CS06 - Concurrent: - Crossovers and Repeated Measures

Repeated measures analysis is one of the most commonly used statistical methods, but our approach to analysis of this type of data often differs in practice. In addition, debate remains as to whether baseline values should be adjusted for in the analysis of RCT data with repeated outcome measures. Furthermore, when adjustment is made, there is often misunderstanding regarding the way this should be done.

On reading a paper by Twisk et al*, who evaluated three statistical models for analysis of RCT data with continuous outcomes (longitudinal analysis of covariance, repeated measures analysis and analysis of changes), I realised that the repeated measures analysis makes no adjustment for baseline values, as I had previously assumed. The authors advocated that baseline values should always be explicitly adjusted for in these analyses, and gave recommendations as to how this could be implemented under each of the three models discussed. Interested to evaluate these models further, I investigated in more detail using real data from an intervention study of a mobile phone application designed to improve wellbeing in young people in New Zealand.

In this talk, I will present my findings from application of these three models under different scenarios using a real-life dataset and linear mixed effects models. Analyses were of original study data (both including and excluding individuals missing data at any time point) as well as assessing the effect of inducing a baseline imbalance between groups (in both directions). I present results and discuss considerations for implementation in statistical software. Overall recommendations and implications for analysis of RCT data with continuous outcomes in practice will be made. This talk will help people to understand the differences in a number of repeated measures analysis methods and enable them to make appropriate analysis decisions.

*Twisk J, Bosman L, Hoekstra T, Rijnhart J, Welten M, Heymans M. Different ways to estimate treatment effects in randomised controlled trials. Contemp Clin Trials Commun. 2018 Mar 28;10:80-85. 

In randomized controlled trials with parallel arms, it is often necessary and beneficial for the participants randomized to placebo arm to continue the trial by crossing over to the treatment arm after the primary endpoint measurement at the pre-specified timepoint is recorded. In scenarios like this, it is necessary to develop an analysis methodology to integrate data from the main period of analysis involving parallel arms with the data after crossing over. One of the problems that arises in the cross over design is that the underlying conditions could change after cross over and thus making it difficult to control for when comparing with the base period. The results from the cross over period may have to be adjusted in some ways to account for these differences. Also, the homogeneity of results between the parallel arms and the cross over arms after adjustment need to be evaluated and the extent to which heterogeneity can be tolerated is assessed. Multiple methodologies for performing the integrated analysis are proposed, and the power gained under different scenarios are provided here. Repeated measures with different levels of correlations due to multiple measurements of a participant are taken into consideration in the integrated analysis. Several simulated scenarios are performed to adequately evaluate the performance and robustness of different analysis. Practical considerations that need to be considered in the analysis are also provided here.