Causal Inference Methodology for Social Work Research: A Sample Analysis Using Federal Administrative Child Welfare Data

Purpose: Drawing causal conclusions is often the ultimate goal for social science research, yet this can present a particular challenge to social work researchers, for whom ethical or practical concerns may limit the use of experimental or quasi-experimental designs.  

Advances in the theory of causality have made the isolation of causal effects theoretically feasible through statistical analysis even in the absence of experimental or quasi-experimental research designs (Pearl, 2000; van der Laan & Rose, 2011). Causal inference statistical methods are increasingly being used in public health, medicine, and other fields, but are not prominent in social work research to date. This presentation provides a heuristic example of how to apply new statistical methods to address questions of causality in social work research

Objectives: In this illustrative analysis, the authors examined the causal effect of maltreatment type (neglect vs. abuse) on the likelihood of reunification as a foster care exit outcome. The authors will present this study as a simple example to demonstrate causal inference methodology using a seven-step “causal roadmap” based on Pearl’s (2000) methodological framework.

Methods: This analysis used data from the Adoption and Foster Care Analysis and Reporting System (AFCARS) FY 2009 Foster Care Data file. The study population included all foster children in the United States whose case reached closure in 2009 and whose outcomes were reunification, adoption, kinship care/guardianship, or emancipation (N=209,169). The primary outcome variable was reunification and the exposure variable was maltreatment type (abuse or neglect). Covariates included child race, child sex, number of placements, and age at removal.

Applying causal inference methodology, we used non-parametric structural equations and a counterfactual framework to translate the scientific question (and its related background knowledge) into a causal model. We then identified a target causal parameter, assessed the identifiability of that parameter given the available data, chose a statistical model, and implemented targeted maximum likelihood estimation to quantify our target parameter.

Results: Results indicate that entering foster care due to neglect vs. abuse results in a nearly 8% lower likelihood of reunification for children with a low number of placements and a 2.5% lower likelihood of reunification for children with a high number of placements.

Implications: Our causal conclusions are only valid under the very specific assumptions we had to make to complete the analysis. Because some of these assumptions were implausible, the causal conclusions of the analysis are limited. These limitations will be presented in detail as the authors walk through the steps of the methodological roadmap. Nevertheless, this analysis provides a demonstration of how to apply causal inference methods to social work research in a systematic and rigorous manner.

Key advantages of this method are that it requires researchers to make interpretations transparent by making assumptions explicit as part of the research process, and that it respects the contributions and limits of background knowledge in the statistical modeling. As such, causal inference holds substantial promise as a methodological approach for social work researchers.