Machine Learning a Probabilistic Structural Equation Model to Explain the Impact of Climate Risk Perceptions on Policy Support

While public opinion is a strong driver of policy change in democratic societies, the complex interactions of climate risk perceptions, belief in climate science, knowledge, political ideology, demographic factors, and their combined effects on support for policies aimed at mitigating climate change are not very well understood. This study applies an unsupervised machine learning approach to learn a “probabilistic structural equation model (PSEM)” for understanding such complex interactions. With foundations in Bayesian Network theory and information theory, PSEMs use the principle of Kulback-Leibler divergence to learn the relative importance of latent variables that explain structural dynamics of support for climate policy. A PSEM with R2 of 92.80% is derived from publicly available mixed-pool “Climate Change in the American Mind” (CCAM) dataset collected between 2008 and 2018 (N=22,416). The estimated PSEM predicts that 27.38% of the US population strongly supported climate policy action, while 59.46% were lukewarm supporters and 13.15% strongly opposed climate policy interventions. The conditional probability distributions of lukewarm policy supporters reveal a novel finding: Lukewarm supporters are more likely to be ambivalent about human induced climate change, less likely to be worried about climate change and more likely to be moderates and independents. Poor adoption of climate policy proposals in the US can be attributed to this silent majority of lukewarm supporters. Consistent with previous studies, we also find that strong supporters of climate policy are more likely to be alarmed and worried with relatively high and moderate risk perceptions and likely to be very liberal or somewhat liberal. In contrast, strong opposers of climate policy are more likely to be climate deniers, skeptics or doubtful, not concerned, risk deniers and very conservative or somewhat conservative. Theoretically we discover strong support for dual processing theory: while analytical risk perceptions have the largest effect size, this effect is mediated through affect/emotions, beliefs and ideology. We argue that data-driven machine learning models can account for complex interactions among latent variables to explain climate policy preferences. Future experimental research may be implemented to test whether emotionally sensitive communication of climate change induced risk may trigger a significant change in the policy preferences of lukewarm policy supporters to become strong supporters of climate policy.