Research That Matters (January 17 - 20, 2008)

Children's paths through the child welfare system to a permanent exit (adoption, reunification or guardianship) are complicated. Currently, federal policy and the child welfare literature suggest that reunification with the birth parents is the preferred type of permanent exit. Traditionally the analysis of exits over time is conducted using survival analysis. One of the primary assumptions necessary for survival analysis is that the censoring be uninformed (i.e. that the censored data looks like a random subset of the uncensored data; or, there is no underlying mechanism driving the censoring events). This assumption is violated when the probability of one of a number of events occurring is not independent of another outcome (when there is informative censoring). This presentation will compare the results of a standard survival analysis for reunification with a two-stage statistical model accounting for the informative censoring present in child welfare data systems. This study, based on data abstracted from a longitudinal extract of the California Child Welfare Case Management System, examined all children who entered the California Child Welfare System for the first time in 2001 through 2003. Children were followed for two years from the date of entry to document instances of exits to reunification. Time to reunification was modeled using a standard survival analysis and using a two-stage model designed to account for informative censoring. Results suggest that the use of survival analysis to examine time to reunification miss-specifies the amount of time for children to reunify. The two-stage model removes children from the reunification risk set based on characteristics of the children that are known in the system. By incorrectly maintaining children in the risk set who have a low probability of reunification survival analysis methods overstate the time to reunification for the population. Comparisons of each model along with graphical representations of the survival analysis are presented to highlight the differences between the models. The results from this analysis suggests that the information currently being used on median time to an event such as reunification or adoption are most likely overstated because of the presence of informative censoring in the child welfare data. The implication being that scarce resources are being set aside to move some children toward reunification when this outcome is not possible. Policy and practice implications of the results will be discussed in detail.