Background

Stepped-wedge trials are used to evaluate the impact of interventions. Researchers often use mixed effects regression to estimate effects. This method includes within-cluster - non-randomised - comparisons that requires assumptions about the secular trends.

Methods

We simulated data from stepped-wedge trials with different characteristics. We analysed these data using a within-step only approach and mixed effects regression, and evaluated their performance. The within-step only approach preserves randomisation by combining estimates of effect from within steps using a weighted average; we used non-parametric bootstrapping to generate inferential statistics. We introduced violations of the mixed effects model assumptions and investigated the effects on the two methods.

Results

When the assumptions were met, the mixed effects method was more sensitive and specific than the within-step approach. Bias was introduced to the mixed effects results by interaction of the secular trend with the clusters, and with the intervention. The within-step approach remained unbiased even in extreme violations of these assumptions. Comparing the mixed effects estimate of effect with the within-step estimate helped identify violations of the assumptions.

Discussion

We confirmed that mixed effects methods are more powerful than a within-step method when assumptions are met. Moderate to severe violations of assumptions led to bias, supporting the need for clear reporting standards and sensitivity analysis for stepped-wedge trials. Estimating the within-step effect can be useful for identifying bias.

Conclusion

Within-step analyses that preserve the randomisation should be used as a diagnostic to assess the validity of common mixed effects methods for analysing stepped-wedge trials.