Extending Heckman's Treatment Model to Nonadditive Treatment Effects

• Activity Number: 33
• Type: Contributed
• Date/Time: Sunday, August 9, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Andrew Justin Spieker* and Joseph Chris Delaney and Robyn McClelland
• Companies: University of Washington and University of Washington and University of Washington
• Keywords: Biomarker ; Causal Inference ; Endogenous ; Observational study

Abstract

Often, we wish to evaluate the association between an exposure and a biomarker (e.g., race/ethnicity and blood pressure) using cross-sectional observational data. Depending on the age range of the study, a large portion of the participants may be on medication given to lower their biomarker value. Medication use is endogenous to both the untreated biomarker value and observable risk factors, which can result in substantial bias when using simple methods such as ordinary least squares. Alternative models such as censored normal regression and inverse probability weighting do not adequately remove this bias. In 1978, James Heckman proposed a likelihood-based approach which jointly models the outcome and the treatment. This hybrid model shows great promise when treatment effects are fixed and additive, but can perform poorly if the magnitude of treatment effect depends on the untreated biomarker value. We describe and present an extension of Heckman's hybrid model in which treatment effects are modeled as proportional to the untreated biomarker value. Part of our evaluation of this extension includes a simulation-based comparison to the simple approaches and the additive-treatment bas