Many challenges arise in the analysis of pulsatile, or episodic, hormone concentration time series data. Among these challenges is the determination of the number and location of pulsatile events and the discrimination of events from noise. Analyses of these data are typically performed in two stages. In the first stage, the number and approximate location of the pulses are determined. In the second stage, a model (typically a deconvolution model) is fit to the data conditional on the number of pulses. Any error made in the first stage is carried over to the second stage. Furthermore, current methods, except two, assume that the underlying basal concentration is constant. We present a fully Bayesian deconvolution model that simultaneously estimates the number of secretion episodes, as well as their locations, and a non-constant basal concentration. This model obviates the need to determine the number of events a priori. Furthermore, we estimate probabilities for all ``candidate'' event locations. We demonstrate our method on a real data set.



Included in

Biostatistics Commons