Intention-to-treat vs. on-treatment analyses of clinical trial data: Experience from a study of pyrimethamine in the primary prophylaxis of toxoplasmosis in HIV-infected patients

Randomized clinical trials analyzed by the intent-to-treat approach provide unbiased comparisons among treatment groups. To avoid dilution of treatment effect, many people also perform an analysis by treatment actually received, although this method may introduce bias into the results. This paper presents several approaches used for analyzing data of a recent trial and the difficulties encountered in interpreting the results of each approach. The ANRS 005/ACTG 154 Study was a double-blind, placebo-controlled, randomized, international (French, U.S., and Spanish) multicenter trial designed to assess the effectiveness of pyrimethamine for the primary prophylaxis of cerebral toxoplasmosis (CT) in HIV-infected patients with advanced immunodeficiency. In the intention-to-treat analysis, the cumulative probability of CT at 1 year did not differ significantly between the pyrimethamine arm (11.9%) and the placebo arm (13.1%), Hazard Ratio (HR) = 0.94 (95% Confidence Interval (CI) = 0.62-1.42), whereas an on-treatment analysis resulted in a significant difference: 4.2% in the pyrimethamine arm and 12.4% in the placebo arm, HR = 0.44 (95% CI = 0.24-0.80). The data showed a significant interaction between compliance and treatment outcome; and side effects were more frequently cited as reasons for compliance violations in the pyrimethamine group. Several different analytic approaches (censoring data at the time patients discontinued the study medication only for selected reasons) failed to explain the disparity between the estimation of effect of pyrimethamine by the intention-to-treat and on-treatment analyses. This experience led us to believe that comparing the results of both analyses was the best method to convince clinicians that intention-to-treat was the only interpretable analysis. We were concerned that even if pyrimethamine had a beneficial effect, it was very difficult (1) to quantify and (2) to apply to clinical practice unless one could predict the occurrence of study drug discontinuation for each patient at the time of treatment assignment. Although exploratory analyses may yield clinically relevant information and useful clarifications in the evaluation of treatments, intention-to-treat remains the only interpretable analysis of clinical trials.