The effect of different extreme weather events on attitudes toward climate change

interpreted as a treatment effect [41]. Importantly, such a design can strengthen causal claims regarding the relationship between extreme weather and public opinion about climate change and allows us to provide more credible inferences than designs that control for an array of cross-sectional covariates alone. We use an extension of difference-in-differences designs called dynamic or event study DiD. A common approach to studying disasters’ effects would be using a generalized differ- ence-in-differences design (i.e., two-way fixed effects). However, recent research has shown those designs can generate biased estimates when treatment effects are not homogeneous across time [42]. A design that allows us to draw inferences about the impact of disasters but does not present this shortcoming is a dynamic or event study difference-in-differences design [35]. In a dynamic DiD, estimates are aggregated at the length of exposure (e.g., at first expo- sure), which is particularly helpful when having multiple treatment periods. We have three waves of panel data in this particular case. Consequently, we can study whether the parallel trend assumption holds in one pre-treatment period and learn the effects of disasters after initial exposure and one year after initial exposure. An event study or dynamic DiD uses a staggered treatment, meaning that when a unit is considered exposed, it stays in that group and does not go back to the control group. Also, we incorporate key individual-level covariates that should not be affected by exposure

to a disaster as controls (i.e., education, gender, and age in the 2010 wave). Covariates were included for efficiency reasons, so they help to reduce standard errors. Findings are consistent when covariates are included (see Appendix B in S1 File). We use bootstrapped-based standard errors, which is recommended when exposure happens in a cluster design [43]. Table 1 provides a summary of the main variables used in the analysis. The outcome follows a 1–5 point scale; the exposure indicators can take four values: 0 means never exposed, 1 exposed for the first time in the first wave, 2 in the second, and 3 in the third. Education, gen- der, and age are the placebo covariates used in Appendix B in S1 File. Our design is based on comparing a control and treated group over time. The control group is composed of never-treated respondents (i.e., subjects that were not exposed to a disas- ter) and the treated group of exposed respondents (i.e., subjects that were exposed to a disas- ter). Exposure to a disaster can happen at different times, such as during the first, second, or third wave of the panel study. Subjects who live in an exposed county but before exposure are called eventually exposed or eventually treated. For example, a respondent living in a county exposed to a disaster during the 2014 wave would be considered eventually exposed during the 2010 and 2012 waves. Results Fig 1 reports the main results of exposure to the four common types of extreme weather on attitudes toward climate change using the five-point scale version of the outcome for the

Table 1. Descriptive statistics. Statistic Mean St. Dev. Min Max Outcome 3.493 1.330 1 5 First exposure to fire 0.243 0.658 0 3 First exposure to flood 0.575 1.026 0 3 First exposure to hurricane 0.377 0.802 0 3 First exposure to severe storm 0.647 0.911 0 3 Education 3.897 1.437 1 6 Gender 1.445 0.497 1 2 Birth year 1,954.212 11.603 1,919 1,992 https://doi.org/10.1371/journal.pone.0300967.t001 PLOS ONE Extreme weather events and attitudes toward climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0300967 May 15, 2024 5 / 10