Background: To detect changes in biological processes samples are oftenmeasured at several time points. We observe expression data measured atdifferent developmental stages, or more broadly, historical data. Hence, the mainassumption of our proposed methodology is the independence between theobserved samples over time. In addition, the observations are clustered at eachpoint in time. The clustering is caused by measuring litter mates from relativelyfew mother mice at each development stage. The examination is lethal.Therefore, we have an independent data structure over the entire history, but adependent data structure at a particular point in time. Over the course of thehistorical data, we want to identify abrupt changes in the outcome - a changepoint.
Results: In this paper, we demonstrate the application of generalized hypothesistesting using a linear mixed effects model as one possible method for detectingchange points. The coefficients from the linear mixed model are used in multiplecontrast tests. The effect estimates are then visualized with simultaneousconfidence intervals. The figure of the confidence intervals can be used for thedetermination of the change point. Multiple contrast tests depend on the choiceof the used contrast. A variety of possible usable contrasts exists. In smallsimulation studies, we model different courses with abrupt changes and illustratedifferent contrasts. We found two contrasts, both capable of answering differentresearch questions in change point detection. Sequen contrast to detectindividual points of change or McDermott contrast to illustrate overallprogression. In addition, we show the application on a clinical pilot study.
Conclusion: Simultaneous confidence intervals estimated by multiple contrasttests using the model fit from a linear mixed model are usable to determinepossible change points in clustered expression data. The confidence intervalsdeliver direct interpretable effect estimates on the scale of the outcome for thestrength of the potential change point. Hence, scientists can define biologicallyrelevant limits of change depending on the research question. We found tworarely used contrast with the best properties to detect a possible change: theSequen and McDermott contrast. We provide R code for the direct applicationwith examples