5220.0: Wednesday, November 15, 2000 - 3:30 PM

Abstract #16054

Classification trees to account for selection bias for evaluating effects of a large public health program

Victoria G Lazariu-Bauer, MSc1, Howard Stratton, PhD2, Robert Pruzek, PhD2, and Mary Lou Woelfel, MA1. (1) Division of Nutrition, Evaluation and Analysis Unit, New York State Department of Health, 150 Broadway, 6th Floor West, Albany, NY 12204, 518-4027310, vgl01@health.state.ny.us, (2) Department of Biometry and Statistics, University at Albany, One University Place, Rensselaer, NY 12144

This study develops a methodology to adjust for selection bias in observational studies in the context of evaluating the Supplemental Foods Program for Women, Infants and Children (WIC). WIC provides supplemental foods and nutrition services to low-income pregnant women. Birth weight is used as a measure of success. Prenatal WIC records were matched to NYS 1995 birth records and WIC check redemption records, resulting in a database of 77,601 women. The variables explored to predict length of time on WIC included organizational characteristics, mother's demographic and medical characteristics, socio-economic characteristics of neighborhood, travel distance to WIC site. Rosenbaum and Rubin's propensity score analysis (PSA) was used to account for selection bias in this observational study. PSA estimates differences in outcome by conditioning on the probability of time on WIC. Decision to remain on WIC was described by a classification tree developed using a training, validation and test partition of the dataset. Predictions from tree are used to group women with the same likelihood to participate in WIC for similar lengths of time. Women were stratified using length of gestation. WIC effects were assessed within each strata. For each gestational stratum, women were grouped by time on WIC. The 10th percentile of the birthweight distribution for each gestational age was used to define low birth weight for that gestational age. The positive effects of time on WIC were substantial. Results were compared with more standard five-strata Rosenbaum and Rubin's PSA and with classical two-equation Heckman selection model.

Learning Objectives: Define the concepts of propensity score analysis and classification tree, describe a new way to evaluate a large public health program in the context of accounting for selection bias. Describe two methods for adjusting for selection bias and the differences between them, discuss issues related to implementing propensity score analysis methodology

Keywords: Statistics,

Presenting author's disclosure statement:
Organization/institution whose products or services will be discussed: New York State WIC Program
I have a significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.
Relationship: NYSDOH employee.

The 128th Annual Meeting of APHA