Directed Acyclic Graphs in Social Work Research: Frequently Occurring Types

Background & Purpose. Directed acyclic graphs (DAGs) are visual tools that aid in the specification of causal relationships between variables. Social work researchers can benefit from using DAGs to guide the design of prospective and retrospective studies. Researchers can use these DAGs to clearly state assumptions about how the putative causal effect of a non-randomized treatment can be confounded by factors not under the researcher’s control, leading to selection of measures and methods for addressing this confounding and estimating a causal effect without bias. Constructing a DAG is simple—doing so requires us to make assumptions about the causal relationships among variables and drawing unidirectional arrows between them, with the direction of the arrow indicating causal influence of one variable on another.

The rules needed to “deconstruct” and interpret a DAG, and determine the relevant confounders, can be more challenging. Rules pertaining to colliders (variables that have multiple causal antecedents) and selectors (variables that indicate a selection mechanism) are challenging, as are rules pertaining to the causal outcomes (descendants) of colliders. Notably, variables can play multiple roles, which suggests that it is the causal path rather than the variable that is relevant for researchers to understand. Researchers would therefore benefit from a simple typology of common DAGs that could easily be applied in most research contexts.

Method. I conducted a cross-disciplinary literature review based on permutations of the terms “directed acyclic” and “causal” graph. Based on this review, I classified these DAGs into a typology of DAG forms, connecting these forms to patterns typically seen in social work research, and to solutions that researchers might use to address the confounding inherent in these forms.

Results. Nearly all observed DAGs fall into one of six major types: the randomized design; confounder on a backdoor path; backdoor collider path; descendant of treatment on causal path; terminal (non-causal path) descendant of confounders or treatments; descendants of outcome. The challenges with causal inference inherent in each type are discussed, as are any known solutions. Clarifications are made for critical variations within each type, e.g., descendants of colliders that are not on the causal path that may result in virtual collider bias. Methods (such as the front-door path and instrumental variable design) are discussed.

Conclusions & Implications. Social work researchers are starting to rely on DAGs as a tool for causal inference. They can aid in measuring or identifying appropriate variables and choosing a method that might best eliminate these problematic relationships. I have identified a set of six major forms that describes most DAGs in the extant literature. Each form has its own attendant challenges and solutions, and forms can be combined into more complex DAGs. Foundational rules, as needed, can then be applied in situations that do not easily fit into one of these types. These forms may help support adoption of DAGs as a tool of social work research. Specific examples of these forms based on extant research support their utility for furthering research aims across social work research.