Conditional Inference Methods for Incomplete Poisson Data With Endogenous Time-Varying Covariates

原文摘要

We investigate the effect of protease inhibitors (PIs) on the rate of emergency room (ER) visits among HIV-infected women from a longitudinal cohort study. One strategy for accounting for serial correlation in longitudinal studies is to assume that observations are independent, conditional on unit-specific nuisance parameters. It is possible to estimate these models using unconditional maximum likelihood, where the nuisance parameters are assigned a parametric distribution and integrated out of the likelihood. Alternatively, we can proceed using conditional inference, where we eliminate the nuisance parameters from the likelihood by conditioning on a sufficient statistic for these parameters. An advantage of conditional inference methods over parametric random-effects models is that all patient-level time-invariant factors (both measured and unmeasured) are accounted for in the analysis. A limitation is that standard conditional inference methods assume that missing data are missing completely at random and do not allow endogenous time-varying covariates (i.e., past ER visits cannot predict future PI use). Both assumptions are unlikely to be met for these data, because one would expect “sicker” patients would be more likely to receive treatment and/or drop out of the study. We develop new estimation strategies that allow endogenous time-varying covariates and missing-at-random dropouts. The analysis shows that PI use reduces the rate of ER visits in patients whose CD4 cell count was <200 cells/ml at baseline. The size of the effect is substantially smaller than that estimated using a random-effects approach.

摘要来源：OpenAlex 索引摘要（稿件版本未注明）