Causal Inferences for treatment moderators on Zero-Inflated outcomes of HIV risk

University Of Pennsylvania, Philadelphia PA

Investigators

Abstract

DESCRIPTION (provided by applicant): This proposal addresses two statistical issues in substance abuse research (1) the development of causal inference model for count response variable with a high preponderance of zeros (i.e. zero-inflated outcomes), and (2) accommodation of treatment moderators to the causal inference modeling paradigm for zero-inflated outcomes. Our recent collaborations with the University of Rochester has exposed us to data consisting of zero-inflated outcomes as examined within two NIDA Clinical Trials Network (CTN) studies: CTN0018 (HIV/STD Safer Sex Skills Groups For Men In Methadone Maintenance Or Drug-free Outpatient Treatment Programs) and CTN0019 (HIV/STD Safer Sex Skills Groups For Men In Methadone Maintenance Or Drug-free Outpatient Treatment Programs). These studies examined the effectiveness of a five session gender-specific HIV sexual skill intervention relative to a one session treatment as usual standard HIV education intervention. Our previous research focused on standard modeling methods for assessment of moderation within a combined CTN0018 and CTN0019 database. Our colleagues at Rochester are focused on development of distribution-free models to assess moderation for zero-inflated outcomes in the combined database. These standard and distribution-free methods, however, do not address potential confounding between the moderator and other unmeasured variables. This in turn, has led to our interest in extending the causal inference modeling structure to accommodate moderation investigations with zero-inflated outcomes. Our general goals in this proposal are to (1) develop a causal inference statistical model for analysis of zero-inflated count data using Rubin's counterfactual framework under a specific Structural mean models (SMM; Robins, 1994; Robins, 2003) called the rank preserving model (RPM; Ten Have et al., 2007), (2) implement methods of the analysis of moderation under this causal inference framework, and (3) create software for implementing these new causal inference methods for zero-inflated outcomes. Within the context of the CTN0018/19 pooled database, we will test the new models by examining several clinically relevant moderators that were evident using standard moderation analyses: drug use severity, poly-drug use, whether or not the primary drug abused was cocaine, age, knowledge and understanding of male condom usage, and mini-mental status score. The new statistical methods and associated software will be disseminated so that other investigators can test for causal effects of moderator variables when study outcomes are zero-inflated.