04-conclusions.Rmd

# Conclusions

This body of work introduces study design methods and to improve the effective sample size, allow for greater personalization of intervention to subgroups, and adaptively follow studies until reaching a clear clinical conclusion.  Software and practical recommendations are provided for developing adaptive monitoring study designs.

First, we provide novel extensions to Sequential Matched Randomization which achieve greater covariate balance and efficiency than existing Sequential Matched Randomization and traditional methods such as Stratified Block Randomization. A dynamic and empirically-estimated matching threshold allows all patients to match and relaxes an assumption that baseline covariates are normally distributed. We allow matches to break and rematch if a better matches enrolls in the study.  Under our method, randomization-based inference achieves nearly the same efficiency as fitting an adjusted linear model to adjust for baseline covariates.  And, with greater covariate balance, an investigator may better investigate the response of subgroups to intervention.

Second, we introduce adaptive monitoring with the second-generation p-value which allows for following a study until ruling out effects deemed trivial or until ruling out effects highly actionable to change clinical practice.  This grounds the clinical trial not only on statistical significance but also clinical relevance, and can help reduce the risk of a trial ending with inconclusive findings.  We provide a case study through the REACH, a Vanderbilt University Clinical Trial, aimed to help patients with diabetes increase glycemic control and better adhere to medications.

Finally, we develop statistical software and provide practical recommendations for designing an adaptive monitoring trial using the second generation p-value.  The R package sgpvAM simulates data to estimate operating characteristics for these trials; the required simulation may be a barrier of entry to use this adaptive trial design.  We recommend a wait time before applying monitoring rules to control the classical Type I error.  When outcomes are not immediately observed relative to enrollment, we recommend increasing the number of observations before affirming a stopping alert.  This reduces the risk of stopping based on observed observations then finding a study inconclusive when observing the remaining observations.