Methods for handling treatment switches


In a recent statistical seminar I attended, there was a discussion on statistical strategies to cope with treatment switching and the estimation of survival.

The presentation went around slides from Ian White(a)

Methods for handling treatment switching: rank-preserving structural nested failure time models, inverse- probability-of-censoring weighting, and marginal structural models

which are taken from the HTMR network workshop on Methods for adjusting for treatment switches in late-stage cancer trials.

In two words, the question is how to estimate overall survival (and treatment effect) when only progression-free survival is available and there was a treatment switch. ITT is way too conservative while PPA implies loose of randomization; regarding censoring at the time treatment switch occurs, it is not better because this would lead to informative censoring. Alternative methods that are discussed and illustrated in the talk are

  • Inverse-probability-of-censoring weighting (IPCW), which considers propensity scores, censoring at switch, and a Cox model with time dependent weights;
  • Rank-preserving structural failure time model (RPSFTM), which ignores censoring and relaxes covariate requirements (no need to account for factors predicting switch or outcome), assumes a common treatment effect (unlike IPCW).

The second strategy has been implemented in Stata, see White, IR and Walker, S (2002). strbee: Randomization–based efficacy estimator. The Stata Journal, 2(2), 140-150. It is also discussed in White, IR (2005). Uses and limitations of randomization-based efficacy estimators. Statistical Methods in Medical Research, 14(4), 327-347.

As discussed during the meeting, it is interesting to note that none of the proposed approaches actually take into account or even discuss interim analyses.

Be sure to check also the Clinical Trials Methodology Conference 2011 (all slides are available on-line).(b) Evaluation of methods that adjust for treatment switching in clinical trials is another talk on the same topic.

Here is another paper that looks interesting:

Morden, JP, Lambert, PC, Latimer, N, Abrams, KR, and Wailoo, AJ (2011). Assessing methods for dealing with treatment switching in randomised controlled trials: a simulation study. BMC Medical Research Methodology, 11: 4.

Finally, here is a recent review on adapative designs:

Chow, S-C and Corey, R (2011). Benefits, challenges and obstacles of adaptive clinical trial designs. Journal of Rare Diseases, 6: 79.


(a) Ian White has written a lot of handy Stata stuff for the analysis of RCTs with missing data or treatment switches.

(b) There are also a series of nice workshops with full access to lectures notes, handouts or code.


Articles with the same tag(s):

Academic teaching
Data cleaning techniques
Data Science from Scratch
Writing a book
Stata for health researchers
R Graphs Cookbook
Bad Data
Data science at the command-line
Reproducible research with R
Twenty canonical questions in machine learning