iIn many randomized clinical trials, the primary analysis is an intention-to-treat (ITT) analysis, an approach based on the treatment assignment as randomized rather than the actual treatment received. One rationale for the ITT approach is that it evaluates the real-world effects of the intervention. However, a common misconception is that the ITT analysis will be unbiased regardless of crossover or missing data.
To understand the effects of crossover and dropout in an ITT analysis, it is useful to understand the 2 types of treatment effect that are generally of interest: the ITT effect and the average causal effect. The ITT effect measures the intervention effect as randomized; the average causal effect measures the intervention effect as actually received. In the ideal situation with perfect compliance and no missing outcome data, the ITT effect and the average causal effect are identical. This section of the Living Textbook considers the population-level causal effects in situations in which there is noncompliance or missing outcome data. Missingness in covariates may require further consideration.
In the presence of treatment noncompliance, the ITT effect and the average causal effect usually are not the same. In the absence of study dropout, the ITT effect can be estimated using standard methods and ignoring noncompliance. However, the ITT effect is diluted by crossover. A large crossover rate diminishes the ITT effect and reduces the statistical power of the analysis.
In the presence of dropout, the validity of a complete-case ITT analysis (ie, a standard analysis ignoring missing data) requires an untestable assumption that there is no selection bias by study dropout. This assumption will be violated, for example, if those who drop out of the study are "sicker" than those who do not. In such situations, even when the dropout pattern does not differ across treatment arms, the resulting naive estimators ignoring missing data would be biased for the originally targeted population-level ITT effect, provided the ITT effect in "sicker" participants is different from the general population. When the dropout pattern differs across treatment arms, the resulting naive estimators would be biased even for the ITT effect in the population represented by participants remaining in the trial. This assumption can be weakened to no selection bias by study dropout within levels of a set of measured baseline factors. Under such assumptions, valid ITT effect estimates can be obtained through methods that adjust for measured baseline selection bias due to study dropout (eg, inverse probability weighting or g-estimation).
For a detailed explanation using the causal counterfactual framework to understand these issues, see "Analyses of Randomized Controlled Trials in the Presence of Noncompliance and Study Dropout," a white paper from the NIH Collaboratory’s Biostatistics and Study Design Core.