|Friday, February 19|
|CS11 Multivariate Analytic Methods||
Fri, Feb 19, 2:00 PM - 3:30 PM
Multivariate Inverse Prediction (Calibration) Using Standard Software for Mixed Models (303086)*Lynn Roy LaMotte, Louisiana State University Health Sciences Center
Jeffrey D Wells, Florida International University
Keywords: multivariate calibration, classification, inverse prediction
Use of insect evidence to estimate postmortem interval (time since death) motivates this presentation. More generally, given a multivariate response from a mystery specimen of unknown age, the objective is to compare it to training data from specimens of known ages in order to estimate the age of the mystery specimen. The problem can be viewed generally as classification or identification. The relations between response and age (or other factors) can be modeled as mixed models, with adaptive models for the mean and variance-covariance matrix. Comparison of the mystery-specimen response to the model at each age gives a p-value for the test of the response as a multivariate outlier at that age. Methodologically defensible interval estimates of the age of the mystery specimen can be based on these. This presentation will illustrate the formulation, implementation, and computations of multivariate inverse prediction in terms of widely-available software for analysis of mixed models. The presentation will benefit practitioners by providing a general framework and specific outline of steps that can be adapted to different settings and different software.