Rapidly declining remarkability of temperature

anomalies may obscure public perception of

climate change
Frances C. Moorea,1, Nick Obradovichb, Flavio Lehnerc, and Patrick Baylisd
a
Department of Environmental Science and Policy, University of California, Davis, CA 95616; bMedia Lab, Massachusetts Institute of Technology, Cambridge,
MA 02139; cClimate and Global Dynamics Laboratory, National Center for Atmospheric Research, Boulder, CO 80305; and dVancouver School of
Economics, University of British Columbia, Vancouver, BC, Canada V6T 1Z4

Edited by Edward W. Maibach, George Mason University, Fairfax, VA, and accepted by Editorial Board Member Hans J. Schellnhuber January 14, 2019
(received for review September 25, 2018)

The changing global climate is producing increasingly unusual (7). Possible reference periods such as an individual’s lifetime
weather relative to preindustrial conditions. In an absolute sense, (8), a recent 30-y period (9, 10), or a trailing mean (11) have
these changing conditions constitute direct evidence of anthropo- been hypothesized, but no empirical evidence has yet been
genic climate change. However, human evaluation of weather as presented as to how individuals implicitly define normal condi-
either normal or abnormal will also be influenced by a range of tions or how quickly or slowly that definition changes over time.
factors including expectations, memory limitations, and cognitive This question is not only of theoretical interest. Past work has
biases. Here we show that experience of weather in recent years— shown that public policy tends to advance during “windows of
rather than longer historical periods—determines the climatic opportunity” provided by, among other things, focused public at-
tention (12). Without public perception of a problem, the ability of

ENVIRONMENTAL
baseline against which current weather is evaluated, potentially
scientific experts and policy analysts to advance a policy agenda will

SCIENCES
obscuring public recognition of anthropogenic climate change. We
be limited (13). This potentially poses a challenge for addressing
employ variation in decadal trends in temperature at weekly and
chronic environmental problems such as climate change. If base-
county resolution over the continental United States, combined lines describing “normal” conditions adjust rapidly, the public may
with discussion of the weather drawn from over 2 billion social not perceive there to be a problem requiring policy intervention,
media posts. These data indicate that the remarkability of partic- even as environmental conditions steadily deteriorate.
ular temperatures changes rapidly with repeated exposure. Using Here we provide evidence that the definition of “normal
sentiment analysis tools, we provide evidence for a “boiling frog” weather” shifts rapidly over time in a changing climate. We show

SUSTAINABILITY
effect: The declining noteworthiness of historically extreme tem- that the remarkability of particular weekly temperature anoma-

SCIENCE
Downloaded from https://www.pnas.org by 190.62.18.145 on February 10, 2023 from IP address 190.62.18.145.

peratures is not accompanied by a decline in the negative senti- lies, measured as the volume of social media posts about weather
ment that they induce, indicating that social normalization of they generate, adjusts on approximately a 5-y timescale. We find
extreme conditions rather than adaptation is driving these results. no evidence, however, that the declining noteworthiness of un-
Using climate model projections we show that, despite large in- usual temperatures is accompanied by reductions in their nega-
creases in absolute temperature, anomalies relative to our empir- tive effects on sentiment, implying social normalization of these
ically estimated shifting baseline are small and not clearly conditions without adaptation. Using climate model projections
distinguishable from zero throughout the 21st century. we show that, despite large increases in absolute temperature,

climate change | perception | Twitter | baseline | temperature Significance

A nthropogenic climate change involves the shifting of weather

beyond bounds historically experienced by communities and
ecosystems. Global average temperatures are now significantly
Climate change exposes people to conditions that are histori-
cally unusual but that will become increasingly common over
time. What kind of weather do people think of as normal or
higher than preindustrial levels, an effect that cannot be explained unusual under these changing conditions? We use the volume
without greenhouse gas emissions, while modeling studies have of social media posts about weather to measure the remark-
shown that local temperature anomalies will statistically emerge ability of different temperatures and show that remarkability
from the noise of natural variability in the relatively near term changes rapidly with repeated exposure to unusual tempera-
(1–4). However, how are temperatures that are extreme in a tures. The reference point for normal conditions appears to be
long-term, historical sense understood and interpreted by people based on weather experienced between 2 and 8 y ago. This
exposed to them? Will increasingly unusual temperatures con- rapidly shifting normal baseline means warming noticed by the
stitute direct, experiential evidence of a changing climate, or will general public may not be clearly distinguishable from zero
changing conditions be rapidly normalized so that even large over the 21st century, with potential implications for both
absolute temperature anomalies are not perceived as particularly the acceptance of global warming and public pressure for
unusual? mitigation policies.
Answers to these questions depend on how the subjective
definition of normal temperatures evolves over time as the cli- Author contributions: F.C.M., N.O., F.L., and P.B. designed research; F.C.M. performed
research; F.C.M. analyzed data; N.O. and P.B. contributed data; and F.C.M., N.O., F.L.,
mate changes: What baseline do people use to evaluate the and P.B. wrote the paper.
weather? In a nonstationary climate, the question of what the
The authors declare no conflict of interest.
appropriate climate reference window should be is not obvious.
Various baselines, ranging from the preindustrial period to the This article is a PNAS Direct Submission. E.W.M. is a guest editor invited by the
Editorial Board.
last 30 y, are used in the scientific literature, reflecting the in-
herent ambiguity in choosing a stable reference period in a Published under the PNAS license.
1
nonstationary series (5, 6). The baseline actually used by non- To whom correspondence should be addressed. Email: [email protected].
scientists in evaluating weather as either normal or abnormal is This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
even harder to theoretically specify, since it may be affected by 1073/pnas.1816541116/-/DCSupplemental.
generational turnover, memory limitations, and cognitive biases Published online February 25, 2019.

www.pnas.org/cgi/doi/10.1073/pnas.1816541116 PNAS | March 12, 2019 | vol. 116 | no. 11 | 4905–4910

 anomalies relative to our empirically estimated shifting baseline across space and within the year. This variation is what we use to
are small and not clearly distinguishable from zero throughout test whether the response to historically unusual weather con-
the 21st century. ditions changes with repeated exposure to those conditions.
Our principal empirical model regresses the logarithm of the
Approach number of weather tweets in each county–week on functions of
To investigate our research questions, we employ social media the reference and more recent temperatures. The model in-
data from Twitter. These data consist of all posts on Twitter cludes controls for precipitation, relative humidity, and cloud
between March 2014 and November 2016 geolocated within the cover (to isolate the effect of temperature) as well as differences
continental United States, for a total of 2.18 billion tweets (SI in Twitter use in counties and over time using the logarithm of
Appendix, Fig. S1). Tweets about weather are identified using a the number of Twitter users. County indicator variables (fixed
“bag-of-words” approach (Methods), and the classification was effects) control for all time-invariant difference between coun-
validated manually for 6,000 selectively sampled tweets (Methods ties while state by month of year indicator variables (e.g., De-
and SI Appendix, Table S1). The sentiment of all tweets that cember in California) flexibly control for any regional differences
included no weather terms was measured using two classification in seasonality. Finally, year fixed effects control for common
schemes [Valence Aware Dictionary for sEntiment Reasoning time trends across the United States over the sample period. The
(VADER) and Linguistic Inquiry and Word Count (LIWC)] and residual variation used to identify the causal effect of tempera-
a composite sentiment score calculated as the difference be- ture fluctuations on social media posts about weather is illustrated in
tween positive and negative sentiment (14, 15). We draw data on SI Appendix, Fig. S2. Standard errors (SEs) are clustered at the state
daily maximum temperature and total precipitation for the pe- level, allowing for spatial and temporal autocorrelation within a state
riod 1981–2016 from the PRISM Climate Group and aggregate (more details and the regression equation are given in Methods).
these data to the county or core-based statistical area level from
a 0.25° grid (16). We combine the PRISM data with cloud cover Results
and relative humidity data from the NCEP Reanalysis II (17). Using the full sample, we first show that the number of social
We then aggregate our social media and weather data to the media posts about weather is affected by temperature and that
weekly level. We employ weekly rather than daily resolution as this effect differs depending on the reference temperature for
weeks are a plausible period over which individuals might resolve that county and time of year. People are more likely to comment
the seasonal climatology of their area (e.g., “end of March” or on weather that is unusual for a particular place and time of year
“middle of November”) (18). For each area–week combination, than on the same weather if it is typical (SI Appendix, Fig. S3A).
a 10-y “reference” period is defined as the average of that area’s At the center of the temperature distribution (22 °C), both un-
temperature across the years 1981–1990 for each week of the usually hot and unusually cold temperatures are equally re-
year, a period defined based on the earliest available PRISM markable. At both hotter and colder ends of the distribution,
data. The effect of gradual changes on perception of tempera- however, the response is asymmetric so that more extreme
ture anomalies is identified using the spatial and seasonal vari- temperatures (i.e., colder-than-usual cold temperatures or
Downloaded from https://www.pnas.org by 190.62.18.145 on February 10, 2023 from IP address 190.62.18.145.

ation in temperature change since this baseline period. Fig. 1 hotter-than-usual hot temperatures) are most remarkable. More
illustrates one measure of this variation—the difference between recent temperatures also affect the pattern of comment on the
reference temperatures and the 2011–2015 mean—for the third weather. Counties that have experienced temperatures warmer
week in each calendar month across the United States. It shows than the reference period in the last 5 y are more likely to
substantial variation in exposure to temperature changes, both comment on cold temperatures and less likely to comment on

Fig. 1. Spatial and seasonal variation in the change in average temperatures between the reference period (1981–1990) and the 2011–2015 mean (in degrees
Celsius). Values shown are averaged for the third week in each month.

4906 | www.pnas.org/cgi/doi/10.1073/pnas.1816541116 Moore et al.

 warm temperatures than counties without recent warming, con- health, and economic productivity (20–23). It is therefore perhaps
sistent with an adjustment of expectations in response to recent unsurprising that they should generate more comment than more
conditions (SI Appendix, Fig. S3B). The converse is also true: typical and less consequential temperatures.
Counties experiencing recent cooling are less likely to comment Fig. 2 B and D, however, show that the remarkability of these
on cool weather and more likely to comment on warm weather. SI temperatures decays rapidly with repeated exposure. In the cold
Appendix, Table S2 summarizes these regressions and shows that sample, cold anomalies experienced between 2 and 8 y ago re-
models that allow a county’s response to temperature to differ duce the remarkability of contemporaneous temperatures (SI
depending on its history of temperature in recent and earlier time Appendix, Fig. S4A). This means that cold anomalies in a county
periods are preferred to those that do not allow this heterogeneity. that has experienced these anomalies for more than 5 y in a row
We use a finite distributed lag model to more precisely esti- are no longer remarkable (Fig. 2B). This effect is driven by a
mate the influence of past exposure to temperature anomalies on precisely estimated effect of lagged temperature anomalies be-
the remarkability of current temperatures. For each county– tween 2 and 8 y ago operating in the opposite direction of the
week in our sample we use its 15-y history of temperature contemporaneous effect (SI Appendix, Fig. S4A). An equal and
anomalies, defined relative to the 1981–1990 reference period, to opposite response is observed for warm anomalies at cold tem-
estimate how behavior adjusts in response to repeated exposure peratures (SI Appendix, Fig. S4B). The decline in remarkability
to altered temperatures. The model allows for nonlinear effects of hot temperatures with repeated exposure occurs even more
of temperature anomalies that change smoothly over time (for rapidly (Fig. 2D), although the lagged effects for this smaller
additional details see Methods). Given the asymmetry in the re- sample are less precisely estimated (SI Appendix, Fig. S4C).
sponse to temperature anomalies noted above, we split our SI Appendix gives information on the robustness of these re-
sample and estimate responses separately for the hottest and sults to alternate specifications. Results at both hot and cold
coldest quarter of baseline temperatures (greater than 28.3 °C temperatures are robust to alternate specifications of the tem-
and less than 13.6 °C, respectively). Since humidity is known to perature response and lag structure (SI Appendix, Fig. S5). Not
be important in driving adverse physiological effects of hot limiting the hot sample to locations with high humidity, however,
temperatures, we further restrict the hot sample to county–weeks results in a smaller contemporaneous effect and large error bars,

ENVIRONMENTAL
with relative humidity greater than 80% [corresponding to a heat particularly for the lag coefficients (SI Appendix, Fig. S6). This
index of 32 °C or greater (19)]. may be due to the importance of humidity in driving physiolog-

SCIENCES
Fig. 2 A and C show the effect of contemporaneous temperature ical discomfort at hot temperatures, or because subsetting the
anomalies in the cold and hot parts of the sample, respectively. In sample removes hot states in the southwest such as Texas and
both cases, more extreme temperatures (i.e., cold anomalies at cold Arizona with high penetration of air conditioners.
temperatures and hot anomalies at hot temperatures) are more Two mechanisms could be driving the rapid decline we ob-
remarkable than reference temperatures. These extreme tempera- serve in the noteworthiness of unusual temperatures with re-
tures have been shown to be socially consequential along several peated exposure. One possibility is that people are able to
dimensions, including mortality risk, emotional state and mental quickly adapt so as to lower the psychological or physiological

SUSTAINABILITY
SCIENCE
Downloaded from https://www.pnas.org by 190.62.18.145 on February 10, 2023 from IP address 190.62.18.145.

Fig. 2. Effect of current and past temperature anomalies on social media posts about weather. (A) Contemporaneous effect of temperature anomalies for
the cold sample (lowest quartile mean weekly maximum temperature). (B) Effect of a −3 °C temperature anomaly in the cold sample (20th percentile of the
distribution) as a function of number of years of exposure to that temperature. (C) Contemporaneous effect of temperature anomalies for the hot and humid
sample (highest quartile mean weekly maximum temperature and relative humidity >80%). (D) Effect of a +3 °C temperature anomaly (95th percentile of the
distribution) in the hot and humid sample as a function of number of years exposure to that temperature. Shaded areas show the 95% confidence interval.
Arrows are for visual reference and show the same effect plotted across two graphs: the instantaneous effect of a −3 °C temperature anomaly in the cold
sample (blue arrow) or a +3 °C anomaly in the hot sample (red arrow).

Moore et al. PNAS | March 12, 2019 | vol. 116 | no. 11 | 4907
 negative effects of extreme temperatures: Temperature anoma- (Fig. 3 B and D). Thus, our data suggest the rapid decline in
lies become unremarkable because they are less consequential. remarkability is a result of changing expectations of weather with
This would be a positive interpretation of the results shown in little adaptation to the adverse effects of weather extremes.
Fig. 2 since it implies that private, individual adaptation is highly Based on the empirical results shown in Fig. 2, we derive a
effective and occurs quickly. An alternative possibility is that learning model that describes how baselines of normal weather
declining remarkability is due to altered expectations but not adjust in response to experienced temperatures. Lagged periods
adaptation: Temperature anomalies become unremarkable be- during which the instantaneous effect of temperature anomalies
cause they are less surprising, but no less consequential. This are reversed are a “learning period” that defines the baseline
would mean that the adverse effects of extreme temperatures are against which instantaneous temperatures are evaluated. We use
rapidly normalized in that they quickly become socially un- the estimated lagged coefficients to define this period and the
remarkable. Moreover, since changing expectations should drive weighting of years within it (Methods and SI Appendix, Fig. S4).
people to adapt to new conditions, if declining remarkability is We use results from the cold rather than the hot and humid
not accompanied by reduced impact it would suggest adaptation sample to define the learning process because they are more
options available to individuals are limited. precisely estimated and indicate a longer time for the updating
To distinguish between these two divergent interpretations, we process, meaning our findings on the rate of adjustment will err
look for evidence of adjustment in the sentiment associated with conservatively. The pattern shown in Fig. 2 is consistent with
extreme temperatures under repeated exposure. Previous work baselines being determined based on weather between 2 and 8 y
has shown that average sentiment expressed in all social media ago (SI Appendix, Fig. S7). Weather experienced between 2 and
posts is more negative at both very high and very low tempera- 4 y ago is particularly important, providing empirical support for
tures, implying these conditions negatively affect people’s mood, the hypothesized “recency bias” (24, 25).
well-being, and emotional state (21). We use this as our measure We apply this learning model to climate model projections for
of the consequences of extreme temperatures because it can be the 21st century under RCP 8.5. Fig. 4 shows the annual,
measured for the same population and at the same geographic population-weighted temperature anomalies over the continen-
scale and spatial and temporal resolution as our remarkability tal United States, for 40 realizations of internal variability (26).
measure, allowing for direct comparisons between the two. Anomalies are defined relative to both a fixed 30-y baseline
Fig. 3 shows the change in sentiment associated with tem- (1981–2010) and to a shifting baseline defined using our em-
perature anomalies. The more remarkable temperatures identi- pirically estimated learning model. While persistent warming
fied in Fig. 2 are associated with negative sentiment. Cold over the 21st century results in very large temperature anomalies
anomalies at cold temperatures and hot anomalies at hot tem- defined relative to a fixed historical baseline, the empirically
peratures with high humidity both result in more negative derived, rapidly shifting baseline results in much smaller tem-
expressed sentiment (Fig. 3 A and C). However, we find no ev- perature anomalies, only slightly above zero. Moreover, given
idence of adaptation to these adverse effects on the 15-y time- internal climate variability, anomalies relative to the shifting
scale examined here. Temperature anomalies continue to have baseline are not clearly distinguishable from zero: Across the 40
negative effects on sentiment even after 5–10 y of continuous realizations, these temperature anomalies are less than zero (i.e.,
Downloaded from https://www.pnas.org by 190.62.18.145 on February 10, 2023 from IP address 190.62.18.145.

exposure, long after those anomalies have become unremarkable cooler than expected) in 26% of years on average.

Fig. 3. Effect of current and past temperature anomalies on sentiment expressed in social media posts using two sentiment measures. (A) Contemporaneous
effect of temperature anomalies for the cold sample. (B) Effect of a −3 °C temperature anomaly in the cold sample as a function of number of years of
exposure to that temperature. (C) Contemporaneous effect of temperature anomalies for the hot and humid sample. (D) Effect of a +3 °C temperature
anomaly in the hot and humid sample as a function of number of years exposure to that temperature. Shaded areas show the 95% confidence interval.
Arrows are for visual reference and show the same effect of plotted across two graphs: the instantaneous effect of a −3 °C temperature anomaly in the cold
sample (blue arrow) or a +3 °C anomaly in the hot sample (red arrow).

4908 | www.pnas.org/cgi/doi/10.1073/pnas.1816541116 Moore et al.

 Fig. 4. Effect of shifting baselines on the remarkability of temperature anomalies. Population-weighted annual average temperature anomalies over the
continental United States under RCP 8.5 with 40 realizations of internal variability (26). Anomalies are defined relative to a fixed 30-y period (1981–2010) and
relative to a shifting baseline defined using our estimated learning process. Population weighting uses population density fixed at 2015 values (37).

ENVIRONMENTAL
SCIENCES
Discussion calculated relative to the shifting baseline even when controlling
Here we show that the remarkability of temperature depends not for warming since the reference period and for state-level vari-
just on its absolute value but that it is affected by past experience ation (model 4, SI Appendix, Table S3). These initial results
and resultant expectations. More specifically, the subjective suggest a role for relatively recent experience of weather in
baseline against which temperature is evaluated appears to be shaping climate change beliefs, similar to findings by other au-
dominated by recent experience. Temperatures initially consid- thors (24). Further work is needed to more fully establish the
ered remarkable rapidly become unremarkable with repeated connection between our metric of remarkability, stated belief in

SUSTAINABILITY
exposure over a roughly 5-y timescale. Since expectation ad- climate change, and support for climate change policy.

SCIENCE
Downloaded from https://www.pnas.org by 190.62.18.145 on February 10, 2023 from IP address 190.62.18.145.

justment is rapid relative to the pace of anthropogenic climate One question is the role played by the media in driving the
change, this shifting subjective baseline has large implications for effects we estimate. If news coverage responds to, rather than
the notability of temperature anomalies as climate change pro- shapes, public ideas of notable events, then their role is epi-
gresses. Further, we find no evidence for effective adaptation phenomenal. If news coverage drives public perceptions of
over similar timescales, at least as measured using the negative newsworthiness, however, then some of the rapid decline in
effects of extreme temperatures on expressed sentiment. remarkability we estimate might be driven by editorial judg-
Collectively, these data provide empirical evidence of the ments. Inspection of a sample of tweets suggests the media
“boiling frog” effect with respect to the human experience of cli- organizations make up less than 5% of our dataset, meaning
mate change. This apocryphal metaphor describes a phenomenon they alone cannot fully explain the effect we estimate. How-
whereby the negative effects of a gradually changing environment ever, if news coverage influences the likelihood of individuals’
become normalized so that corrective measures are never adopted, commenting on the weather, then their effect will be larger.
even when those affected would have chosen to avoid these im- Irrespective of the mechanism, however, the declining note-
pacts ex-ante. Although casually discussed in regard to climate worthiness of changing temperatures implies short-lived public
change, the potential for normalization of steadily worsening en- attention and therefore that the “windows of opportunity” to
vironmental conditions has been noted in other fields, particularly advance climate policy on government agendas may be severely
with respect to biodiversity decline and ecosystem health (27, 28). limited (12).
Here we provide evidence for this social normalization occurring in Finally, we note that our results pertain only to ambient av-
a large population and show that it can happen at rapid timescales, erage temperatures. It may well be that more acute extreme
much faster than generational turnover. events such as storms, droughts, wildfires, or floods may be both
The question of how the rapidly declining remarkability of more consequential and more salient and therefore less prone to
temperature extremes relates to stated belief in anthropo- normalization (33). Previous work has found that other variables
genic climate change or support for mitigation policies is not such as changes in phenology or snowfall might be more strongly
straightforward. Many studies have identified a link between attributed to climate change in the public consciousness (34). In
local temperature anomalies and stated belief in global warming addition, a high-emissions scenario will produce absolute tem-
(10, 18, 29–31), with evidence that this is driven by individuals peratures that exceed the range of our data. It is possible that
substituting their personal experience for more relevant data on physiological or biological thresholds at these temperatures
global temperatures (32). Our results imply that this effect alone could result in nonlinear responses not accounted for here.
will not necessarily lead to widespread belief in anthropogenic The preindustrial period is often used as a standard reference
climate change with increasing warming, as the notability of local point in both climate science and policy (35), and unmitigated
temperature anomalies will adjust over time. As an initial in- greenhouse gas emissions over the 21st century will result in
vestigation of the role remarkability might play in determining large warming relative to this baseline. Understanding how these
policy-relevant variables, we conduct a simple regression of the historically unusual temperatures are evaluated by people af-
variation in county-level belief in climate change (SI Appendix, fected, and in particular whether these temperature anomalies
Fig. S8) on local temperature anomalies, calculated either using provide direct sensory evidence for the existence of climate
the shifting baseline or a fixed reference baseline (SI Appendix, change, requires knowing how weather is socially determined to
Fig. S9 and Supplementary Methods). We find a relationship be “normal” or “unusual.” Here we present evidence that the
between belief in climate change and temperature anomalies definition of normal adjusts rapidly in response to changed

Moore et al. PNAS | March 12, 2019 | vol. 116 | no. 11 | 4909
 conditions, despite the negative consequences of those changes of the current anomaly (i.e., during which there is evidence for diminishing
persisting. This results in temperatures that are largely unremark- surprise resulting from adjustment of expectations). The learning model is
able over the 21st century, even in a high-emissions scenario. When defined as the weighted sum of temperature anomalies experienced during
coupled with results from the existing literature, our finding sug- the learning period, with weights given by the relative magnitude of the
gests it may be unlikely that rising temperatures alone will be suf- estimated lagged coefficients. In other words, the subjectively defined,
ficient to produce widespread support for mitigation policies. moving baseline is given by

Methods X
8
~ cwy =
B wk Tc,w,y−k
Data Sources and Processing. Twitter data are the set of geolocated tweets k=2
between March 2014 and the end of November 2016 with device locations
within the continental United States. The total sample is 2.18 billion tweets, ^β
coming from 12.8 million unique users. Tweets discussing weather were wk = P8 k ,
^
j=2 βJ
identified using a simple bag-of-words approach. If the tweet contained one
of a list of words (given in SI Appendix, Supplementary Methods) it was where ^βk is the estimated effect of the temperature anomaly k years ago (SI
classified as a “weather tweet.” This classification was validated manually Appendix, Fig. S4A). Weights are calculated for the −3° temperature
for 6,000 tweets (SI Appendix, Supplementary Methods). Results of this anomaly (∼50% of the cold sample has temperature anomalies smaller than
classification are given in SI Appendix, Table S1. Additional information on ~ cwy is a nonlinear function of regression coefficients,
3° in magnitude). Since B
the weather, Twitter data, and sentiment analysis is given in SI Appendix.
SEs are calculated from the estimated variance–covariance matrix using the
Regression Analysis. All regressions include fixed effects for state by month of delta method (36).
year, year, and county, therefore controlling for all region-specific seasonal Temperature anomalies are calculated for the 21st century based on 40
variation, all common changes across years, and time-invariant differences simulations from 1980 to 2100 with the Community Earth System Model
between counties. Residuals are clustered at the state level. Controls for under RCP 8.5 (26). Population-weighted averages are taken over the con-
precipitation, cloud cover, and relative humidity are included to isolate the tinental United States (2015 distribution) (37). Rolling perceptual baselines
effect of temperature. The finite dynamic lag model (Figs. 2 and 3) allows the are calculated for the period 2010–2100 based on the estimated learning model
nonlinear effect of temperature anomalies to vary flexibly over time by and then temperature anomalies are calculated on an annual basis relative
fitting an interaction surface between the anomaly and the lag. Sentiment both to the 1980–2010 average and to the rolling perceptual baseline.
analysis is conducted at the core-based statistical area level with a slightly
different set of weather controls (discussion in SI Appendix). Regression ACKNOWLEDGMENTS. We thank Jeffrey Schrader, Angeline Pendergrass,
equations and more detailed methodology are given in SI Appendix. and seminar participants at Columbia University, National Bureau of Economic
Research Summer Institute, University of California, Berkeley, University of
Applying the Learning Model. We define a “learning period” as the years California, Santa Barbara, and University of Nevada, Reno for comments on
during which experience of past temperature anomalies reverses the effect the paper and Rudy Huezo for stellar research assistance.

1. Lehner F, Stocker TF (2015) From local perception to global perspective. Nat Clim 18. Howe PD, Leiserowitz A (2013) Who remembers a hot summer or a cold winter? The
Downloaded from https://www.pnas.org by 190.62.18.145 on February 10, 2023 from IP address 190.62.18.145.

Chang 5:731–734. asymmetric effect of beliefs about global warming on perceptions of local climate
2. Diffenbaugh NS, Scherer M (2011) Observational and model evidence of global conditions in the U. S. Glob Environ Change 23:1488–1500.
emergence of permanent, unprecedented heat in the 20(th) and 21(st) centuries. Clim 19. NOAA (2018) Heat index. Available at https://www.weather.gov/safety/heat-index.
Change 107:615–624. Accessed November 16, 2018.
3. Bindoff NL, et al. (2013) Detection and attribution of climate change: From global to 20. Deschênes O, Greenstone M (2011) Climate change, mortality, and adaptation: Evi-
regional. Climate Change 2013: The Physical Science Basis. Contribution of Working dence from annual fluctuations in weather in the US. Am Econ J Appl Econ 3:152–185.
Group 1 to the Fifth Assessment Report of the Intergovernmental Panel on Climate 21. Baylis P, et al. (2018) Weather impacts expressed sentiment. PLoS One 13:e0195750.
Change, eds Stocker TF, et al. (Cambridge Univ Press, Cambridge, UK), pp 867–952. 22. Colacito R, Hoffman B, Phan T (2018) Temperature and growth: A panel analysis of
4. Lehner F, et al. (2017) Toward a new estimate of “time of emergence” of anthro- the United States (Federal Reserve Bank of Richmond, Richmond, VA).
pogenic warming: Insights from dynamical adjustment and a large initial-condition 23. Burke M, et al. (2018) Higher temperatures increase suicide rates in the United States
model ensemble. J Clim 30:7739–7756. and Mexico. Nat Clim Chang 8:723–729.
5. IPCC (2013) Climate Change 2013: The Physical Science Basis. Contribution of Working 24. Kaufmann RK, et al. (2017) Spatial heterogeneity of climate change as an experiential
Group 1 to the Fifth Assessment Report of the Intergovernmental Panel on Climate basis for skepticism. Proc Natl Acad Sci USA 114:67–71.
Change, eds Stocker TF, et al. (Cambridge Univ Press, Cambridge, UK). 25. Weber EU, Stern PC (2011) Public understanding of climate change in the United
6. IPCC (2014) Summary for Policymakers. Climate Change 2014: Impacts, Adaptation States. Am Psychol 66:315–328.
and Vulnerability. Working Group 2 Contribution to the IPCC 5th Assessment Report, 26. Kay JE, et al. (2013) The community earth system model (CESM) large ensemble
eds Field CB, et al. (Cambridge Univ Press, Cambridge, UK). project: A community resource for studying climate change in the presence of in-
7. Tversky A, Kahneman D (1974) Judgment and uncertainty: Heuristics and biases. ternal climate variability. Bull Am Meteorol Soc 96:1333–1349.
Science 185:1124–1131. 27. Papworth SK, Rist J (2009) Evidence for shifting baseline syndrome in conservation.
8. Frame D, Joshi M, Hawkins E, Harrington LJ, de Roiste M (2017) Population-based Conserv Lett 2:93–100.
emergence of unfamiliar climates. Nat Clim Chang 7:407–411. 28. Pauly D (1995) Anecdotes and the shifting baseline syndrome of fisheries. Trends Ecol
9. Brooks J, Oxley D, Vedlitz A, Zahran S, Lindsey C (2014) Abnormal daily temperature Evol 10:430.
and concern about climate change across the United States. Rev Policy Res 31: 29. Donner SD, McDaniels J (2013) The influence of national temperature fluctuations on
199–217. opinions about climate change in the US since 1990. Clim Change 118:537–550.
10. Hamilton LC, Stampone MD (2013) Blowin’ in the wind: Short-term weather and 30. Egan PJ, Mullin M (2012) Turning personal experience into political attitudes: The
belief in anthropogenic climate change. Weather Clim Soc 5:112–119. effect of local weather on Americans’ perceptions about global warming. J Polit 74:
11. McCright AM, Dunlap RE, Xiao C (2014) The impacts of temperature anomalies and 796–809.
political orientation on perceived winter warming. Nat Clim Chang 4:1077–1081. 31. Deryugina T (2013) How do people update? The effects of local weather fluctuations
12. Kingdon JW (2010) Agendas, Alternatives and Public Policy (Pearson, London), 2nd on beliefs about global warming. Clim Change 118:397–416.
Ed. 32. Zaval L, Keenan EA, Johnson EJ, Weber EU (2014) How warm days increase belief in
13. Béland D, Howlett M (2016) The role and impact of the multiple-streams approach in global warming. Nat Clim Chang 4:143–147.
comparative policy analysis. J Comp Policy Anal Res Pract 18:221–227. 33. Howe PD, Boudet H, Leiserowitz A, Maibach EW (2014) Mapping the shadow of ex-
14. Hutto CJ, Gilbert EE (2014) VADER: A parsimonious rule-based model for sentiment perience of extreme weather events. Clim Change 127:381–389.
analysis of social media text. Proceedings of the Eighth International AAAI 34. Akerlof K, Maibach EW, Fitzgerald D, Cedeno AY, Neuman A (2013) Do people
Conference on Weblogs and Social Media (Assoc for the Advancement of Artificial ‘“personally experience”’ global warming, and if so how, and does it matter? Glob
Intelligence, Palo Alto, CA). Environ Change 23:81–91.
15. Pennebaker JW, Booth RJ, Boyd RL, Francis ME (2015) Linguistic Inquiry and Word 35. Schurer AP, Mann ME, Hawkins E, Tett SFB, Hegerl GC (2017) Importance of the pre-
Count: LIWC2015 (Pennebaker Conglomerates, Austin, TX). industrial baseline for likelihood of exceeding Paris goals. Nat Clim Chang 7:563–567.
16. PRISM Climate Group (2018) Northwest Alliance for Computational Science and 36. Oehlert GW (1992) A note on the delta method. Am Stat 46:27–29.
Engineering (Oregon State Univ, Corvallis, OR). 37. CIESIN; Columbia University (2018) CIAT Gridded Population of the World, Version 3:
17. Kanamitsu M, et al. (2002) NCEP-DOE AMIP-II reanalysis (R2). Bull Am Meteorol Soc Population Density Grid (NASA Socioeconomic Data and Applications Center, Pallisades,
83:1631–1643. NY). Available at sedac.ciesin.columbia.edu/gpw. Accessed July 27, 2017.

4910 | www.pnas.org/cgi/doi/10.1073/pnas.1816541116 Moore et al.