Propensity score matching provides an alternative framework for causal inference when random assignment is not possible. The technique draws on core data science skills of predictive model building and algorithm development. Data scientists who need alternatives to experiments will find this a useful and accessible addition to their methodological toolbox.
Determining whether a marketing campaign caused an increase in sales, or a change to the landing page caused an increase in clicks, requires creating a counterfactual condition – a condition that can tell us what would have happened if the campaign or site change was not present. Random assignment, the critical process underlying A/B testing, is the preferred method for constructing valid counterfactuals. Random assignment is not always feasible, however.
This talk provides an introduction to propensity score matching, a method for creating a counterfactual condition that can be used to approximate an experiment when random assignment is not an option. Propensity score matching is built on the core data science skills of building predictive models and matching algorithm development. I will discuss when and how to implement this method in a business context, with a focus on best practices that have emerged from research and personal experience. Important limitations and special considerations for model building will also be covered. The material will be illustrated using examples taken from real-world projects.