Dropout due to loss to follow-up or death is common in prospective longitudinal cohort studies, such as the Women's Interagency HIV Study (WIHS), the Multicenter AIDS Cohort Study (MACS) and the Acute Infection and Early Disease Research Program (AIEDRP). Patients that drop out may be more likely to have disease progression, use drugs and alcohol or engage in high risk behaviors. Over time, the cohort evolves to be biased towards healthier subjects with lower risks. As a result, when estimating the relationship between drug and alcohol use (DAU) and longitudinal outcomes, analyses must consider subject losses or the impact of DAU will be underestimated. When the probability of dropout depends on the unobserved outcomes, even after conditioning on observable data, the missing data are missing not at random (MNAR) and therefore nonignorable. Despite the likelihood of nonignorable dropout, traditional methods, such as mixed- or random- effects models are frequently used. This may be partially due to the complexity level of existing statistical methods and the inability to implement methods using standard software. In addition, many investigators are na?ve to possible biases and the resulting loss of power. Mixture model methods account for the dropout mechanism by factoring the joint outcome-dropout distribution into the dropout-time distribution, f(u), and f(y|u), the distribution of the outcome given dropout. The resulting complete data distribution, f(y), is +f(y|u)dF(u). Misspecification of a parametric form of f(y|u) can lead to bias. Recently developed varying-coefficient mixture models can be used to semi-parametrically model the outcome- dropout relationship for a continuous outcome. The method is computationally stable, highly flexible and relatively simple to implement using standard software. We propose extensions to this method to accommodate differential dropout and new cohort enrollment effects;accommodate nonlinear relationships; and provide a more integrated method of B-spline knot selection through a Bayesian approach utilizing Markov Chain Monte Carlo methods. We will additionally develop a semi-parametric pattern mixture model to accommodate patterns of intermittent missingness. These innovative methods will be applied to the WIHS, MACS and AIEDRP data to accurately estimate patterns of active DAU in HIV-1 infected subjects by sex; determine predictors of needle sharing by DAU and sex;estimate the relationship between DAU and Highly Active Antiretroviral Therapy adherence by sex;and estimate the association between DAU and clinical HIV outcomes (CD4+ T cell count and HIV-1 RNA) in treated and untreated HIV-1 infected subjects.
Individuals with disease progression, drug use and other risk behaviors are more likely to drop out of the Women's Interagency HIV Study, the Multicenter AIDS Cohort Study and the Acute Infection and Early Disease Research Program due to death or loss to follow-up, such that over time, remaining subjects are healthier with lower risks. Analyses exploring the patterns of drug/alcohol use and the association of drug/alcohol use with HIV outcomes and medication adherence must therefore account for this subject loss. We will utilize new statistical methods to more accurately determine these patterns and associations.
|Forster, Jeri E; MaWhinney, Samantha; Ball, Erika L et al. (2012) A varying-coefficient method for analyzing longitudinal clinical trials data with nonignorable dropout. Contemp Clin Trials 33:378-85|