Dichotomizing partial compliance and increased participant burden in factorial designs: the performance of four noncompliance methods

Noncompliance to treatment assignment is an inevitable occurrence in randomized clinical trials (RCTs). Intention to treat (ITT) is generally considered the best method for addressing noncompliance in RCTs. Alternatives to ITT exist, including per protocol (PP), as treated (AT), and instrumental variables (IV). These three methods define participant compliance dichotomously, but partial compliance is a common occurrence in RCTs. By defining a threshold, above which a participant is called a complier, PP, AT and IV can be used, but the resulting loss of information may affect their performance. Trials with factorial designs may experience higher rates of noncompliance due to the heavier burden that participants experience by being assigned to multiple experimental treatments.

Using simulations, we assessed the performance of ITT, PP, AT, and IV in both the partial compliance setting and in a 2-by-2 factorial design with increased participant burden for those randomized to both active treatments.

The bias, mean squared error, and type I error rates of the IV method after dichotomizing partial compliance were heavily inflated. The performance of all four methods depended on the level of noncompliance present, with higher average noncompliance leading to poorer performance. PP and AT showed improved bias and power relative to ITT without inflating the type I error beyond acceptable limits. However, the PP and AT heavily inflated the type I error rates when participant compliance was affected by the participants’ general health.

There are consequences for dichotomizing compliance information to make it fit into well-known methods. The results suggest the need for a method of estimating treatment effects that can utilize partial compliance information.