Online Program

Estimating the Causal Effect of a Latent Class Treatment on Binary and Count Outcomes

Bethany C. Bray, Pennsylvania State University 
*Donna L Coffman, The Methodology Center, Penn State 
Stephanie T Lanza, The Methodology Center, Penn State 

Keywords: latent treatments, causal inference

Causal inferences methods, including propensity scores, are used more and more frequently in health policy research to estimate the causal effect of an observed intervention or exposure on an outcome. However, exposures may be latent variables in that the exposure may be measured by several questions, which are indicators of latent classes. The investigator may then be interested in estimating the causal effect of the latent class exposure on an outcome. One approach to estimating this causal effect was proposed by Kang and Schafer for continuous outcomes and implemented in an R package (lcca.r), and is called latent class causal analysis (LCCA). LCCA is based on Rubin's causal model and the treatment or exposure is multinomial. In addition, the treatment or exposure is latent, and we do not know which level of the treatment or exposure an individual received. LCCA makes use of both maximum-likelihood estimation and estimating equations; it imputes all unobserved potential outcomes and calculates the causal effects of interest directly. LCCA involves a model for selection into the latent classes and a model for imputation of the potential outcomes. In this talk, we will describe the LCCA approach, review its assumptions, and propose an extension to binary and count outcomes. We believe this approach will be valuable to public health scientists for estimating the causal effect of latent class behavioral profiles on a variety of outcomes.