Measuring “dose” and “response” with multtvariate data using partial least squares techniques

Abstract

Existing statistical methodology is weak in exploiting the multiplicity of measurement in complex uncontrolled human teratology studies of “dose” and “response”. We describe a study of the neurobehavioral effects of moderate levels of prenatal alcohol exposure which examines 13 alternate measures of alcohol exposure (reflecting timing and variability of maternal alcohol consumption) with respect to hundreds of outcome measures over seven years of elapsed time. We demonstrate how one fundamental statistical tactic, the modeling of crosscorrelation matrices by Partial Least Squares, has permitted us to address the problem of jointly studying dose and response in terms of the salience of alcohol exposure for the outcome measures and of the outcome deficits for estimation of dose. Using this technique we summarize the apparently distinct mechanisms of alcohol exposure (in regard to timing and variability of dose) with respect to 474 outcomes organized in 15 blocks of variables at 5 ages.

Keywords:

• Fetal alcohol effects
• Indirect measurement
• Latent variables
• Singular value decomposition
• Teratology