Pragmatic cluster-randomized trials are widely used in health intervention research. Their use includes the assessment of quality care improvement programs in hospitals or primary care medical practices and the evaluation of interventions to improve health and behavioral outcomes in vulnerable populations (e.g., institutionalized elderly, people with mental illness, prisoners, individuals experiencing trauma, and victims of violence). The planning of cluster trials requires the judicious allocation of resources—that is, patients and costs associated with conducting a trial—while maintaining internal validity and providing sufficient statistical power to detect the hypothesized treatment effects. Given such finite resources, researchers choose a particular study design to maximize statistical information that will produce generalizable findings. The objective of this study is to formally quantify trade-offs involved when logistical, resource, and patient-centered considerations are balanced against methodological implications in cluster-randomized trials employing incomplete stepped-wedge designs in health intervention research. In these trials, clusters (e.g., hospitals, medical practices, nursing homes, prisons) start out in the control condition and transition to the intervention at randomly assigned periods, or “steps.” Most study planning methods are for complete stepped-wedge trials in which all clusters have outcome data collected in all periods. In pragmatic trials, however, considerations may require the use of intentionally incomplete designs in which all clusters are not followed in all periods.
Examples include staggered enrollment of clusters at the trial start and an implementation period needed to roll out the intervention. Our study will for the first time characterize the efficiency and validity of incomplete stepped-wedge designs with categorical patient-centered outcomes as well as continuous ones. First, we will develop statistical methods and software tools for planning stepped-wedge trials, based on the generalized information content of data and statistical power for incomplete designs. The new procedures will inform choice of efficient study designs for cross-sectional and cohort stepped wedge trials using generalized linear models for continuous, binary/binomial, and count patient-centered outcomes. We will also develop statistical power procedures, based on cluster period summary statistics for multiple outcome types in complete and incomplete stepped-wedge designs, to be analyzed by generalized estimating equations for fitting marginal generalized linear models that have population-averaged interpretations.
Second, we will develop statistical and computational tools for the planning and analysis of complete and incomplete stepped wedge cluster-randomized trials. The analysis software will include methods to identify outlying data elements (cluster, period, or cluster period) with large influence on study results, to establish robustness as well as statistical options for finite-sample bias corrections to account for the small numbers of clusters common to stepped-wedge trials. The software will consider four within-cluster correlation structures to address the Improving Methods PCORI Funding Announcement topic of “Methods to improve the design and conduct of cluster-randomized trials, including methods to account for the dependence of observations within clusters.”