←Back

439 – 439 - Topics in Marketing

Causal Inference and Machine Learning for Outbound Strategy

at Fidelity Investments

Telemarketing via outbound phone channel has been a popular channel for organizations to improve customer experience or increase revenue for decades. Randomized control trial (RCT) or A/B testing is often used to measure program effectiveness, and in this case, the impacts of phone conversations. However, because only a small portion of customers tend to pick up such calls, RCT can only adequately measure the average treatment effect (ATE) in an Intent-To-Treat (ITT) setting. In our talk, Instrumental Variable (IV) method is proposed to estimate the effect of phone interactions only for the reached audience which is formally referenced as the Local Average Treatment Effect (LATE). Armed with these additional insights (LATE), marketing managers can make more informed decisions on how to optimize future program effectiveness by either improving the phone scripts (if the treatment effects are small) or increasing the reach (if those effects are large but are diluted by a low reach rate). In addition to measurement and attribution, organizations are naturally motivated to improve program effectiveness by targeting a subset of customers who would generate relatively high individual treatment effects (or lift values). To address this second objective, a new approach is proposed to optimize the lift values by combining a call pickup propensity model (to increase reach) and an uplift modeling approach (to identify heterogeneous treatment effect). This integrated analytics framework can be leveraged to optimize the overall effectiveness of any intervention in an environment where it has limited reach. The reached audience is often referred to as compliers in RCTs. Finally, a realistic example will be used to illustrate the proposed framework.