Organizations in many areas of society, private and public, are eager to improve decision making using machine learning applied to records of past decisions and outcomes. Healthcare is one of many examples: electronic healthcare records are constantly updated with decisions on tests, treatments, procedures, and drug prescriptions. If used appropriately, machine learning has the potential to use this data to personalize and improve medicine. However, there are many hurdles on this path. In particular, current machine learning systems have been found to pick up on associations that are not causally related to the results of decisions. This leads to poor decisions when the system is applied to new problems or in new domains. This Ph.D. project will focus on developing machine learning methodology and theory for learning decision-making policies based on historical data.