全球变化知识资源中心

Study sites.

We conducted social vulnerability assessments of 10 communities along the Kenyan coast in 2008, which were randomly sampled from a list of pilot sites undergoing changing coastal governance arrangements^{8, 28}. In 2012, we revisited eight of these communities (pirate and terrorist activity prevented us from re-visiting the other two sites).

Sampling.

We employed a combination of surveys targeted at resource users’ (fishermen and fish traders) households and semi-structured interviews with community leaders and fishermen. We conducted a total of 293 household surveys, and at least two key informant interviews per site. All interviews were conducted in Swahili by trained interviewers. We employed a simple random sampling strategy within a defined group (for example, resource users) for the household surveys, whereby respondents were randomly selected from lists of resource users provided by local leaders. Lists were cross-referenced with other fishermen and fish traders for accuracy.

Data collection.

On the basis of these surveys, we generated 9 socioeconomic indicators of social adaptive capacity (Table 1), which are adapted from previous studies^{8, 9, 14}. In addition, we used our household survey to examine four covariates: age of respondent, measured in years; fortnightly expenditures (converted to US$, and adjusted for inflation); migration status, measured as a binary metric of whether the respondent was born in the village (0) or somewhere else (1); and whether the respondent was actively involved in local decision-making (that is, attended meetings regularly and spoke), passively involved (attended but did not voice opinions), or not involved.

Analysis.

We conducted two types of analysis. First, we examined whether there were statistical differences in adaptive capacity indicators between groups based on: age; fortnightly expenditure; migrant/non-migrant status; and those who were actively, passively, or not involved in local decision-making processes. We fitted linear mixed models for our continuous response variables, binary logistic mixed models with a logit link for our binary response variables, and multinomial logistic mixed models with a cumulative logit link for our ordinal response variables (Table 1 and Supplementary Table 1). We generally included year as a fixed effect and community as a random effect (except when community was a fixed factor; Supplementary Table 1). We used spider plots to visualize these relationships.

Second, we examined whether each adaptive capacity indicator had changed over time in the eight communities where we had data from both 2008 and 2012. Our study examined data from independently drawn cross-sections of a population sampled at two points in time, in what are referred to as pseudo-panel data (as opposed to panel data, which track individuals over time). Pseudo-panel data are generally analysed by comparing the mean or median between the time periods, or between ‘cohorts’ of individuals within each time period^{29, 30}. Drawing on the former, we used linear mixed models for continuous adaptive capacity indicators, and generalized (binomial family) linear mixed models for binary or ordinal indicators: each adaptive capacity indicator was the dependent variable, with year as the independent variable, and ‘community’ as a random factor (Supplementary Table 1). This tested whether the mean of each indicator varied significantly over time, while explicitly accounting for the differences between communities.

Our research was an initial attempt to use applied field data to inform key debates about adaptive capacity, but the applied nature of the data and research design means that there are several caveats that should be considered: our design does not allow us to infer causation, which would require a proper experimental design, and/or the use of instrumental variables; there is potential for biased and inconsistent estimates because some of our indicators relied on recall data by respondents.