Search

Abstract: In repeated measures data, large or small values at the initial measurement tend to be followed by values that are closer to the mean at the follow-ups measurements. This tendency is called regression to the mean (RTM). The presence of the RTM effect is inevitable in repeated measures data because of less than perfect correlation (correlation coefficient < 1) between the repeated measurements. Despite the growing evidence of the presence of RTM effects in clinical and public health studies based on repeated measures data, very few studies have evaluated and considered them when interpreting observed changes over time. In intervention studies, an RTM effect is mixed with an intervention effect in observed changes. It is extremely important to separate the RTM effect from the observed change in order to isolate any intervention effect and thus to make valid inferences about the effect of the intervention. In studying changes in outcome variables in repeated measures studies, RTM effects should always be adjusted for the valid interpretation of the changes and unbiased assessment of the intervention effects. The choice of methods to control for the RTM effect should be based on the type (continuous, count) and shape (normal, non-normal distribution) of the outcome variables of interest. A new method of estimating RTM effects for non-normal data using simulation is proposed. The method is a combination of bootstrap sampling from the standardized outcome variable and matrix decomposition of the correlation matrix between the repeated measurements. The method is applied to adjust for the RTM effects in studying changes in mean drinks in a typical week in a study evaluating the impact of a brief alcohol intervention on youth. In the study, mean drinks followed a positively skewed distribution. The proposed method estimated the RTM effects considering the true distribution (positively skewed) of the outcome and in doing so, provided more accurate estimation of the intervention effects compared to other methods considered in the thesis. The method ensured valid interpretation of the observed changes in the outcome by providing the most accurate estimation of the RTM effect and then removing it from the data. The proposed method could be applied to adjust for the RTM effect in non-normal repeated measures studies.