Statistics Colloquium : Dr. Erniel B. Barrios

Abstract: In vivo clinical trials with free-living subjects are often confronted with limited number of volunteers to participate in the experiment. This is typically mitigated by considering the experimental units as their own control in a crossover design. A substantial washout period is planned to ensure that a possible carryover effect is eliminated prior to the administration of the next treatment level. Recent use of non-invasive devices to measure physiological functions of mobile subjects at regular intervals (often, real-time) facilitate collection of measurements at higher frequencies than a typical repeated measurements setup using conventional instruments. These devices may even collect similar responses even during washout periods. With mobile subjects freely doing their daily chores, covariates can easily skip the control of researchers and contaminate the responses. In analysing repeated measurements (with higher frequency) in crossover design, we propose a clustering methodology based on kernel density estimates to verify possible presence of treatment effects. The methodology includes a pre-processing technique designed to smoothen the noise attributed to unknown covariate effects in freely moving experimental subjects (possibly, with minimally controlled experimental conditions). Presence of clustering among observations in repeated measures is taken to indicate evidence of possible treatment effect. The method is evaluated using a simulation study to evaluate the algorithm’s sensitivity to presence of clustering. Increments in drift, autocorrelation parameter (representing possible carryover effect), and length of washout periods affect accuracy of clustering as well as computing time of the algorithm. The algorithm was used to measure treatment effect from baseline measurements in an endurance experiment using a corn-based energy bar.