Controlling for Population Density Using Clustering and Data Weighting Techniques When Examining Social Health and Welfare Problems

Abstract

Clustering techniques partition the unit of analysis or study subjects into similar groups by a certain variable, thus permitting a model to run on cases with related attributes as a control for sociodemographic differences. Though large-scale national surveys often provide a raw weight variable, when applied without transformation, yields no change in statistical results, thus leading to spurious conclusions about relationships between predictors and outcomes. For example, most research studies using various components of the National Longitudinal Survey on Youth (NLSY) data sets to test hypotheses do not employ a weighting technique or post-stratification procedure to normalize the sample against the population from which it is drawn. Therefore, this chapter will illustrate how an algebraic weight formula introduced by Oh and Scheuren (Weighting adjustment for unit non-response. In: Incomplete Data in Sample Surveys, Chap. 3. Academic Press, New York, 1983), can be used in path analysis to elucidate the relationship between underlying psychosocial mechanisms and health risk behaviors among adolescents in the 1998 NLSY Young Adult cohort. Using the NLSY sample originally surveyed from US population, the association between self-assessed risk perception or risk proneness and how that perception affects the likelihood of an adolescent to engage in substance use and sexual behavior is investigated, separated into clusters by mother’s race/ethnicity and educational attainment. To control for oversampling of under represented racial/ethnic groups, mathematically adjusted design weights are then implemented in the calculation of the covariance matrices for each cluster group by race and educational attainment, comparing non-normalized vs. normalized path analysis results. The impact of ignoring weights leading to serious bias in parameter estimates, with the underestimation of standards errors will be presented illustrating the distinction between weighted and non-weighted data. As an innovative statistical approach, this application uses a weighted case approach by testing the model on discrete cluster samples of youth by race/ethnicity and mother’s educational attainment. Determining public health policy initiatives and objectives requires that the data be representative of the population, ensured by transforming and applying the weight formula to the sample.