Identification and Estimation of Causal Effects by Cross-Temporal Design

• Activity Number: 247 - Causal Inference and Statistical Learning of Intervention and Policy Effects
• Type: Contributed
• Date/Time: Wednesday, August 11, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Health Policy Statistics Section
• Author(s): Yi Cao* and Roee Gutman
• Companies: Brown University and Brown University
• Keywords: Causal inference; Encouragement design; Instrumental variable; Cross-temporal; Bayesian analysis

Abstract

In estimating the causal effect of hospice on end-of-life expenditures, one obstacle is the possibility of unmeasured confounders due to data limitation. Specifically, important factor such as the self-preference for aggressive care is not collected from the Medicare claims data. To address this issue, we formulate a causal inference framework by taking advantage of the expansion of hospices between 2004 and 2009. We construct an encouragement design, treating the year (in 2009 or not) as a binary instrument variable (IV). Three types of hospice users: new-users (compliers), traditional-users (always-takers), and non-users (never-takers) are defined by the hospice enrollment status in 2004 and 2009. The stochastic exclusion restriction assumptions are violated due to the temporal effect introduced by IV. We propose alternative identifying assumptions that account for the temporal effect between the IV and outcomes. A Bayesian hierarchical model is constructed to estimate the temporal effect explicitly. We perform a simulation study to assess the estimation performance under different temporal trends and violation of assumptions.