Exploring New Statistical Methods for Causal Inference in Longitudinal Studies

This study proposes a novel method to provide unbiased effect estimates in the presence of time-dependent confounding and applies the method in an existing merged multi-source nursing home dataset to examine the effect of antipsychotic medication use on all cause mortality. Using standard methods, effect estimates of time-varying exposures from observational data will be biased in the presence of time-dependent variables that are simultaneously confounders and intermediate variables. (Robins et al. 2000) This situation is common in clinical practice whenever treatment initiation, treatment choice, or treatment dose change over time and are based on surrogate markers that are affected by prior treatment. For example, in the treatment of human immunodeficiency virus (HIV)-positive patients with highly active antiretroviral therapy (HAART), CD4 cell count and HIV RNA levels represent time-varying surrogate markers that are affected by prior treatment and are strong risk factors for acquired immunodeficiency syndrome (AIDS) and death. In this context, only methods that appropriately adjust for such timedependent covariates are able to show the beneficial effect of HAART in observational data, while standard methods fail to find significant benefit. (Cole et al. 2003) Unbiased estimation of the mortality risk of antipsychotic treatment of elderly nursing home residents for behavioral symptoms of dementia from observational data is likely subject to similarly problematic conditions. For example, observed behavioral symptoms may simultaneously predict antipsychotic treatment initiation and Department of Statistics, Rutgers University. E-mail: [email protected]. Institute for Health, Health Care Policy, and Aging Research, Rutgers University.