Analysis of antiretroviral immunotherapy trials with potentially non-normal and incomplete longitudinal data

For many HIV-infected patients, use of antiretroviral therapy (ART) results in a sustained suppression of plasma viral load to undetectable levels. However, due to lack of antigenic stimulation, this may also result in a gradual loss of cell-mediated immune (CMI) responses that help control HIV infection. In concept, augmenting ART with periodic administrations of an HIV vaccine that boosts CMI responses could enhance control of viral replication. Researchers are designing 'antiretroviral immunotherapy' (ARI) trials to test this hypothesis. In a typical ARI trial, HIV-infected patients with sustained viral suppression will receive inoculations of an experimental HIV vaccine or a placebo, and subsequently stop taking their antiretroviral drugs. The goal is to assess whether plasma viral loads during the ART interruption phase are generally lower in the vaccine group. Assessment of a vaccine effect will be challenging if some subjects resume ART or drop out before the end of the treatment interruption phase. To tackle this 'missing' data problem and potential non-normality of the viral loads in ARI trials, we propose a two-step approach: multiple imputation of the missing values followed by use of the Wei-Lachin method with Wilcoxon scores. We use a numerical example and extensive simulations to illustrate the robustness and power advantages of our proposed method compared with other methods for incomplete longitudinal data, including REML, weighted GEE, last observation carried forward, and 'worst-rank' methods. Our proposed method is general enough for the robust analysis of longitudinal data in other therapeutic areas as well.