ITT analysis under equally high non-adherence in both arms in RCT

data analysis causal inference
,
1

Devereaux et al. ran an RCT comparing dabigatran vs. matched placebo. Patients were supposed to take the drug daily for 2 years. Analyses were done according to the intention-to-treat principle (ITT).

Notably, the study drug was permanently discontinued in 401 (46%) of 877 patients allocated to dabigatran and 380 (43%) of 877 patients allocated to placebo. This is the highest non-adherence rate I have ever seen.

image

Among patients who permanently discontinued study drug, the median time that patients took the drug was 80 days (IQR 10–212 days) in the dabigatran group and 41 days (6–208) in the placebo group. Among patients who did not permanently discontinue study drug, the median time that patients took the drug was 474 days (237–690) in the dabigatran group and 466 days (261–688) in the placebo group.

image

The primary outcome hazard ratio was 0·72, 95% CI 0·55–0·93; p=0·0115.

We all know that ITT analysis can be problematic in RCTs with long follow-up and/or differential non-adhrence between groups (ref 1 and ref 2). How should one interpret these results given such high non-adherence, albeit equally high in both groups?

3 Likes
2

I’d like to see an ordinal outcome analysis where inability to continue the drug or placebo counts as a moderately bad event. But directly to your question, the ITT estimates properly estimate the effect of the policy of prescribing the drug vs. prescribing placebo. An instrumental variable analysis could be attempted to estimate efficacy were adherence 100% in both groups.

1 Like
3

Thanks. But do you feel confident in such a classic interpretation even with high non-adherence? > 50%

4

I think the policy interpretation works under fairly extreme conditions. It’s good for informing initial treatment decisions. Once a physician sees the adherence trajectory for a patient, the new information may call for a revision of projections, e.g., using an instrumental variable-type analysis to estimate the likely treatment benefit for someone who adheres like the patient being treated.

1 Like
5

Sadly for the patients with blood clotting disorders - the unmet need remains unmet.