Session: Directed Acyclic Graphs: The Bridge from Theory to Methods in Quasi-Experimental Social Work Intervention Research (Society for Social Work and Research 26th Annual Conference - Social Work Science for Racial, Social, and Political Justice)

343 Directed Acyclic Graphs: The Bridge from Theory to Methods in Quasi-Experimental Social Work Intervention Research

Sunday, January 16, 2022: 11:30 AM-1:00 PM
Marquis BR Salon 10, ML 2 (Marriott Marquis Washington, DC)
Cluster: Research Design and Measurement
Roderick Rose, PhD, University of Maryland, Baltimore, John Cosgrove, PhD, University of Maryland at Baltimore and Allison Deitz, MSW, University of Maryland at Baltimore
Significance. Social work researchers conduct causal intervention research to determine if an intervention has its intended effect. For causal research, randomized controlled trials (RCTs) are the standard. By assumption, the causal effect in an RCT is not confounded by unobserved characteristics. However, RCTs are often not possible due to the research setting or question, and generalizability is often poor in the stylized settings that RCTs are conducted. In these situations, observational or quasi-experimental (QE) designs are typically utilized. When using QE designs, methods that account for unobserved confounders of the causal effect must be used.

In this context, researchers have typically relied on regression, but are increasingly using advanced methods such as propensity score analysis (PSA), regression discontinuity (RD), or instrumental variable estimation (IVE). Each of these methods addresses the confounding problem differently. Choosing the most appropriate method in a QE study requires leveraging the researcher's knowledge about the conditions of the research setting, existing theory and knowledge about the problem and intervention, and understanding of how each of these methods "fixes" the confounding problem.

A directed acyclic graph (DAG; or "causal" graph) is a visual aid that leverages theories and hypotheses to specify how a putative cause is connected to an outcome, including through confounders. For example, an intervention M may be proposed as a causal effect for outcome Y, but theory suggests a third variable "C" causes both M and Y (C->M and C->Y). This produces a confounding path M<-C->Y that leads to non-causal association and confounds estimation of the causal effect (M->Y). Once a DAG is constructed, specific rules govern how the causal effect can be estimated in the presence of these potentially complex relationships. Researchers should then use these rules to identify the best statistical method for "fixing" the confounding problem in their QE study.

Content. Connecting theories and hypotheses to DAGs and methods can be challenging in practice. This workshop on QE methods in social work research will demonstrate how to do this. First, we will cover the DAG rules governing the estimation of causal effects in the presence of confounders. Second, we will cover the confounding scenarios for PSA and RD, which both require a very specific DAG form. We will then cover scenarios for regression adjustment and IVE. Finally, we will offer some concrete steps for using DAGs to make the critical choice of method selection. Simulated and real-world examples will be used as illustrations.

Implications. DAGs are under-utilized in social work research. Although PSA is widely used, other methods such as RD and IVE are not. There are research settings where these methods are appropriate, and connecting the theories to the methods through DAG specification may help. Social work scholars have the content expertise to leverage knowledge about the social problems and their solutions, and the prevailing conditions in the research setting. This workshop will help build up the third piece of this strategy, which is improving how they match the method to the problem.

See more of: Workshops