Analyzing Causal Effects with Observational Studies for Evidence-based Marketing at IBM

Sound marketing decisions often require understanding the cause-and-effect relationships between treatment and outcomes. Market research traditionally approaches such questions by designing randomized experiments that aim to isolate the effects of the specific treatment from other effects. We review an alternate methodology that is well suited to observational studies, where the analyst cannot control how treatment is applied. The methodology uses propensity scoring and matching to emulate the randomization of treatment. It is well established in other fields, but not widely known among marketers in spite of the fact that non-experimental data is common in marketing studies. We present two applications as case studies to illustrate the value of the methodology and to describe how we addressed some of the practical issues, in sufficient detail for readers to be able to use the methodology in similar studies.