Multilevel predictors of climate change beliefs in Africa

that people do not care about climate change and its consequences, but rather, they are seen as distant occurrences that can be dealt with when they disrupt or exacerbate more pressing mat- ters [44]. Demographics. Aside from the four previously discussed pathways that account for the climate change information and beliefs linkages, previous research also finds that gender, age, and race can play a role in shaping CCBs—yet a lack of consensus remains within the litera- ture. When it comes to gender, some studies find that females are less aware of climate change than males but have higher risk perception [16, 18, 30, 45], while others find no statistical rela- tionship between gender and climate change beliefs [8, 11, 46, 47]. Some studies have sug- gested that differences in the access to climate information could explain this difference [48]. Regarding age, many studies find that young people have more accurate climate change beliefs, however, this relationship is opposite in rural areas, where young rural dwellers are less aware of climate change [8]. Arguably, in rural areas agricultural experience gained with time pro- vides knowledge about changes in the climate cycle [41, 49]. Finally, race only appears to have a consistent influence on climate change beliefs in North America, where non-white individu- als show more concern and a higher risk perception than whites, a phenomenon known as “white-male effect” [7, 45, 50]. Nevertheless, this effect is not generally supported by cross-

country evidence [8]. In conclusion, evidence about demographic effects on CCBs remains elusive, and these effects seem largely eclipsed by the aforementioned information approaches [8, 18]. The discussion above identifies multiple factors that shape individual climate change beliefs. However, little research has included African countries in their analysis [10, 11, 49]. Those that do often include a limited number of predictors, which in turn, limit the under- standing of what predicts CCBs in the African continent. This study aims to fill these gaps within the literature. Research design Methods We analyse what predicts climate change beliefs in Africa using a random forest approach [51]. Random forest is a machine-learning approach that uses non-parametric recursive parti- tioning to produce models with high predictive accuracy [10, 52]. It can handle high-dimen- sional (large number of predictors) multilevel datasets with high-level interactions and non- linear relations [53], making it ideal for our dataset. For each dependent variable, we grow a classification random forest composed of 1,000 trees with a minimum node size of five. We use the ranger package in R, which is particularly well suited for high-dimensional data and offers a fast implementation that minimizes computational burden [54]. Its optimized compu- tation and parallelized processing are most suitable for our high-dimensional and extensive dataset. To the best of our knowledge, this is the first study to jointly combine survey and cli-

mate data at the second administrative level to explore what predicts climate change beliefs across 33 African countries. Random forest models are not easily interpretable on their own. To interpret them, we use some additional measures. First, we compute the variable importance measure, which ranks predictors by their predictive power (including direct and indirect effects on the dependent variable). We use the corrected Gini method to do so, because it shows no bias towards predic- tors with more classes, in contrast to the impurity importance, at a similar computational cost [55]. This measure shows which are the most important predictors of CCBs. Second, we use partial dependence plots to illustrate the magnitude and direction of the direct effects of signif- icant predictors. Partial dependence plots work like marginal effects in logistic regression PLOS ONE Multilevel predictors of climate change beliefs in Africa PLOS ONE | https://doi.org/10.1371/journal.pone.0266387 April 5, 2022 4 / 14