Leveraging unsupervised machine learning to examine women's vulnerability to climate change

Abstract

We provide an application of machine learning to identify the distributional consequences of climate change in Malawi. We compare climate impact estimates based on drought indicators established objectively from the k-means algorithm to more traditional measures. Young women affected by drought were 5 percentage points more likely to be married by 18 than those living in nondrought areas. Our approach generates robust results when varying the number of clusters and definition of treatment status. In some cases, we find the design using k-means to define treatment is more likely to satisfy the assumptions underlying the difference-in-differences strategy than when using arbitrary thresholds. Projections from the estimates indicate future drought risk may lead to larger declines in labor productivity due to women's engagement in early age marriage than other factors affecting their participation rates. Under the extreme representative concentration pathway scenario, drought exposure encourages the exit of 3.3 million women workers by 2100.