Differentiating placebo response from a true drug response in longitudinal depression studies gives rise to a variety of problems associated with functional data analysis. For instance, suppose the active treatment arm of the study can be modeled by two latent classes corresponding to responders and non-responders. These two latent classes may each consist of a variety of distinct functional response profiles. We propose a mixture modeling approach followed by a prototype analysis (e.g. k-means, principal points). The mixture analysis will determine the different latent classes and the prototype analysis will determine the variety of different response profiles in each latent class. Another issue to consider is the selection of basis functions to represent the functional data. We shall investigate the use of customized basis functions using a Gram-Schmidt procedure that will identify primary modes of variation in response profiles in longitudinal studies.
Available at: http://works.bepress.com/thaddeus_tarpey/39/