Causal Discovery: Its Potential for Use in Social Work Research

Causal graphs, also sometimes called "Directed Acyclic Graphs" (DAGs) have been receiving increasing attention among social work researchers (see Rose, Cosgrove, and Lee, 2022; Lewis and Kuerbis, 2016; and Lewis, 2015). These graphs are often used to encode assumptions regarding which variables in a system are believed to cause others. Assuming a given set of causal assumptions are true and that there's interest in the causal effect of a specific variable, causal graphs can be used to determine, in order to identify the effect of interest, which variable(s) can be controlled for as well as which one(s) shouldn't be controlled for. This is the use of causal graphs that so far among social work researchers has received the most attention. Another use of these graphs, however, is for causal discovery.

Instead of starting with causal assumptions, encoding them in a causal graph, and using that graph to identify causal effects, causal discovery proceeds in the opposite direction. It begins with a set or subset of observations and uses search algorithms to try to obtain the causal graph that generated those observations. That is, causal discovery seeks to obtain information about causation from the correlations (or absence of them) found in a given data set or subset of a data set. Although causal discovery has been used outside of social work (see Ebert-Uphoff and Deng, 2012 and Lee, et al., 2022), it hasn't seen much use in our field. My paper will focus on three things.

First, I'll introduce causal discovery by focusing on the kinds of problems it can address and contrast it with the more typical use of causal graphs. Second, I'll review some of the basic concepts in this area. And third, I'll explore it's potential for social work researchers by indicating how it might be applied to a topic which is of interest to the profession.