Correct power for cluster-randomized difference-in-difference trials with loss to follow-up
Cluster randomized trials with measurements at baseline can improve power over post-test only designs by using difference in difference designs. However, subjects may be lost to follow-up between the baseline and follow-up periods. While equations for sample size and variance have been developed assuming no loss to follow-up ("cohort") and completely different subjects at baseline and follow-up ("cross-sectional") difference in difference designs, equations have yet to be developed when some subjects are observed in both periods ("mixture" designs). We present a general equation for calculating the variance in difference in difference designs and derive special cases assuming loss to follow-up with replacement of lost subjects and assuming loss to follow-up with no replacement but retaining the baseline measurements of all subjects. Relative efficiency plots, plots of variance against subject autocorrelation, and plots of variance by follow-up rate and subject autocorrelation are used to compare cohort, cross-sectional, and mixture approaches. Results indicate that when loss to follow-up to uncommon, mixture designs are almost as efficient as cohort designs with a given initial sample size. When loss to follow-up is common, mixture designs with full replacement maintain efficiency relative to cohort designs. Finally, our results provide guidance on whether to replace lost subjects during trial design and analysis.
READ FULL TEXT