Increase in concerns about climate change following climate strikes and civil disobedience in Germany

Climate Protest Data

We construct a database containing salient climate protests organized by various climate movements in Germany. We first identify the relevant groups organizing climate protests, focusing on the recent cycle of climate activism11. The groups attracting the most participants and news coverage are the youth-led Fridays for Future (FFF) and Ende Gelände (EG) (German saying for “here and no further”) movements, as well as to a lesser extent the group Extinction Rebellion (XR)47. While FFF mobilizes global climate strikes12, EG organizes mass actions of civil disobedience13, such as blocking coal infrastructure48. XR plays a minor role in Germany, mainly focusing on acts of non-violent civil disobedience14, such as blocking roads.

We assume that media reports are the main transmission channel for how climate protests affect the population. Sisco et al. 18 and Wasow (2020)19 find that media attention mediates the effect of protests. Consequently, we only select climate protests that are salient to the general public through this channel. This selection is based on whether a specific climate protest is reported on in the evening news of the two leading public service broadcasters, ARD and ZDF. The Tagesschau and heute are the most trusted49 and viewed news formats in Germany, respectively reaching 11.8 m and 4.6 m viewers in 202050 and are precursors for news topics in other media in Germany (see Supplementary Fig. 1). We include global climate strikes and confrontational protests in our protest database if they appeared in the reporting of either the ARD or ZDF news formats or both. Table 1 displays the climate protests identified with this method.

German Socio-Economic Panel

The main empirical analysis relies on the SOEP, a representative longitudinal household survey of about 15,000 households conducted yearly since 198451. The survey covers various topics and socioeconomic characteristics, including indicators of attitudes and concerns. The data contain information from all household members aged 12 years and over, which includes information on approximately 35,000 individuals. Our outcome variable asks, “How concerned are you about the consequences of climate change?”. The categorical variable has been part of the questionnaire since 2009 and has three possible answers, ranging from no concern over a few to large concerns. Based on this variable, we construct a dummy for “concern,” which equals one if a person is concerned to any extent and zero if they are not. Thus, we can measure the extensive margin of climate change concerns. Figure 1 shows the weekly averages of this variable from 2016 to 2020. We also show that our results remain the same when all three levels of the outcome variable are used (Supplementary Table 8).


The quasi-random occurrence of protests relative to survey dates makes it possible to identify the causal effect of climate protests on individuals’ climate change concerns in a before-and-after research design. As the SOEP surveys respondents over the entire year, the timing may coincide with climate protests. Whether respondents are interviewed immediately before or after a protest is plausibly random. Thus, comparing responses shortly before the protest with those shortly after the protest can, under certain assumptions, identify the causal effect. This approach is also called Unexpected Event during Survey Design25 and has been applied in a variety of settings52,53. In our main specification, we compare a window of 14 days before (N = 12,633) and after (N = 11,933) a protest. We present our findings using multiple alternative time windows to ensure this choice does not affect the results. Supplementary Fig. 2 demonstrates how this time window looks around an exemplary protest and identifies the treatment and control group for a protest. If the post- and pre-treatment periods of two protests overlap, we end the respective periods in the middle.

Muñoz et al. 25 argue that identification relies on two main assumptions: temporal ignorability and excludability. First, temporal ignorability implies that the timing of the survey is independent of the timing of climate protests. Thus, we must assume that comparable population groups are interviewed before and after a protest. This assumption is unlikely violated since the timing of the interview is determined by a long-standing panel survey in which the implementation logistics are decided well in advance. To assess the validity of this intuition, we conduct balance tests on respondent characteristics (age, gender, education, etc.) by regressing them on the treatment indicator and a complete set of protest dummies. The dummies ensure we compare whether the characteristics differ before and after a particular protest and not across all protests. Supplementary Table 1 presents the results of these tests, which suggest that the assumption is plausible, given no statistically significant differences in observed respondent characteristics before and after climate protests (age: p = 0.764, female: p = 0.130, household size: p = 0.140, education: p = 0.895, employment status: p = 0.731, teenagers in household: p = 0.285, household labor income: p = 0.409, interest in politics: p = 0.703, political orientation: p = 0.172). Excludability implies that the survey interview’s timing does not impact the outcome except through the analyzed event. In our case, the timing of the SOEP interview should only affect climate change concerns through the respondents’ exposure to the protest. This identifying assumption may be violated by simultaneous events or time trends in the outcome variable25. Although we cannot conclusively test this assumption, we provide evidence that it holds by conducting placebo tests. We test the validity by measuring i) the effect of climate protests on other concerns that are part of the standard SOEP54 and ii) the effect of hypothetical protests on climate change concerns (Supplementary Table 2). We do not find systematically statistically significant effects in the placebo exercise on other concerns. In particular, we test the effects on respondents’ concerns about the general economic development (p = 0.217), own economic situation (p = 0.071), own pension (p = 0.659), own health (p = 0.054), peacekeeping (p = 0.808), job security (p = 0.910), immigration (p = 0.674), and xenophobia (p = 0.556). We only observe statistically significant effects on crime-related concerns (coef. = 1.09pp, p = 0.042). Some statistically significant coefficients are expected when not controlling for potentially confounding events relevant to each alternative outcome and testing a broad range of hypotheses. Interestingly, climate protests may have a statistically significant impact on concerns about crime, as respondents might become more sensitive to the issues of protesters breaching the law once they are exposed to information about these protests. Importantly, we do not detect statistically significant effects of the hypothetical protests (p = 0.484), indicating that our method does not simply pick up time trends in climate change concerns.

We estimate the effect of protests on climate change concerns (Concern) with the following model:

$${{{{{{\rm{Concern}}}}}}}_{i,s,p,d,t}= \alpha+\beta {{{{{{\rm{Post}}}}}}}_{i,p,d,t}+\gamma {{{{{{\bf{X}}}}}}}_{i,t}+\delta {{{{{{\bf{I}}}}}}}_{i,t}+\epsilon {{{{{{\bf{C}}}}}}}_{s,d}+{\zeta }_{p}+{\eta }_{t}+\theta {{{{{{\bf{D}}}}}}}_{d} \\ +{\varepsilon }_{i,s,p,d,t},$$


where i denotes the individual living in state s, p the respective protest, d the date of the SOEP interview and t the year. Post represents the treatment effect of protests. The dummy equals one if the individual is interviewed after the protest and zero otherwise. Xi,t is a vector of several (socioeconomic) individual-level characteristics in year t that have been shown to be associated with beliefs and concerns about climate change33,34. We include the respondent’s age, self-reported sex (dummy), number of years in education, employment status (dummy), the 2-digit industry code of the respondent’s work (categorical), household size, the number of children aged 14 to 18 in the household, household labor income and the respondent’s interest in politics as well as their political orientation. Political orientation is elicited on a Likert scale ranging from 0 (far left) to 10 (far right). We include a categorical variable indicating “left- leaning” (values 0–4), “right-leaning” (values 6-10), and “center” (value 5 which is the largest category). Similarly, interest in politics is elicited on a 1-4 scale, where we include “(very) strong” (values 1-2) and “weak or none” (3-4). Both variables are pre-treatment values. Ii,t is a vector of interviewer characteristics. It controls for education, sex, and age, and the variables are prepared equivalent to the respondent characteristics. To avoid the loss of observations due to missing values in these variables, we include dummies indicating missing information in a variable in Xi,t and Ii,t.

We further control for external factors possibly correlated with the treatment indicator and climate change concerns in Cs,d. It includes variables related to weather anomalies and relevant political events (federal elections and UN COPs). Weather data is obtained from Germany’s National Meteorological Service (DWD). We operationalize temperature anomalies by taking the absolute deviation of the mean precipitation (temperature) in the month of the interview in state s from the historical mean precipitation (temperature) in that state and month, and this absolute deviation of the mean precipitation (temperature) squared. Historical averages for each state and month are calculated between 1950 and 2000. The dummy for political events is equal to one in the month (week) before and after federal elections (UN COPs). ζp are protest fixed effects, meaning we estimate the effect of Post by comparing individuals around a particular protest. ηt are year fixed effects and Dd is a vector of dummies for the day of the week and month of the year the interview took place. It controls for potential systematic differences in responses across weekdays and months since the timing of the protest is likely correlated with certain weekdays or months (especially with FFF events). In further specifications, we also control for unobserved differences across states where the respondents live by including state fixed effects or differences across states for each protest by including state-by-protest fixed effects. In our main results, we rely on robust standard errors clustered at the level of the protests. The degrees of freedom (df) equal N-k-155, where k is the number of variables and equals 380 in our preferred specification resulting in df = 24,185.

Robustness checks

To test the robustness of our results, we first show that our main results in Table 2 are independent of the chosen estimation method. In line two of Table 2, we adjust any remaining imbalances between respondents before and after protests using entropy balancing of means27. This method is frequently applied conjointly with the methodology used in this study25,56. Entropy balancing is a re-weighting scheme that calibrates unit weights to improve the balance of covariate means between the treatment and control groups further. We use all except the pre-treatment covariates to create the weights since the relatively large number of missing values would reduce the sample size (see Supplementary Table 1). In line three of Table 2, we use alternatively a Probit model instead of the Linear Probability Model to address the common concern that linear probability models might not fit binary outcome data well and predict unreasonable values outside of the zero to one range. Given that the coefficients in these models are complex to interpret, we transform the estimates into marginal effects on the probability of being concerned, making them comparable to our main estimation. Next, we adapt the time window around protests. Moreover, we exclude up to seven days before a climate protest to rule out anticipation effects (Supplementary Fig. 3a). Furthermore, we iteratively exclude climate protests to confirm that single protests do not drive the results (Supplementary Fig. 4). Lastly, we exclude events when two protests happened in three days, which we count as one in our main specification (Supplementary Table 6).

Our main results report aggregate effects across all selected climate protests. To further check the validity of our aggregate results, we analyze the individual effects of each protest. First, we split the sample into sub-samples for each protest and investigate whether the identifying assumptions of our method are still likely to hold around each protest to estimate credible effects. Supplementary Fig. 5 displays the covariate balance for each protest to test the temporal ignorability assumption for protest-specific estimations. The covariates are relatively balanced for certain protests (Supplementary Fig. 5a). However, given the reduced number of observations, the covariates of respondents around some protests are not entirely balanced. Therefore, going beyond controlling for respondent characteristics in the regressions, we adjust for remaining covariate imbalances by applying entropy balancing27. Supplementary Fig. 5b displays the well-balanced covariates with entropy balancing weights. Consequently, we include these weights in the protest-level estimations. We further test the excludability assumption at the protest level. Supplementary Table 7 presents the placebo tests that estimate each protest’s effect on other concerns. The outcomes of these tests show that, for each protest, our empirical approach is not systematically driven by time trends in the outcome variable. These results suggest that the excludability assumption may be reasonably valid for protest- specific estimations.

The study was granted ethics approval by Hertie School’s Research Ethics Officer under the application ID 20230220-27. To perform the empirical analysis, we have used Stata MP 16 64-bit (packages: reghdfe version 5.7.3, ebalance version 1.5.4, estout version 3.17, coefplot version 1.8.5, mlogit version 11.4.2, and gologit2 version 3.2.5) and R 4.3.1 (packages:.ggplot2 version 3.4.3, lubridate version 1.9.2, readtext version 0.90, dplyr version 1.1.2, gridExtra version 2.3, haven version 2.5.3, and patchwork version 1.1.3). For further details please refer to the replication package of this study57.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

You may also like...

Leave a Reply

Your email address will not be published. Required fields are marked *