Research That Matters (January 17 - 20, 2008)

Overview: Hierarchical data are commonly confronted in social science research: employees are nested within families, neighborhoods, and countries; students are situated within the broader context of the classroom, grade, and school district environments. People or events hierarchically structured within the same higher level unit tend to be systematically more similar than those drawn from another unit. Since a fundamental assumption of most statistical models is that observations are independent, clustered data are often in violation.

This analysis explores foster care exits to reunification using hierarchical modeling. While an ample body of literature has identified factors associated with a child's reunification, extant research has, by and large, either failed to account for the nested structure of the data or employed a random sampling design which discards valuable observations. In this paper we specify a reunification model adjusting for the correlated nature of nested factors at the family and county level. Two main research questions are addressed: 1) Is clustering at the family level as a means of controlling for sibling group membership preferred to randomly sampling one child from each family? 2) Is clustering at the county level necessary due to systematic differences in the likelihood of reunification?

Methods: Reunification data used for this analysis were drawn from a longitudinal extract of the California Child Welfare Services/Case Management System (CWS/CMS), housed in the School of Social Welfare at the University of California at Berkeley. Data were sorted and filtered using the unique child identifier variable and the episode placement count to create a cohort of children first entering foster care during the 2001 calendar year. These data were then merged with a file of county level data compiled from publicly available administrative data reports. Because a number California's counties have extremely small populations, and thus the rates for these counties are extremely unstable, all county level variables reflect a rate averaged across three years of data. The resulting dataset contains 26,963 unique children, 15,715 unique families, and 58 unique counties. Cluster identifiers were established at each of the three levels: child, family and county, with several covariates introduced for each level.

Results: Covariates at all levels were significant and consistent in direction and magnitude with previously published literature on reunification. The intraclass correlation for the reunification of siblings proved to be so high (.93) that it was concluded that randomly sampling a child from each family would produce a reliable estimate of the odds of reunification. At the other extreme, the intraclass, or within county, correlation for reunification of children was estimated to be nearly zero.

Implications: Findings support the method of random sibling selection found in the literature and also indicate that the assumption of independence of individuals across counties is not violated in these data. This analysis lends support to present methodological practices, while also highlighting the relative ease with which researchers are now able to explore the clustered nature of data.