As described in Chapter 5, an intent-to-treat analysis is often employed in randomized trials. In this type of analysis, the random assignment at the outset of the trial determines the treatment group in which a subject will be included for the analysis, regardless of whether the subject adhered to that treatment assignment. Therefore, patients who get assigned to a new therapy but for various reasons decide to discontinue it, or never to begin taking it, will still be considered as part of that treatment group for the analysis. This approach maintains the benefits of random assignment for the comparison of a new treatment against an older treatment, but at the cost of misclassification of actual treatment. Those who "cross over” from their assigned treatment to the other treatment group, for example, will be analyzed with their assigned treatment, ignoring the crossover. As a result, the analysis using the intent-to-treat principle incorporates some mis- classification of actual exposure. To the extent that the misclassification is independent of the study outcome, the misclassification will be nondifferential and will lead to underestimation of the effect of actual treatment.
Underestimation of the actual treatment effect is often considered acceptable, because it implies that a successful treatment is even better than the value estimated with the intent-to-treat approach. Nevertheless, as mentioned in Chapter 5, adverse effects of a treatment will also be underestimated by this method. This underestimation of risks is a serious drawback to using an intent-to-treat analysis for trials evaluating the safety of a treatment. In such trials, an analysis that classifies subjects according to their actual treatment may be preferred. Because an analysis based on actual treatment would not have all the benefits of a randomized comparison, the usual array of epidemiologic methods would have to be employed to assess and control confounding in the data analysis.