Back to Annual Meeting

Back to Annual Meeting

APHA Scientific Session and Event Listing

Jichuan Wang, PhD¹, Russel S. Falck, MA¹, Robert G. Carlson, PhD¹, and Peichang Shi, MS². (1) Community Health, Wright State University, 3640 Colonel Glenn Hwy, Dayton, OH 45435, 937-775-2084, jichuan.wang@wright.edu, (2) Math & Statistics Dept., Wright State University, 3640 Colonel Glenn Hwy, Dayton, OH 45435

Missing data or sample attrition is inevitable in longitudinal studies. If data are missing not at random (MNAR), attrition is associated with the response measure (e.g., with the initial value, slope, or true value of the response) and, therefore, associated with the unobserved data also. As a result, the widely cited problems of “selection bias” in linear least squares regression arise. Although MNAR is perhaps the most potentially damaging and frequently-mentioned threat to the value of panel data, applications correcting for non-random attrition-related bias are relatively sparse in longitudinal studies on substance abuse and HIV risk behaviors. This study demonstrates how to apply the sample selection models (Verbeek, 1990; Zabel, 1992; Verbeek & Nijman, 1992) to analysis of an 8-year longitudinal data on drug use among urban cracker users. Information contained in the attrition process will be integrated into analysis of the outcome of interest. Significance testing for non-random attrition will be conducted by assessing the correlation between the residuals in the outcome model and the probit model of attrition. A marginal maximum likelihood estimator will be used for parameter estimation of the sample selection model, using econometrics software LIMDEP (Greene, 2002). Since the sample selection models are a two-step estimator, a correction such as Murphy-Topel adjustment will be used to obtain appropriate standard errors of parameter estimates. Alternatively, the bootstrap method will also be used for such an adjustment for the purpose of comparison.

Learning Objectives: