Causal Inferences with Linear Matching

• Activity Number: 395
• Type: Topic Contributed
• Date/Time: Tuesday, August 2, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and Marketing
• Author(s): Volker Bosch*+
• Companies: GfK SE
• Keywords: Matching ; Bias Reduction ; Propensity Score Matching

Abstract

In a controlled study design experimental and control group must be structurally identical to allow for causal inferences. However, in real life problems structures differ greatly due to selection bias. Thus, the groups must be statistically matched. In academia, state of the art is Propensity Score Matching where bias due to structural variables including interactions thereof is reduced to the same imperfect extent. In contrast, in market research matching restricted to the main effect level is preferred assuming that interaction bias is negligible. In any case, high test power for group comparison should be preserved. We developed a linear matching algorithm that perfectly aligns the structures of the test and the control group but ignores interactions of structural variables unless explicitly specified. Importantly, the algorithm is optimal under the given constraints with respect to statistical power. Studies on customer cards utilizing household panel data and on promotion effectiveness utilizing sales data demonstrate the advantages of the approach. Moreover, the results indicate that bias due to unspecified interactions is indeed negligible.