Session: Building & Using Directed Acyclic Graphs to Inform Causal Analysis in Social Work Research (Society for Social Work and Research 27th Annual Conference - Social Work Science and Complex Problems: Battling Inequities + Building Solutions)

Symposium Theme and Importance. Social work is a change-oriented profession. Social work research is undertaken to identify the leverage points for change-a population's risk and protective factors-and to develop and test interventions that promote change. These concepts-leverage points, interventions, and change-require us to make causal claims from statistical analysis. Typically, strong designs such as randomized control trials (RCTs) are used for causal inference, but social work researchers are making greater use of quasi-experimental (QE) studies. QE studies can be less costly, may lead to more generalizable findings than RCTs, and may address ethical concerns that are typical of RCTs, but they require stronger assumptions to support causal inference.

Importantly, when using QE methods, unobserved confounders of the hypothesized causal effect must be accounted for. A framework for demonstrating how specific observed and unobserved confounders can bias the causal effect can help. Directed acyclic graphs (DAGs) are a highly structured rule-based visual framework for representing the causal influences between variables. Built around the hypothesized causal effect, a DAG represents the myriad ways in which other characteristics of participants as well as the research design and setting affect estimation of the treatment effect.

Description of Studies. In this symposium of four studies, we will do an end-to-end demonstration of the use of DAGs in research design and analysis, including how they are constructed, how they are interpreted, and how they can be used to inform analysis with QE methods.

First, DAGs must be built primarily from the subject area knowledge about a particular social problem and, if applicable, the theory of change by which an intervention is hypothesized to have an effect. Using their own DAGs, our presenters will provide a grounded description of the approach taken to including and excluding variables and the causal paths between variables.

Second, in order to understand which confounders are critical to causal effect estimation, we must interpret the DAG to identify a "sufficient set". These are the confounders that, when controlled for, eliminate all possible confounding. Our presenters will describe a structured approach to identifying sufficient sets.

Third, in order to use advanced QE methods for causal inference, such as propensity score analysis (PSA) or instrumental variable estimation (IVE), certain key assumptions must be met. Research shows that these assumptions take very specific forms in a DAG, which makes it possible to select a method by looking for alignment between our DAGs and the form needed for a given method. In addition to choosing the method, these DAGs can be used to aid in model specification once a model is chosen. Two oral presentations show how to design and carry out analyses using DAGs. In the first, the DAG supports PSA as the method, and the DAG is used to minimize the risk of inducing additional bias from poorly conceived predictors of treatment assignment. In the second, a DAG is used to choose IVE as the method, to identify the instruments, and to justify that the assumptions of IVE were met.