Adaptive Design for Observational and Experimental Studies

Classical scientific studies of humans tend to follow a standard pattern. First, an appropriate sample size is determined, often via a power analysis based on somewhat hazy assumptions. When the prespecified number of participants have been recruited, possibly randomized, and measured, recruitment is halted. Only after all data are collected, analyses are run; at most, one or two interim analyses may be run to ensure that safety norms are met.  While this process provides clean data for easy statistical inference, that cleanliness comes at a cost.  Errors in the power analysis’s assumptions lead to under- or over-estimates of the required sample size.  If power is lower than predicted, the study may show no findings at all.  If power is higher than expected, the question of interest could be answered early in data collection, and continuing the study is at least inefficient.  In some cases, it may even be unethical to continue to randomize patients into an ineffective treatment group or a control group where effective treatment is withheld.

This presentation will discuss adaptive design approaches for scientific studies that enable researchers use the data from their study while it is being collected, rather than waiting until the study is completed. Tradeoffs of these approaches for the researcher and  applications in clinical trials and observational studies will be highlighted.