Complier average causal effect is a better estimate of the true causal treatment effect than intention-to-treat estimand in the presence of low treatment compliance: Insights from an instrumental variable approach

Abstract Background In the presence of substantial non-compliance to intervention, intention to treat (ITT) analysis may underestimate the effect of intervention on the outcome. In contrast, the complier average causal effect (CACE) provides a valid causal estimate of receiving intervention in the presence of non-compliance. This paper provides a user’s guide to estimating CACE following the instrumental variable approach in a simulated dataset of an individual-level randomized controlled trial (RCT) with non-compliance. Methods We simulated data for an RCT utilizing the simstudy R package, which generated potential outcomes and observed outcomes based on the assigned intervention and subgroups (Always Takers, Compliers, and Never Takers). Using the instrumental variable approach, we performed 20 simulations against all possible proportions of compliance ranging from 5 to 99% to generate an average for the true causal effect, ITT, and CACE. We created a graph comparing the averages of the three parameters i.e., true causal effect, ITT, and CACE by the proportion of compliance. Results We found that at an extremely low proportion of treatment compliance (i.e., below 10%), neither CACE nor ITT equaled the true causal effect. Above 10% compliance, the CACE (0.79–0.80) very closely approximated the true causal effect (0.80). However, the ITT (0.08 to 0.68) greatly underestimated the true causal effect at compliance of 10 to 85%. As the proportion of compliers increased above 85%, the ITT moved closer to the true causal effect, but it only became equivalent to the true causal effect when compliance was 99%. Conclusion The CACE estimand may be an alternative approach to ITT for estimating the causal effect of interventions in RCTs, particularly in the context of non-compliance. We recommend against the exclusive use of ITT when the proportion of compliers is below 85%. Instead of relying exclusively on the ITT to estimate the causal treatment effect in RCTs, we recommend using CACE in addition to ITT to allow for less biased estimates.