Session: Applications of Multilevel Models for Child Welfare Research (Society for Social Work and Research 15th Annual Conference: Emerging Horizons for Social Work Research)

The phenomenon of nested structures is pervasive in child welfare research data and advocates the use of multilevel analytic methods. Specifically, research data on children in the child welfare system are frequently observed within broader contexts that require the use of corrective multilevel models to account for the correlation of observations taken within groups. For example, children are served by agencies and within jurisdictions that give them a shared experience with other children. As multilevel methods are increasingly being applied in child welfare research, there is a need to explore how different types of multilevel models can successfully be used to answer a variety of child welfare research questions.

This symposium will explore the application of three distinct multilevel analytical models for studying common child welfare research questions. Two of the analytic models are multilevel survival models using different longitudinal approaches to assess the rate of achieving permanency outcomes, and one of the analytic models uses a generalized linear model used to project the number of children who will age out of foster care.

The first multilevel model is a corrected Cox proportional hazard survival model used to account for the correlated data of children nested within county child welfare agencies. Specifically, the LWA marginal model is used to demonstrate how individual and county-level contextual factors relate to the rate of achieving permanency outcomes. To account for nested data, the LWA model estimates marginal distributions of the distinct failure times to produce a robust and optimal estimation of the variance-covariance matrix, which corrects for biases in standard errors.

The second multilevel model is a discrete time survival model applied to examine whether there are children who enter foster care having come from places where children generally have a better chance of going home, and, if so, how those places differ from others from a social structural perspective. This multilevel survival model application relies on discrete time so that one can more carefully study what happens within six months of placement.

The third model is a multilevel generalized linear model (HGLM) using a Poisson log-linear regression model to project the number of children who will age out of foster care at age 18. The data come from the Foster Care Data Archive with data from seventeen states. Since state child welfare agencies have varying policies with respect to care for children after age 18, we can expect the model coefficients to vary by state. An HGLM with children nested within states accommodates this aspect of the data structure. These three examples of multilevel models demonstrate the variety of uses for these analytic approaches for child welfare research. For example, multilevel models are beneficial for obtaining projections of child welfare caseload dynamics, which are essential for child welfare agency policy and planning. Using multilevel models to assess patterns in outcomes for children nested within broader contexts such as agencies or counties allows for a much needed understanding of the role of structural, organizational, and community factors in regard to child welfare outcomes.