Adjustments to propensity score matching for network structures

摘要

Causal inference from observational data rely on similar treatment and control groups to isolate for variation, in addition to adjustments in estimates to account for the remaining uncontrollable variation. Propensity score matching and statistical inference are established tools to achieve for these two requirements respectively. Network structures in the underlying data of the experiment challenge this convention since they question assumptions of independent observations and increase the risk of unobserved variables. In this paper we approach the problem with the intent of preserving propensity score matching and inference, while accommodating network information. Multiple experiments are re-evaluated with network information. All experiments were intended to create organic growth through referrals in a financial services business. We offer first, the Propensity Score Layout; a rapid visualization approach to scan data from multiple studies that potentially may require re-evaluation due to network structure. Second, the Propensity Score Network Risk; a metric that captures the extent to which network structure interferes with the treatment of the experiment. And third; variables constructed from network information that to our surprise estimate the propensity score significantly better than node attributes. We also present a set of interesting problems for researchers in academia and industry. To the best of our knowledge network methods have not been studied thoroughly in this domain. We feel the combination of technique, results and domain are novel.