climate change itself [37, 57–59]. In some studies, survey questions about personal experience are also combined with external observational data on weather conditions or trends [15, 20, 32, 36, 38, 39, 46, 50, 60, 61].

Based on this literature, there is a fairly robust relationship between perceived or subjective experience and related climate opinion questions. Of 32 papers that address the association between self-reported experiences and climate opinion, 27 (84%) find evidence for such a relationship. For example, a large study of surveys from 119 countries found that perception of rising temperatures at the local level was an important predictor of climate risk perceptions [14]. In general, these studies suggest that people who think they have experienced the effects of global warming (or have experiences of extreme weather) also tend to believe that global warming is happening and to be more concerned about it. However, these studies alone do not establish a causal relationship between experience and opinion. Despite the associations identified in these papers, climate opinions may also shape the weather-related experiences that people self-report, as we discuss in section 3.1.4 below.

Of the reviewed articles, 51 (70%) examine the association between externally measured weather variables or extreme events and climate opinion. Of these, 46 (90%) used some measurement of temperature or temperature-related extreme events (e.g. heat or cold waves). Temperature is typically operationalized as either the absolute air temperature over a certain time period, or a temperature anomaly: the difference in the absolute temperature from a long-term base period. A small number of studies also examine trends in temperature over a set time period. The prevalent use of temperature as an independent variable likely arises from its conceptual salience with global warming or climate change, as well as the ease with which temperature observations from weather stations or gridded data can be joined with survey responses by location.

Among studies that have tested temperature as an independent variable predicting climate opinion, some provide evidence for a 'local warming' effect (e.g. [5, 6, 10, 11, 15, 18, 62]), showing that elevated temperatures in the short term (daily to monthly) are associated with increased concern about climate change, belief that it is happening and human-caused, or policy support. For example, in one analysis, temperature anomalies in the week prior to a survey predict opinion about whether global warming is happening [9]. These same studies indicate that the magnitude of the effect is relatively low and not persistent as the experience of elevated temperatures recedes over time (e.g. [9, 63]). Other studies that have examined associations between short-term temperature and climate opinions have found no effect [7, 19, 64].

There are mixed findings among studies that have investigated the relationship between longer-term temperatures or temperature trends and public opinion. The largest recent study of US survey data (N = 348, 500) from 1999 to 2017 finds that higher average annual temperatures at the state level are associated with a small but robust increase in worry about climate change [65]. This finding is reflected in similar studies that have examined seasonal-to-annual temperatures, e.g. [7, 8, 17, 19, 20, 63, 64, 66–68]. For example three studies find that 10 year summer temperature trends are positively related to beliefs about human-caused global warming in the US [19, 64, 66]. In addition, an index representing the ratio of previous record high temperatures to record low temperatures over several years (with more record high temperatures expected under a warming climate) is associated with estimates of county-level climate opinions in the US [12, 69]. However, other studies that have focused on longer-term temperatures or trends have found little to no effect on climate opinions [28, 30, 32, 34, 70, 71].

The few longitudinal studies that have examined how changes in climate opinions may be associated with local temperature observations have generally found little to no effect. For example, monthly US temperature anomalies do not predict changes in opinions over 2.5 years from 2008 to 2011 [31], and there were minimal effects of monthly temperature anomalies in a large panel analysis of Cooperative Congressional Election Study data over a four-year period from 2010 to 2014 [72].

While temperature is the most commonly used observed weather variable, other studies examine precipitation observations (e.g. [10, 64, 65]) or derived climatic variables (such as drought indices) (e.g. [30, 67]). Of studies that use objective weather data, 21 (41%) use data related to precipitation. In contrast to temperature, most studies have identified little to no association between precipitation or related variables alone and climate opinions [10, 31, 32, 64]. For example, a very large study in the US [65] shows no relationship between median annual precipitation and climate opinion. By contrast, an earlier US study [20] does find that the seasonal snowfall anomaly (as compared to the 30 year average) predicts beliefs about whether global warming is happening in the US However, this variable is not independent from temperature, since the occurrence and amount of snow is related to air temperature.

Beyond temperature and precipitation, some studies have also examined a range of impacts and occurrences of various extreme weather events. Of studies using external measures rather than self-reported experiences, 20 (39%) include measures related to extreme weather events. Several of these studies have used published aggregated indices of weather extremes as predictors of climate opinions, such as the US Climate Extremes Index that combines temperature, precipitation, drought severity, and landfalling tropical storms [28, 30, 33, 73]. These studies do not find a relationship between these aggregated extremes indices and climate opinions.

A further set of studies focus explicitly on the relationship between the impacts of weather-related extreme events and climate opinions, with mixed results. Several US studies utilize the National Oceanic and Atmospheric Administration's Storm Events Database [74] (or derived products like the Spatial Hazard Events and Losses Database [75]). This dataset records the impacts of weather-related events that cause loss of life, injuries, or major property damage, or are unusual enough to generate media attention (georeferenced at the county or forecast area level). For example, one study [76] examines weather-related property damage at the county level and does not find a direct effect on climate opinions. Another recent example [46] finds that hurricanes, tornadoes, floods, and droughts are unrelated to climate opinions, in contrast to self-reported experience for certain types of events. This finding is echoed by several other studies [17, 71, 72]. Yet, a few other studies do find that indices of extreme weather events predict climate opinion, such as a study of US Gulf Coast counties [67], a national US study using natural hazard fatalities [70], and a national US study using the number of extreme weather events by region [22, 23].

A final set of papers involve case studies of certain communities affected by extreme weather events, or the effect of a specific event on changes in climate opinion. A set of studies in the UK [21, 24, 53] focus on the experience of floods; these find that flood impacts predict climate opinion. However, an earlier UK study [54] finds that flood experiences does not affect climate opinion, and a recent study suggests that coping capacity moderates the negative emotions from a flood event that would motivate opinion change [52]. A recent study in the US [77] surveyed four communities exposed to tornadoes or wildfires and finds that event proximity does not predict climate opinion, as opposed to subjective harm from the event. Similarly, case studies of particular events like a flood event in Boulder, Colorado [78] or a drought in the US Midwest [29] show no effect on climate opinions. By contrast, a study in New Jersey finds that hurricane exposure predicts support for pro-climate political candidates [27].

A subset of studies focuses on how people perceive weather or climate conditions at the local level. Rather than using beliefs about global climate change as a dependent variable, these studies examine whether people perceive the climate in their local area to be getting warmer, whether recent seasons are warmer or colder than normal, or related local climate trends [25, 35, 36, 38, 39, 50, 60, 61, 71, 79–81].

Multiple studies have found that these subjective experiences of local weather or climate conditions are associated with broader climate opinions and political affiliation or ideology. In short, people who are already more concerned about global warming are more likely to think they have experienced its impacts, or have experienced weather conditions consistent with global warming [35, 36, 38, 39, 61, 71]. These trends are typically attributed to motivated reasoning or related phenomena.

Although studies of local climate perceptions show that climate opinions shape such perceptions, they also show that people are often able to detect a signal of local weather or climate despite biases created by motivated reasoning. For example, a study of Oklahoma residents found that they were able to perceive local seasonal temperature and precipitation anomalies despite biases introduced by political ideology [81]. Similarly, a study in Norway found that perceptions of seasonal temperature and precipitation were strongly associated with measured conditions while also exhibiting biases associated with climate opinions [36]. Detection of climate trends at the local level may be limited to larger-magnitude variations across large areas; however, a study in Florida found that five-year trends in temperature were not associated with perceived temperature trends across the state, although trends in precipitation were faintly detected [32]. The types of self-reported weather experiences shaped by climate opinions have not yet been fully explored, but there are suggestions that temperature-related experiences are more sensitive to biases driven by climate opinions than experiences related to precipitation or other extreme events [36, 38, 50].

Measurements of climate and weather (the treatment variables) vary considerably across the studies we review. As noted previously, a first-order distinction is between the 32 papers that measure the effect of subjective or reported experiences of weather and climate, versus the 51 papers that examine the influence of objectively measured weather and climate changes. Ten papers examine both subjectively (self-reported) and objectively measured weather or climate changes.

Local temperature measurements used in these studies vary in their use of absolute versus relative values, spatial extents and time intervals, and data sources. Temperature data for many studies are taken from the Global or United States Historical Climatological Network, which is an integrated database of climate summaries from land surface stations subjected to a common suite of quality assurance review. Station data are limited, however, in their ability to represent the continuous range of weather conditions across the Earth's surface. As distance increases from a station, its accuracy in capturing local conditions declines. Thus, it may be difficult to accurately represent weather experiences for populations living distant from a weather station (a challenge that is particularly prevalent in areas of the developing world with sparse station networks). To address this issue, other studies use gridded datasets derived from station and/or satellite data that create a continuous surface of modeled weather or climate variables. One example is the parameter-elevation regressions on independent slopes model (PRISM) dataset [82], which combines data from a variety of US monitoring networks. PRISM employs a range of modeling techniques that incorporate the influences of topography, for example, to derive spatially accurate estimates of climate parameters. Whether the use of these divergent data sources explains divergent findings has not been systematically assessed. In particular, it is not clear whether one type of data source provides climate data closer to what individuals in a particular location actually perceive. As a result, it is impossible to carefully diagnose the source of observed differences in treatment effects between studies employing these different datasets and associated indicators derived from these datasets. Similarly, there are multiple different sources of data and ways of calculating long-term temperature trends and extremes, including a heat stress index [32], climate extremes indices (e.g. [46]) and annual or seasonal trends based on temperature minimums, maximums, or means (e.g. [61, 67]).

Other weather and climate measurements are also characterized by these measurement inconsistencies. For example, heavy precipitation must be measured relative to some 'normal' base period or distribution, whether as deviations or percentiles. Heavy precipitation is also highly localized. In contrast, the extended absence of precipitation (i.e. drought) has a much larger and more homogeneous spatial imprint than heavy precipitation events, temperature anomalies or heat waves; it is also measured very differently. Drought can be measured through the duration of consecutive dry days, but thus far analyses of drought perceptions have relied on readily available indices such as the Palmer Drought Severity Index [28, 73, 79] or the US Drought Monitor [29, 68, 72]. Such indices, however, were not originally designed from the perspective of understanding how people experience weather and climate change but rather were designed for use in climatological, agricultural, and similar purposes. It remains unclear whether drought measurements operationalized in this fashion are consistent with how individuals perceive or experience drought.

Many papers that consider subjective experience focus on experience with extreme events or disasters, especially flooding [24, 47, 48, 52–54], but also tropical cyclones, drought, wildfire, and other changes [47, 49]. Such differences are non-trivial, as each event carries very different risks, is associated with different economic costs, and affects individuals and communities in very different ways. We need to better understand the different interpretations of subjective experiences in these varied contexts. At least one study explores this topic with open-ended responses, which provides insight into the diversity of interpretations of personal experience [59]. Additional studies using open-ended responses would be productive to understand the variety of ways in which climate change is expressed and perceived locally.

As with the independent variables, the variety of dependent variables is also substantial (table 1). Studies examine a diverse range of climate experiences and opinions, from whether individuals accurately detect changes in their local area (measured instrumentally) [32, 35, 83], to whether measured (or reported) changes are associated with increased awareness, a change in affect or emotion [48, 52], beliefs (e.g. [25, 76]), belief certainty [13, 37], risk perceptions (e.g. [6]), self efficacy, mitigation or adaptation policy preferences (e.g. [23, 62]), and intended behaviors. We identified no studies that measured actual behavioral changes. However, some studies measured behavioral intentions, including political behaviors, such as support for green politicians [27], media use [83], and intended reductions in energy use [24]. Most studies rely on individual question items but some construct narrow [70] or broad [65] risk perception indices from multiple items.

Examining the intersection of weather or climate and perceived experience or climate opinions requires careful attention to spatial considerations. Researchers make decisions about the scale of the study (from local to global), the distribution of individuals within the study domain, the distribution of weather or climate trends and events considered, and how climate data will be matched with individuals in the study. The studies reviewed span the full range of local to global analyses, but the vast majority are conducted within the United States, and many are national (table 2). The spatial distribution of people and climate are given varying degrees of attention. Samples may be selected based on particular extreme weather events in order to directly test the effects of those events (e.g. flooding). In other cases, an exploratory approach is taken. In some cases, respondents' locations are known (i.e. household addresses) and so analyses are performed at the individual level (e.g. [31, 32, 38, 66]), whereas in other cases respondents are aggregated by zip code, county, or even state levels (e.g. [8, 12]). The varying degrees of mismatch between respondent locations and weather or climate 'treatments' introduces uncertainties into any assessment of treatment effects.

Based on the geographic scope of studies identified in the review, it is clear that more non-US research is necessary, especially in China, India, and the global South, which represent some of the areas most vulnerable to climate change impacts. In addition, for studies that focus on the effects of climate changes that have relatively well-defined spatial signatures (e.g. wildfires, flooding, hurricanes), careful attention should be paid not only to the location of respondents within the geographic area, but also to how those respondents are matched with their respective weather or climate treatments. Weighting climate indicators according to the spatial distribution of population densities within a county or state, for example, is a relatively straightforward way to account for the uneven distribution of both climate and population in any area (e.g. [1, 84]). Population density grids are publicly available and would improve estimates of treatment effects in places that encompass large rural or sparsely-populated areas in particular.

At one pole, a number of papers describe associations between weather experiences and climate opinions without explicit attention to causal identification. These papers rely on multivariate controls to estimate the the effect of weather on diverse outcomes.

However, these studies face certain limitations. First, weather conditions are spatially autocorrelated; in other words, conditions are likely to be more similar for participants located closer to each other. Spatial dependence is inherent in many human and physical processes. Traditional regression models assume independence between observations. However, when studying the relationship between weather and public opinion, it is particularly important to account for spatial dependence in predictor variables. This will reduce the chance of underestimating standard errors and the likelihood of Type I errors (or incorrectly rejecting a null hypothesis). Multi-level models, clustered standard errors, or geographically weighted regression are several methods to account for geographic structure in predictors. However, most papers focused on the relationship between weather and climate opinions do not attempt to account for possible errors introduced by spatial autocorrelation using these or other methods, nor do they attempt to measure the extent to which spatial autocorrelation is present in their modeled regression residuals (though see [38]).

Second, studies that rely on self-reported measures of weather exposure face an additional inferential threat. In these cases, it can be difficult to tell whether public opinion responds to weather exposure, or perceptions of weather exposure are motivated by underlying opinions (see section 3.1.4 above).

Third, because weather varies geographically, geographic patterns of a particular weather variable may sometimes coincide with the geographic patterns of other unmeasured social, cultural, political, or demographic predictors. This means that meteorological variables may be correlated with other latent phenomena that are also causal drivers of a particular dependent variable. For example, if a period of cold weather strikes the central US while coastal areas are hotter than normal, then this weather pattern will be strongly correlated with underlying patterns of American political geography. This correlation will bias a cross-sectional multivariate analysis that predicts opinion using cold weather experiences. A related complication arises due to the possibility of multiple comparisons. Weather and climate datasets contain a wide range of variables that may be used as predictors of, for example, individual survey responses. These variables can be aggregated over multiple time periods selected by the researchers. Without adjustment, multiple comparisons among predictors can increase the risk of inferential errors.

A limited set of papers address these methodological issues head-on. These papers use as-if random variation in weather as the basis for their research design. Of course, most weather patterns are not randomly assigned across a large country or region. Instead, causal identification claims rest on the idea that, conditional on a particular area or geography, variation in weather is as-if random.

These papers thus exploit randomness in short-term or local weather conditions to test how weather extremes shape public opinion [7, 9, 65]. According to the logic of these papers, weather patterns vary more arbitrarily within a given county or local area. These differences can be causally identified. By contrast, when everyone in a given area is simultaneously treated with a large event like a hurricane or regional heat wave, then our efforts to understand the causal effect of weather will be compromised by non-random factors that are simultaneously associated with both weather trends and climate opinions.

These threats to inference can be managed through the inclusion of geographic fixed effects, preferably at the local level. Yet, only a handful of papers covered by this review include fixed or random effects at any level, including regional [20, 33, 46, 68], state [7, 9, 39, 91], or geographies below the state level (e.g. county, city, weather station) [9, 22, 36, 49, 65, 76, 77, 79].

Further, these studies would then benefit from demonstrating that, conditional on geographic fixed effects, populations which receive a particular weather treatment are identical to those that do not (but see [9] who do show that observed covariates cannot predict weather fluctuations). Instead, some articles simply assume that 'weather fluctuations are as good as random once geographic controls are included' ([7], p 406). More work needs to be done by the research community to understand whether or when this assumption is founded.

When pooling data over time, articles should also include time dummies to control for secular shocks [9, 31, 33, 62, 65, 81] or other more sophisticated controls such as linear time trends [65]. Similarly, between-country analyses should utilize country-level fixed or random effects [34, 63].

While the majority of observational studies reviewed relied on cross-sectional statistical analysis, a growing group of papers exploit panel data. These papers provide traction in estimating whether shifts in weather conditions are linked to shifting public opinions [8, 19, 28, 30, 31, 34, 37, 63, 65, 81]. Their results are not always consistent with cross-sectional studies (see above). In part, the drivers of shifts in climate opinions may not be the same as the drivers of absolute climate opinion levels. At the same time, panel analysis controls for a greater number of potential omitted variables, including any factors that are time-invariant in two-way fixed effect specifications.

An even smaller set of papers (n = 5) involve researcher-controlled experimental tests. Since researchers cannot control weather itself, these papers tend to test the causal effect of messages about extreme weather or exposure to particular extreme weather prompts. For example [18, 92], use a survey experiment to test whether public opinion in the aftermath of an extreme event changes whether a survey gauges their opinion on 'climate change' or 'global warming'. Another study [93] evaluates how shaping the prompt used to encourage evaluation of weather experiences shapes opinion. Other studies examine the effect of information, heat-related, or problem-severity primes [11, 18]. The lone article reviewed that used lab experiments [16] finds that warmth and thirst both increased subject belief in climate change and desertification risks. Beyond direct experience, climate-related experiences may also occur indirectly through exposure to media coverage or communication with other people such as friends or family. In addition, media coverage may interact with direct experience in influencing climate opinions. Future research should examine how such indirect experiences, and communication about direct experiences, might influence climate opinions. For example, such research could combine techniques to measure exposure to weather-related events with emerging techniques to measure exposure to media content [94] or online activity [95].

More broadly, the vast majority of articles reviewed here are quantitative in their approach. These articles all rely on cross-sectional or panel analysis of survey datasets. Only two articles we reviewed approached the study of opinion through qualitative interview-based or ethnographic lenses [57, 78]. For example [57], draws from 50 interviews with landowners in Oregon. These studies remind us that public attitudes towards climate change should not simply be reduced to survey scales. Instead, there is value in understanding the rich and multidimensional content of climate opinions, and understanding the ways in which weather experiences construct these beliefs. In general, this approach is under-represented in the existing literature. Some excellent work on this topic fell outside this review's sampling strategy, since it has been published in book form. For example [96], presents an ethnographic study on public experiences with an unusually mild winter in a Norwegian town. The book outlines how public engagement with climate change is shaped by social efforts to regulate emotions. In line with works like these, scholars with expertise in interview, focus-group. Focus-group and ethnographic methods could make a major contribution to this literature by expanding the scope of research on weather and public opinion.