Thread by @p_tronc

What to do when you're analysing data arising from a Randomised Controllled Trial (RCT) and you have noncompliance in the treatment group?
A thread on Complier Average Causal Effect (CACE) modeling... /1
#Stata #RCT #CACE #gsem

In RCTs, we compare the outcome in the treatment and the control group. When everyone in the treatment receives the treatment in full or (comply), trial arm assignment is enough to estimate the effect of the intervention. We call this "intention-to-treat" (ITT). /2

The problem is that some participants may not receive the treatment in full or comply with all the requirements of the intervention itself. This produces bias in the ITT estimate, "diluting" the effect of the intervention, which is why we need to account for noncompliance. /3

An intervention effect doesn't occur because a treatment is offered, but because it's received. This is the main principle behind CACE, where we estimate the diff. bw participants who complied with the intervention and those who would've complied if assigned to treatment. /4

But why and how do you identify the individuals in the control group that would have complied? Why: because they are the counterfactual for those who complied in the intervention. How: by making assumptions and adopting a probabilistic approach (latent class). /5

That was a quick intro to the latent class approach to estimate CACE. For a more detailed discussion and an example using the Stata command "gsem", here are the slides for my presentation at the 2021 @Stata Conference: stata.com/meeting/us21/s… /6
@neilhumphreyUoM

Also, stay tuned for the paper that @A_moralesgomez and I wrote for the Stata journal, where we illustrate CACE using open data, with an audience of applied researchers in mind in the hopes this would be a useful resource for trial evaluators. /end
@SCADR_data