Abstract:

Diabetes (mellitus) is a chronic disease affecting hundreds of millions of people globally, a gateway disease that will substantially increase the risk of cardiovascular complications. To prevent or delay the onset of these complications for diabetics is challenging, in part because healthcare providers lack treatment guideline at clinic operational level. The current guidelines for these providers can be overly broad for medicating patients, or formed with a cohort of patients who have different socio-demographic/genetic profiles, leaving providers to rely greatly on their own experience and intuition in medication. A scientific and practical solution to this problem would be localizing an evidence-based medication guideline, namely, taking into account evidence from observational data (e.g., patient registries) specific to a given country or region. However, existing methods for causal inference with observational data have limitations that hinder them from being used in this purpose. In this paper, we are motivated by this setting, specifically, a provider has to decide among three treatment options: no-medication and two classes of medications (statins and fibrates) to control cholesterol for a newly-diagnosed diabetic. We develop a new approach that obtains causal effects of multiple treatments from observational data. We discuss how our approach overcomes several limitations with existing methods; in particular, the approach enjoys robustness to some model mis-specifications. By implementing this approach on a national registry of diabetics in Singapore, we empirically corroborate our theoretical findings.