Directed Acyclic Graphs in Social Work Research

Significance. Social work research is conducted to test programs that promote social change. Research designs and methods that provide causal inference are important to this mission. All statistics are based on association only, however. Causal inference requires untestable assumptions that render these associations causal. For example, we might make the dubious assumption that there is no other way for income and education to be related other than through education causing income. A statistically significant positive association between education and income would then imply a causal effect.

To ensure that our assumptions are credible and that they are exhaustive in addressing threats to causal inference, a causal framework should be used. Most researchers are familiar with the nine threats to internal validity (Campbell framework), and many are also familiar with the Rubin potential outcomes framework. A pictorial framework developed by Pearl, the directed acyclic graph (DAG), has received less attention in social work research. This framework, when applied to researchers’ expertise in their fields, results in powerfully concise depictions of the challenges inherent in estimating a causal effect for a social work intervention. In this workshop, I will provide a complete description of DAGs for social work research.

Content. This is an applied workshop focusing on the uses for DAGs in non-randomized social work research designs. A DAG identifies the variables that need to be accounted for, and describes how these variables are related to the cause, the outcome, and to each other. I begin by breaking down the DAG into its constituent parts – nodes (variables), directed and bi-directed edges (arrows); and how these combine into chains, forks, and colliders. Subsequently I discuss the relationship between DAGs and statistical adjustments such as regression and propensity score analysis, including blocking, backdoor paths, and front door paths. I provide several concrete examples of DAGs.

I provide guidance on the use of DAGs in research design and analysis, contrasting this framework with the Campbell and Rubin frameworks, and demonstrating how they can be used in complementary roles to develop and conduct stronger research. I also cover important differences between each framework. I give several practical examples and I describe important challenges that researchers face when using DAGs. I will interact with the audience and develop a DAG in real-time from their suggestions.

Implications. Social scientists are often not clear about their causal assumptions, often obscuring the conclusions of otherwise well-designed studies behind assumptions that may not be credible. A causal framework translates the real world into credible assumptions that have statistical meaning. The Campbell framework for example describes the consequences of the different environments experienced by the treatment groups in a cohort study (history effects). DAGs go farther by providing a way to specify how these environments can be controlled for by measured variables in order to estimate a causal effect. DAGs not only provide a comprehensive framework for model specification and testing, but they also serve the pedagogical role of explaining more generally the challenges associated with causality in social research.