Contextualizing cross national patterns in household climate change adaptation.

Understanding social and behavioral drivers and constraints of household adaptation is essential to effectively address increasing climate-induced risks. Factors shaping household adaptation are commonly treated as universal; despite an emerging understanding that adaptations are shaped by social, institutional, and cultural contexts. Using original surveys in the United States, China, Indonesia, and the Netherlands (N=3,789) - we explore variations in factors shaping households' adaptations to flooding, the costliest hazard worldwide. We find that social influence, worry, climate change beliefs, self-efficacy, and perceived costs exhibit universal effects on household adaptations, despite countries' differences. Disparities occur in the effects of response efficacy, flood experience, beliefs in governmental actions, demographics, and media, which we attribute to specific cultural or institutional characteristics. Climate adaptation policies can leverage on the revealed similarities when extrapolating best practices across countries, yet should exercise caution as context-specific socio-behavioral drivers may discourage or even reverse household adaptation motivation.

Worldwide, escalating climate-induced hazards inflate economic damages , undermine livelihoods , and force migration . The approaching new climate reality calls for an urgent and ambitious adaptation at all levels: from government-led actions to household climate change adaptation behavior . Understanding how and why households adapt is critical for diminishing adaptation deficits and overcoming socially-constructed adaptation limits, , for fostering societal resilience, and risk communication . Recent research on households’ adaptation behavior to climate-induced hazards provides valuable insights in factors shaping individual motivations to adapt . Growing empirical evidence indicates that perceptions, experience, and self-efficacy could facilitate or inhibit households’ adaptation to hazards .

Flooding is the most widespread and costliest climate-induced hazard worldwide . Previous work has advanced our understanding of the empirical drivers of household flood adaptation, but has primarily focused on single countries; with rare exceptions that utilize non synchronous and non-identical surveys . Furthermore, while climate change disproportionately impacts Global South countries, most surveys on households’ flood adaptation are conducted on the Global North . Yet, adaptation is locally shaped, and social, institutional, and cultural factors likely affect individual adaptation behavior . In exploring these influences, past work has faced data limitations , with the result that household adaptation and its drivers and constraints are often discussed uniformly across diverse contexts.

Household adaptation involves different actions, ranging from seeking information, to hazard-proofing one’s property. Previous studies suggest that households’ adaptation behavior that varies in effort and costs could trigger different decision pathways . Yet, research that specifically tests to what extent the drivers of different adaptations vary is notably missing. Hence, extrapolating a universal theoretical and empirical understanding of household adaptation behavior in diverse and understudied contexts remains a key challenge in the field of climate adaptation .

To address this gap, we question to what extent commonly theorized factors of household adaptation have analogous effects across (a) different contexts and (b) adaptation types that require varying degrees of implementation efforts. To gather sufficient data to answer these research questions, in March-April 2020 we conducted identical household surveys (N=3789) in four countries: the United States (USA), China, Indonesia, and the Netherlands. We focus on coastal urban areas, which are vulnerable to flash, river and coastal floods, and to sea level rise (See Supplementary Material Table S.2 for specific survey location details) . The four countries represent unique social, institutional, cultural, and geographic contexts. USA and the Netherlands are two Global North nations where theories of behavior under risk were developed and advanced , and floods surveys are repeatedly administered . China and Indonesia are two Global South nations where prior surveys on factors motivating households’ flood adaptation behavior are limited. All four however, are front-runners in escalating flooding risk , yet vary in the frequency of flood experiences: from nearly annual (Indonesia) to once-in-a-lifetime (the Netherlands). The four cases differ culturally, and in the role governments take in adapting to climate-induced floods (stronger centralized protection in the Netherlands and China vs. more individual responsibility in Indonesia and USA).

To measure adaptation intention, we examine 18 household-level actions (details in Methods). Drawing on prior findings on the differences in adaptation motivation towards flooding based on the type of measures and potential synergies , we classify our 18 measures into two groups (supported by confirmatory factor analysis, Supplementary Material Table S.6). High Effort measures (8) involve structural, usually irreversible modifications to one’s home, and Low Effort measures (10) comprise less intensive non-permanent protection and communication actions. For both groups we estimate the proportion of intended actions from the remaining actions – not yet undertaken actions per case-study. Adaptation intention for these two groups are the focus of our analysis (‘Dependent Variables’ in Methods).

To determine what drives and hinders households’ adaptation decisions, we build on Protection Motivation Theory (PMT) . Following previous work , our survey examines perceived hazard probability, perceived damage, and worry about flooding driving Threat Appraisal, and self-efficacy, response efficacy and perceived cost shaping Coping Appraisal (Figure 1). We expand the original PMT model to account for preceding engagement with hazards (prior actions, experiences) , external influences (media, peers) , climate-related beliefs , and demographic background . Hence, our 16 explanatory variables (Figure 1; see details in Methods) go beyond interpersonal factors to account for some intra-personal cues considered essential for behavioral adaptation . To quantify the effects of these 16 socio-psychological factors on household adaptation intentions we estimate and analyze the effects from Bayesian beta regression models (details in Methods), separately by country and measure group.

Figure 1

Social and behavioral factors motivating household climate change adaptation in four countries: USA (N=1,139 survey respondents), China (N=842), Indonesia (N=1,080), and the Netherlands (N=728). All 16 variables under the categorical groupings (bold) are included in the Bayesian beta regression models. The circles demonstrate the effects of these variables on households’ adaptation intentions for High Effort and Low Effort Measures. The size of the circle is proportional to the size of the effect, which if negative, is presented by a hollow circle; the colors denote the four countries. The effect sizes and standard errors are presented in detail in Figure 2 and in Section S.4 in Supplementary materials.

We find that while a few drivers have universally consistent effects across countries and measure groups (i.e. social expectations and worry), others exhibit salient difference across countries (i.e. response efficacy) or measures (i.e. self-efficacy and cost) (Figures 1, 2). Key similarities and differences in the drivers across countries, when properly contextualized, could help strategies aimed at extrapolating household adaptations to data-scarce regions.

Patterns in primary drivers of household adaptation

The perception of a greater threat, is generally associated with an increased likelihood to take adaptive action . In line with past empirical works , our analysis affirms that emotional, rather than analytical, reasoning drives household decisions. The former is intuitive and fast , while probabilities requiring cognitive efforts, are abstruse to the public . Perceived probability and damage, offer little power in explaining households’ intentions to adapt across all four countries (‘Fl Prob’ and ‘Fl Damage’ in Figure 2). The effect of perceived damage in Indonesia presents an exception when estimating High Effort measures; possibly due in part to the vulnerable position and high exposure to flood damage many households face in Jakarta annually . Yet, even in Indonesia, ‘Worry’ offers more explanatory power than the calculated risk variables (‘Fl Prob’ and ‘Fl Damage’). ‘Worry’ has a consistently positive relationship with adaptation intention for both High and Low Effort measures across all counties (Figure 2).

Figure 2

The ‘X’s’ report the mean effect sizes of for factors influencing households’ intentions to adapt to flooding. The vertical bars indicate 95% credible intervals. The effects are calculated from Bayesian beta regression models; run separately by adaptation type - (a): High Effort, (b): Low Effort - in four countries (United States (N=1,139), China (N=842), Indonesia (N=1,080), the Netherlands (N=728).

Coping Appraisal is generally a strong predictor of action . Among the three Coping Appraisal variables, the effects of two - self-efficacy and perceived costs - on household intention to take High Effort measures, are universally consistent across the four countries (Figure 2.a). In line with PMT and with past research , households who report greater capability and view the measures as generally less expensive (‘Self Eff’ and ‘Cost’, Figure 2.a), are more likely to intend adaptation for High Effort measures. Notably, in the two economically-wealthier Global North countries, USA and Netherlands, perceived cost is 2-4 times less of a deterrent in house-hold adaptation compared to the two Global South countries - China and Indonesia (Section S.4, Supplementary Materials), calling for innovative climate finance solutions that support adaptive capacity in Global South.

The effect of response efficacy on intending to undertake High Effort measures, differs among countries (Figure 2). In USA and the Netherlands it likely has no effect on adaptation intentions; in Indonesia, the effect is marginally positive. In China however, we observe a negative effect meaning that households that, in general, view these household adaptation measures as less effective overall, paradoxically are more likely to adapt. While a null or negative response efficacy is not unheard of when estimating a grouped adaptation variable , past empirical work usually demonstrates positive effects of ‘Resp Eff’ on adaptation intentions . Chinese culture, in comparison to the other three case-studies, is more long-term oriented . Longer-term oriented cultures situate their beliefs in a broader temporal context, potentially situating the way people assess efficacy in the longer term. Possibly, flood-aware respondents in China, who see property-level adaptions as less effective in the long term, may yet recognize the short-term utility of some measures - and hence are driven to adapt to remedy the more imminent adversities.

For Low Effort measures, in contrast to PMT, perceived costs have a reverse effect on households’ intentions to adapt in all four countries (Figure 2.b; Section S.4, Supplementary Materials). Likewise, compared to High Effort measures, we see a universal substantial decrease in the effect of self-efficacy on intentions for Low Effort adaptations. The change in effects is likely due to the fact that several of the measures in this group are free and require minimal effort, (i.e. coordinating with neighbor or moving expensive furniture to a higher floor). Hence, measures that require less time and resource investments likely have different psychological drivers and/or are made using varying heuristic shortcuts . Further, we also find larger standard errors and slightly greater variance in effects of ‘Resp Eff’ among countries for Low Effort measures compared to High Effort - possibly due to more accurate reporting on intentions to undertake High Effort measures . Intentions to pursue Low Effort adaptations by households in USA and the Netherlands and, to a lesser degree, Indonesia are positively affected by ‘Resp Eff’, while the negative effect in China remains, though lessened.

Role of experience, background, beliefs, and social influence

In Indonesia and USA, 46% and 48%, respectively, of the households included in this analysis reported having experienced a flood, in stark contrast with China (19%), and the Netherlands (15%) (Table S.1, Supplementary Materials). Yet, prior flood experience is a weak predictor of High Effort adaptations among our respondents everywhere, except the Netherlands (‘Fl.Exp’, Figure 2.a). In China, Indonesia, and USA, floods occur annually throughout the country. Dutch residents, by contrast rarely experience them, except occasionally with heavy rainfall or in unembanked areas. Since beliefs and personal baselines are formed in the context of own experiences , for a Dutch household, a flood is a unique experience creating a memorable availability heuristic ) that positively influences (95% likely) adaptation intentions.

Our data demonstrates that 17.7%-39.5% of households in four countries have already undertaken High Effort adaptation measures, and almost twice as many (43.2%-78.6%) have adopted Low Effort measures (Figure 3). The effect of prior adaptation on intending additional Low Effort measures has a strong negative effect everywhere, except the Netherlands (null effect for ‘Undergone’, Figure 2.b). Whereas for High Effort, the likely negative effect is lessened, and is only present in China and Indonesia (Figure 2.a). Both countries suffered major floods in the preceding nine months before our survey: 2019 Typhoon Lekima in China, and 2019 Jakarta Floods in Indonesia. Possibly, households in these countries have more recently undergone High Effort flood adaptation measures - lessening the likelihood that they would need to intend others in the immediate future.

Figure 3

The percentage of households who have previously undergone at least one adaptation in each category.

While the effect is not included in the models (to maintain model independence for comparative purposes), it is worth noting that households who have not undergone Low Effort measures are more likely to intend High Effort measures (Wilcox Rank-Sum: for each individual country, p< 0.001). Still, households who have not undergone High Effort measures - in USA (Wilcox Rank-Sum: p< 0.001), China (Wilcox Rank-Sum: p< 0.01) and Netherlands (Wilcox Rank-Sum: p< 0.01) - are more likely to intend Low Effort measures. This is not the case in Indonesia (p=n.s.), where due to the relatively high flood exposure households that feel at risk, have likely already taken at least some Low Effort measures.

The effects we observe from the demographic variables are mixed and generally weak (Figure 2). In USA, Indonesia and the Netherlands, ‘Age’ has a small negative effect on intentions to pursue High Effort measures, perhaps due to discounting of implementation efforts over the remaining lifetime in own property. Age also discourages Low Effort measures in USA and China. That elder respondents are less likely to intend adaptation than the younger is concerning: they are more vulnerable and require specific attention during and following disasters . ‘Education’ has a positive effect on adaptation intentions only for households in USA (>99% likely for both High and Low Effort measures) and China (98% likely for High Effort measures), while in other countries it matters less. Gender has a likely null effect everywhere except Indonesia where men appear more likely than women (92% certainty) to intend Low Effort measures. Our sample respondents are slightly more educated than the general population, and in China and Indonesia somewhat younger; possibly influencing the effects (Supplementary Material, Tables S.3, S.4). These results however are in line with other international flood adaptation work that found inconsistent explanatory power from personal characteristics .

Across the four countries, between 62%-79% of respondents believe that climate change is happening now (Table S.1, Supplementary Materials). Past work however has shown that belief in climate change often does not translate into action , can deter action , and does not necessarily have a strong cognitive link with extreme weather events . Here the belief that climate change is happening now, (‘C.C.Now’, Figure 2) has a negative direct effect consistently in all four countries. The reason could be that households who believe in urgency of climate change, have already taken some actions - as many in our dataset have (Figure 3). Notably, the belief in climate change is associated with having previously undergone Low Effort measures (χ2=123, p=0.0). While there is no discernible relationship between belief in climate change and previously undergone High Effort measures, as noted with past action, having undertaken Low Effort measures is associated with less intention for both High and Low Effort. Hence, it likely quells protection motivation and possibly explains the negative relationships.

Government adaptation many influence households’ intentions. Previous research often found negative effect on households’ adaptation intentions of trust in governmental protection or of belief that it is governmental responsibility . We go beyond measuring general beliefs and asked specifically whether households think actions already taken by their respective governments were sufficient (Table S.1, Supplementary Materials). In Indonesia and USA the belief that the current government measures are inadequate discourages household adaptation intentions for both measures (>98.5% likely); whereas in China and the Netherlands the effect is small and uncertain, hence likely null (‘Gov Meas. Insuff.’ Figure 2).

Two institutional and experiential differences between countries could explain the observed disparity in effects. First, the negative relationship in USA and Indonesia between the belief that governmental measures are inadequate and own adaptation intention aligns with other work that finds public and private adaptation can go hand in hand, especially for adaptations that entail structural property modifications . This relationship has been previously rationalized by the logic that past flood events or close calls can trigger both public action and private household adaptation . Indeed, everywhere if our respondents have experienced a flood, they are more likely to have already undergone measures (High Effort: χ2=123, p=0.0, Low Effort χ2=61, p=0.0) possibly lessening the intention for further action. In Indonesia and USA more people have experienced floods than in the Netherlands and China (Supplementary Material Table S.1). If a household has experienced a flood, they are also more likely to view the government measures as insufficient (χ2=30, p<0.001). Second, China and the Netherlands have a similar, collectivist approach to flood management - that is in general, trusted by the populaces . In Indonesia and USA many disaster management programs are viewed generally more unfavorably and as insufficient . Our own data reflects these sentiments: 11% of Dutch and 22% of Chinese view flood protection measures already taken by the government as insufficient compared to 30% in Indonesia and 43% in USA.

Norms play a strong role in influencing behavior . Our analysis supports this notion: the perceived expectations of one’s friends, family, and neighbors, as a prescriptive norm, positively influences the intention to implement both High and Low Effort measures across all four countries (‘Social Inf’, Figure 2, and Supplementary Materials). Differences between the four countries appear in the extent of social influence on households’ adaptation, with USA exhibiting the lowest positive effect of social influence on High Effort adaptations, perhaps due to nation’s individualistic culture . With Low Effort measures, we find that social expectations play a higher role in China and the Netherlands compared to USA and Indonesia; in spite of the mean of ‘Social Inf’ being lower in China (2.9) and the Netherlands (2.3) compared to USA (3.3) and Indonesia (3.3) (T-Tests: CN < IN/ USA: p< 0.001; NL < IN/ USA: p< 0.001). This phenomenon could be due to the influence of social norms that often go undetected by the influenced party . Alternatively, the higher effects of social expectations in China and the Netherlands could be due to the confirmation bias , when respondents are more likely to report higher social expectations if they have already undergone a Low Effort Measure (T-value=3.7, p=0.0). In USA and Indonesia, households report higher social expectations, but also are significantly more likely to have already undertaken both High and Low Effort adaptations (Figure 3). As such, while they report a higher prescriptive norm, it is less likely to inspire action as many households already conform to the norm.

The traditional general media has a likely null effect/ slightly negative effect on household adaptation intentions across all countries, except Indonesia. There it distinctly discourages house-holds intending High Effort measures (‘General Media’, Figure 2.a), possibly signaling distrust in information from the media . Conversely, social media has, in general, a positive effect on adaptation intentions for High and Low Effort Measures for the Netherlands and USA and lower/likely null in China and Indonesia. The internet in USA and the Netherlands are among the most ‘free’ and host generally unrestricted content. In Indonesia the internet falls on the lower end of the scale of ‘partly-free’ in terms of content restrictions and China’s is one of most censored in the world . Differences in content restrictions could play a role in influencing what people can post and read on social media, how much they trust the information, and the effect it has on adaptation intention.

Discussion and Conclusions

Across countries, disparities in the effects indicate that the social, institutional, and cultural contexts manifest meaningful differences in the drivers and constraints of household adaptations (Figures 1, 2). Given that effects also vary across measures, four countries are too few to make conclusions about generic differences across regions or cultures. Yet notably, while perceived cost universally discourages households’ adaptation, it is 2-4 times a stronger barrier in our two Global South countries compared to the two affluent Global North. Also, the curbing effect of undergone adaptations on motivation to pursue High Effort adaptations is 3-9 times stronger in Indonesia and China compared to the Netherlands and USA (Figure 1 and Supplementary Materials, Section S.4). However, for other factors the differences are not straightforward, like the revealed stronger negative effect of households’ beliefs in insufficiency of governmental measures in USA and Indonesia compared to the Netherlands and China.

Our unique dataset and analysis extends past research by refining assumptions about what commonly theorized factors of household adaptation are universal versus context-dependent, distinguishing between behavioral cues for High and Low Effort measures. Our work is the first step in comparing socio-psychological models across contexts. However, the coverage of four countries impedes a statistical attribution of cross-country variations in effects with specific contextual factors, limiting us to qualitative arguments of observed differences in effects of behavioral models. To generate large-scale cross-country data on households’ adaptation future work should survey multiple countries and consolidate existing fragmented survey data in joint globally-shared databases to permit quantitative analysis; for example using structural models to additionally tease out intra-variable relationships.

Further, our analysis of this single snapshot of data indicates a complex relationship between beliefs (in the urgency of climate change and in insufficiency of governmental actions), previously undergone adaptations and intentions to adapt. Future research should prioritize longitudinal designs to elicit if and how intentions lead to actions to assist in closing the intention-behavior gap. Panel data will permit monitoring the adaptation progress, its speed and effectiveness at the household level to complement the adaptation tracking of government-led adaptation . Furthermore, while survey length limitations in the present study compelled single-item measures for several socio-psychological constructs previously validated by others , future research would benefit from multi-item measurements.

Finally, a recent review stresses the importance of complementing interpersonal factors with intrapersonal (social networks, cohesion, and norms) when studying households’ responses to climate-induced hazards. Our study partially responds to this call by capturing prescriptive social norms, and finds a positive effect consistently in four countries. Future work could expand to study network and cohesion effects, and deepen to explore related social processes, like social amplification of risk or information cascades in networks . Computational social science methods, like networks and agent-based modeling, are especially pertinent to quantitatively study dynamic feedbacks between intra- and interpersonal factors. Notably, the revealed uniform strong effects in self-efficacy and perceived costs underscore the need to investigate adaptive capacity further. Other elements theorized to constitute households’ adaptive capacity – diversity, access to capitals, institutional capacity, and learning – should be systematically captured in future climate adaptation surveys. Understanding contextually-shaped patterns in household adaptation behavior will enable future empirical models to meaningfully extrapolate to data-scarce regions when projecting households’ adaptation progress.

Our findings have implications for climate change adaptation policies. To prompt household adaptation behavior, personalized narratives appealing to affect should complement communication of climate-driven probabilities and damages. Since social expectations consistently facilitate adaptation, associating desired behavior with a positive group identity could aid households’ adaptation diffusion and soften socially-constructed adaptation limits. Policies aimed at closing the adaptation gap by promoting diffusion of household-level action should target High and Low Effort actions differently: with undergone measures, beliefs and social influence having being key for Low Effort measures, vs. worry, self-efficacy, perceived costs and beliefs dominating in driving households’ intentions for High Effort adaptations. Importantly, knowledge on drivers and constraints of household adaptation should be transferred to new contexts with caution.

Methods

Data collection

In March-April 2020 we ran household surveys in flood-prone coastal cities in the United States of America, China, Indonesia, and the Netherlands. The surveys were conducted online by YouGov and the data analyzed and presented in this paper are from identical, translated questions in the respective languages of each country . The survey was written in English by the authors, one of whom is a native speaker from USA. For the non-USA respondents, the survey was professionally translated by YouGov field experts in each country, and the translation was reviewed by a climate adaptation scientist from each of the four case studies countries to help ensure cross-national relevance of the measures and aid in avoiding cultural bias. Further, YouGov field experts provided relevant information on national context, culture-specific ethical considerations and legislation that aided in the design of the survey.

Based on national statistics, YouGov forms representative panels. In China, Netherlands, and Indonesia we specifically controlled for gender representation, and age and gender in USA (see S.2, Supplementary Material). YouGov has a number of quality assurance measures in place, including excluding “speeding-respondant” (respondents who click through too rapidly to allow reading), inviting future panelists to participate, before announcing the topic - helping avoid the self-selection bias, and the verification of personal details given when respondents are registered for the panel. Further, respondents who consistently click the same (i.e. the first) answer are additionally filtered out. Finally, YouGov limits the number of surveys that respondents participate in monthly to reduce survey fatigue and conditioning [59]. The YouGov platform for online surveys is accessible via mobile phones, as such, according to the field teams, a lack of internet at home is not a barrier to reach the representative sample. As our research was focused on major urban centers, internet access was not a limiting factor . Employing an external company to run a survey of such scale is a necessity when running a large, cross national survey, however it can result in a lack of transparency and replicability. With YouGov’s extensive history conducting surveys for both academic, government, and corporate entities, we feel confident that theses drawbacks were minimized.

Dependent Variables

We study 18 household level flood adaptation measures (Supplementary Material Table S.5). We selected the relevant measures by reviewing prior empirical work guided by several meta analysis , as well as case studies that looked in depth at adaptation in each country i.e. . Here, we analyse adaptation intentions instead of already undergone actions to avoid issues with feedbacks with undergone measures on risk perception . Prompted by recent research , we group the adaptation measures into High Effort group - involving structural modification to ones home and necessitate significant time and financial investments; and Low Effort group - that include non-permanent flood mitigation actions as well as communication and information-seeking behavior. The two groups vary in the effectiveness of reducing hazard impacts and the extent of improving households’ resilience (compare raising ground floor level with seeking hazard-related information). During the survey, within each group, we randomized the order in which the respondents saw the adaptation actions. The grouping on the survey likely contributes to some within group consistency. See section S.3 in Supplementary Materials for factor loading’s and alphas on both groups.

For all adaptation measures, the respondent could select the following options:

I have already implemented this measure

I intend to implement this measure in the next 6 months

I intend to implement this measure in the next 12 months

I intend to implement this measure in the next 2 years

I intend to implement this measure in future, after 2 years

I do not intend to implement this measure

Options 2 - 5 were grouped together, by measure type, to indicate future intention. The questionnaire design allows us to construct a dependent variable based on the proportion of remaining measures a respondent can still pursue (the number not undergone) per measure group (Equation 1). This proportional formulation of the dependent variable helps maintaining consistency across respondents and accounts for the fact that different respondents likely have already undertaken a number of different measures. Already reflected in the reported sample size, our analysis of adaptation intentions excludes all households who had already undergone all measures in a given group as they have nothing left to intend.

This specification of the dependent variable has several advantages over other approaches of modeling intention to take multiple actions. Ordinal logit models, and count models do not explicitly incorporate the fact that many respondents may have already undertaken some of the measures asked and therefore cannot ‘intend’ to do something they have already done. Furthermore, count models such as binomial regression, assume Bernoulli trials, which we deemed potentially inappropriate in light of recent research that notes the connectivity between related measures . Binary logistic/ probit regression (that groups any intention as a 1 and no intention as a 0) overcomes this issue; but in grouping all intended measures together, even if the intended adaptation measures are in subgroups, information about quantity is lost. Therefore, we choose a ratio of the intention to pursue adaptations proportional to the remaining in the corresponding measure group as the dependent variable (Equation 1). While acknowledging that the likelihood of observing differences in effects is subdued and for measures specific variables (i.e. self-efficacy) averages must be used , we argue that this dependent variable is a good representation of household intention to pursue adaptation measures accounting for the ones already taken in the same group most accurately.

Explanatory Variables

The presented analysis focuses on flood adaptation measures and factors driving household intentions to pursue them. The survey design relies on an extensive review of the empirical adaptation literature aided by several meta-analysis . Six of the variables used in our analysis, make up the base PMT variables that often explains household adaptation intentions, and the remaining ten are variables frequently used to explain households’ protective actions against flooding. While not exhaustive of all tested constructs, we identified these 16 variables as drivers of household adaptation motivation that were regularly found to be influential in past work . The list of constructs, the questions used to solicit the variables, and their descriptive statistics are available in the Supplementary Material Table S.1. Leveraging the previous research that has extensively tested various survey items for these constructs, we select single questions to solicit the variables for questionnaire length limitations, common practice, and quality . As all of these variables has been previously studied we were able to compare effects to past work to help ensure the constructs were understood. We checked the variance inflation factors (VIF) of all variables in the model to ensure that multi co-linearity was not problematic (all VIFs < 2).

Data Analysis

To model the proportions of measures that households intend to take from those remaining, we estimate a Bayesian beta regression model. It preforms significantly better, based on WAIC scores (see Table S.5, Supplementary Material for more information), than other models we tested that can accommodate a proportions as a dependent variable (linear and logistic regression). Previous work has found that adaptation intention can occur ‘in concert’ , which can lead to bi-modal distributions. Our data confirms this finding and furthers supports the beta regression model choice selection, as the beta family is very flexible with regards to the array of density forms it can accommodate . Beta regression models cannot contain values exactly equal to one or zero, thus before estimating the model, we scale the dependent variable, the intention proportion values by group (Y), to fit between 0,1 Equation 2, .

The probability density function of the beta distribution is: where a, b >0, and Γ is the gamma function. We run all of our models in Python with the PYMC3 package . We parameterize the beta distribution in terms of its means (m) and standard deviation (σ). All coefficient priors in all models are broadly set as β ∼ N(0, 5) and all intercepts as: β0 ∼ N(0, 10). We set the prior variance as σ ∼ halfN(0.5), bounded at the upper end with , where δ is the minimum value of y, transformed by the inverse logit function for each country that, when input into the above function, determines the upper limit on the value of sigma . Next, we transform the values to the Beta distribution shape parameters (a, b) using:

Constructed from the a and b parameters shown above, Bayesian beta regression models are typically reparameterized and represented with: and γ = a+ b . Thus, where θ is a vector of regression coefficients and intercept, β = (β0,β1…β ) and y = (y 1 ,y 2 …y), the Bayesian beta regression model we consider is: where F(·) is the inverse logit function that transforms our linear combination of independent variables (x).

In various places throughout the paper we compare the relationship of a specific variable between countries via means testing with T-Tests or note the relationship between two variables either with a T-Test, Wilcox-Rank Sum test, or Chi-Squared test. For both epistemological reasons (this type of survey is repeatable) and ease of understanding, we use frequency statistics in these instances. Test scores and p-values are reported in text.