How do we learn from biased data? Historical datasets often reflect\nhistorical prejudices; sensitive or protected attributes may affect the\nobserved treatments and outcomes. Classification algorithms tasked with\npredicting outcomes accurately from these datasets tend to replicate these\nbiases. We advocate a causal modeling approach to learning from biased data,\nexploring the relationship between fair classification and intervention. We\npropose a causal model in which the sensitive attribute confounds both the\ntreatment and the outcome. Building on prior work in deep learning and\ngenerative modeling, we describe how to learn the parameters of this causal\nmodel from observational data alone, even in the presence of unobserved\nconfounders. We show experimentally that fairness-aware causal modeling\nprovides better estimates of the causal effects between the sensitive\nattribute, the treatment, and the outcome. We further present evidence that\nestimating these causal effects can help learn policies that are both more\naccurate and fair, when presented with a historically biased dataset.\n