Making Automatic Intervention Decisions Using Machine Learning

Abstract

Machine learning models are increasingly deployed to make decisions automatically, often with the goal of improving an outcome by means of an intervention. Common examples are influencing someone's purchasing behavior with an advertisement or affecting customer retention with a special offer. An increasingly popular approach to make such decisions is to use machine learning to estimate causal effects at the individual level, and then use those estimates to make intervention decisions automatically. However, estimating individual-level causal effects and making individual-level decisions are not the same thing. Deciding on which individuals to intervene is fundamentally a classification task where the decision maker should assign a positive class to individuals with a high causal effect and a negative class to individuals with a low causal effect. This thesis consists of three studies that show how this has critical implications for the estimation of treatment assignment policies from data because bias and variance in statistical models affect classification errors very differently than conventional regression errors. The first study shows that intervention decisions can be substantially improved by modeling them as a (causal) classification task rather than a causal effect estimation task. The second study shows that confounding bias does not necessarily hurt decision making, and that even when it does, the benefits of larger data may outweigh the detrimental impact of confounding. Finally, the third study shows that causal statistical models may not be necessary at all to make good intervention decisions when there is a suitable proxy variable for causal effects. Overall, this thesis shows that traditionally `good' estimates of causal effects are not necessary to make good intervention decisions, an encouraging insight given that modeling and acquiring data to estimate causal effects accurately is often complicated and expensive.