Better Analysis of RCTs

Background: Standard Intent to treat (ITT) estimators are unbiased in large, randomized trials, but are often imprecise because they ignore covariate information. Augmented inverse probability weighted (AIPW) estimators combine the robustness of standard ITT estimators (the treatment model is known by design as a result of randomization) with the added precision from covariate-adjusted outcome models.

Objectives: Here we aimed to leverage AIPW to improve the precision of ITT estimators and illustrate better analyses of trials.

Methods: This is a reanalysis of 795 participants of the AIDS Clinical Trial Group (ACTG) 5202 phase 3 trial evaluating the impact of HIV treatment with abacavir/lamivudine (ABC/3TC) or emtricitabine/tenofovir (TDF/FTC). AIPW analysis combines inverse probability of treatment weighting with an outcome model to evaluate follow-up CD4 count (cells/mm3). The outcome model included the following baseline covariates: sex, age (categorized as 3 groups, 16-25 years, 26-49 years, and ≥50 years), viral load (log10 copies/ml), and CD4 count (cells/mm3). Inverse probability of treatment weighting used the empirical estimates of treatment to obtain the probability of receiving ABC/3TC versus TDF/FTC.

Results: The standard ITT and AIPW estimates of the difference in follow up CD4 count were -6.9 cells/mm3 and -6.7 cells/mm3, respectively. The standard error was 37% smaller in the AIPW estimate (SE:10.33) versus the ITT estimate (SE:16.34). To obtain a standard error similar to the standard ITT analysis, we would have only needed to enroll 416 (60%) of the 695 patients. Alternatively, to obtain the standard error given by AIPW the standard ITT analysis would require an additional 1030 participants (total n=1725).

Conclusions: AIPW estimation of ITT effects can improve the precision of trial analyses, and facilitate ethical and financially responsible trial design and ‘better' analysis.