Adherence adjustment in randomized trials is possible, here's why

This post began life as a tweetorial, or twitter tutorial, for International Clinical Trials Day 2018. To read the original tweetorial, click here. The goal of this tweetorial was to explain the evidence for and against estimating adherence adjusted effects in randomized trial, and show that modern causal inference tools make adherence adjustment feasible.

We begin with a short trip back in time to 1980: the Coronary Drug Project (CDP), a large 6-arm placebo-controlled randomized trial of lipid-lowering medications published a comparison of adherers & non-adherers in their placebo arm. You can read the original CDP paper here:

Over 5 years, survival was nearly 10 percentage points higher among people who adhered to placebo most of the time compared to those who didn’t, even after adjusting for ~40 baseline covariates. A massive difference and bad news for proponents of the per-protocol effect. These results were interpreted as evidence that per-protocol effects could not be estimated without bias, because people shouldn’t be able to think themselves into living longer!

But that was the 80s and lots have changed since then, including statistical techniques for time-varying exposures. Before we revisit the data, let’s talk a bit about time-varying exposures. The intention-to-treat (ITT) effect is the effect of being assigned to one treatment vs another. In most trials, assignment happens once and that’s it, so the ITT is almost always about a point exposure. But when we want to know the effect of treatment itself, suddenly we have to worry about time – when can people adhere to a treatment protocol, and when did they adhere to the treatment protocol.

Even if the intervention only happens once (e.g. surgery), we might still define adherers based on when (if ever) they get treated. For example, if non-adherers are people were randomized to surgery but who never get surgery by the end of follow-up, that could be a time-varying exposure because we don’t know who the non-adherers are at baseline. If our intervention is a medication, and adherers are people who always take their medication (or take, say, >80%), that’s time-varying too!

Why does this matter? Because if the exposure varies over time, then baseline covariates aren’t going to cut it! If we have a time-varying exposure, we need to deal with time-varying confounding. The CDP didn’t do this, because they couldn’t – methods didn’t exist in 1980.

But it’s not the 80s anymore, and now we can deal with time-varying confounding. And that’s exactly what we did! We got the CDP data, and adjusted for post-randomization (time-varying) confounding.

First, we looked at cumulative incidence of mortality - just like the original paper. Using only baseline covariates, adherers to placebo did seem to have better survival than non-adherers. But once we included post-randomization covariates, the survival difference disappeared! You can read our paper here:

Cumulative incidence isn’t a great way to compare survival, though, so next we used a survival analysis approach. And we even tried lots of different ways to model adherence, to see if we were just getting lucky. We weren’t! When we adjusted correctly for post-randomization confounding, there was no survival difference between adherers and non-adherers to placebo! Correct adjustment meant: (1) IPW to prevent induced bias (2) a flexible function for adherence (not just linear!). You can read the 2nd paper here:

So we can adjust for adherence in randomized trials after all! And why is that good news? Because when we have non-adherence, the ITT is not an estimate of the effect of treatment. Compared to active treatment, the ITT could be an over- or under-estimate of the effect of treatment!

But the per-protocol effect is the effect of treatment. That’s the definition. Whether or not we can estimate this effect in a given trial is a separate issue. So, what are the options? And when should or shouldn’t we use them?

Option 1: (naive) per-protocol analysis
• Approach: throw out information on anyone or any person-time that’s non-adherent. Then just do your regular ITT-style analysis (maybe adjusting for baseline covariates)
• Why it fails: same problem as the 1980 paper — doesn’t account for time-varying confounding!

Option 2: as-treated analysis
• Approach: same as above but allow people to cross-over and pretend they’d been randomized to their cross-over group.
• Why it fails: Time-varying confounding strikes again!

Option 3: instrumental variables
• Approach: estimate the association between randomization and adherence and use to “correct” the ITT estimate.
• Why it fails: it doesn’t (at least, not all the time). But the simple version doesn’t work if adherence is time-varying!

Option 4: estimate the per-protocol effect
• Approach: same data as the naive per-protocol analysis PLUS adjust for time-varying confounding!
• Why it fails: it doesn’t! But it won’t work if you haven’t collected post-randomization confounder data!

So that’s it! Let’s recap
(1) Adherence is a time-varying exposure.
(2) Simple methods don’t work because of time-varying confounding affected by prior treatment (treatment-confounder feedback).
(3) But there are methods that do! If you have enough post-randomization data.


Very interesting write up, thank you!
I have one additional question.
I was always under the impression that one of the main issues with correcting for confounding by indication is that it is virtually impossible to fully correct for because generally there is a lot of unmeasured confounding. I can imagine that this holds for both baseline and time-varying confounders.
However, if your assumptions for an instrumental variable analysis hold, unmeasured confounding is not an issue. Does this make instrumental variable analysis the preferred choice if it’s available since it has the highest validity?

Four points:

First, yes unmeasured confounding is potentially an issue with the methods I described. But in the Coronary Drug Project, we have a good idea of what the mortality difference between placebo adherers and non-adherers should be over 5 years (i.e. zero). When we correct for baseline confounding only, we don’t get a difference of zero so we suspect there is still confounding. However, when we correct for time-varying confounding using an appropriate approach that deals with treatment-confounder feedback, we do get a difference very close to zero and feel more confident that we have accounted for all confounding.

Second, instrumental variables require different, but also strong assumptions. In the context of adherence adjustment for a trial, the most relevant are (a) exclusion restriction: are people aware of their treatment assignment and can they change their outcome risk through other mechanisms because of this?; and (b) monotonicity: no individuals in the trial who would have refused the intervention if they had been assigned to it and would have received the intervention if they had been assigned to placebo, a requirement about unobservable counterfactual adherence behavior. These assumptions are often difficult or impossible to assess. Even when we’re comfortable making them, the resulting causal effect is different than the one we can get using confounding-based causal inference methods – the IV effect applies only to people who would have complied with whichever treatment they were assigned to regardless of which treatment they were actually assigned to (a group defined by another set of unobservable counterfactual behaviors).

Third: The usual instrumental variable approach fails when we have time-varying exposures as is the case when we are interested in adherence to a daily pill regimen as in the CDP trial.

Fourth: Outside of trials, we also have to worry about no unmeasured instrument-outcome confounding, so it’s not always true that instrumental variables don’t have an issue with unmeasured confounding. Rather, the unmeasured confounding that can cause an issue is for a different set of variables.

1 Like

interesting. I guess non adherence often implies the use of other medications. I came across the following paper recently re adjusting for the use of nonstudy medications:

First, I’m not a statistician and apologize if I’m wasting your time. I am interested in adjusting RCT outcome measures for time-dependent adherence collected by smart blister packagers. I have one data point for each tablet taken over the course of a trial (typically 90 -180 / participant). The FDA guidance says I can’t use post randomization covariates to “adjust” primary outcome measures, for fear of treatment confounding adherence. Simplistically, could I not take the average adherence / patient, and test the Tx groups for homogeneity of variance (say an F-test), and regression slopes, and if NS assume minimal Tx - adherence effect? And then run ANCOVA? I read your article with great interest (esp. solution 3) but we’re talking pharma/FDA here and I can’t wait another decade for that to happen.
Thanks and kind regards, Allan Wilson