I am trying to understand the effects of loss to follow-up on treatment effect. Following McMaster University’s nice workshop of how loss to follow-up can influence RR (https://training.cochrane.org/resource/assessing-risk-bias-when-loss-follow-too-high), I was wondering how to handle when a patient withdraws from a trial. Their data is usually retained until the date of withdrawal and used in the analysis. But what if, worse case scenario, all of these subjects die shortly after withdrawal? Should this missing data be treated like a loss to follow-up?
One approach to the dropout problem is to calculate the best-worst scenario. First, assume everyone who drops out from the intervention group dies, while everyone dropping out from the control group survives. That’s the worst case scenario. Then do the same, but assume every dropout from the intervention group survives, while the dropouts from the control group dies. That’s the best case.
You can’t know for sure what happened to the dropouts, but these extremes cover the range of possible values. This approach is discussed here.
you have information on why they withdrew?
Unfortunately not, just that they decided to withdraw
Thanks, that’s what I have seen for loss to follow-up, but haven’t seen any references for patients who actively withdraw from the trial, at any time point. Thank you for the reference.
As a get a slightly better grasp of the issue, I think this is mostly to evaluate the risk of bias due to dropout which is uneven between treatment groups. The workshop video on McMaster provides an example how, in two different settings and depending on measured treatment effect and LTFU, similar values can have a large vs. a small effect on the measured treatement effect. I see this range from worst - measured - best case as an indication of the robustness of the measured effect to LTFU.
with issues like this it’s always worth checking guidelines first: Guideline on Missing Data in Confirmatory Clinical Trials , because they will collate whatever wisdom exists on the topic
here they recommend adopting multiple approaches: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3651592/ including use of a time varying covariate. But maybe this creates ambiguity and confusion and thus pre-specify an approach as primary (with justification)
edit: i should also mention, maybe composite endpoints are used ie if you mean literally that “patients … actively withdraw from the trial”, it sounds like treatment failure, and might be treated as such in a composite outcome
Prospectively, the International Council for Harmonisation’s important estimand framework in the new guideline E9(R1) will help addressing questions about missingness such as withdrawals and intercurrent events: https://ich.org/page/efficacy-guidelines#9-2.