The Use of Causal Dags to Guide Causal Inference and the Identification of Causal Effects

Quantitative researchers in social work are often interested in causal effects of interventions or other types of “independent variables.” The widely recognized “gold standard” for establishing such effects is the randomized controlled experiment (RCE). Yet RCEs are often unethical, infeasible, or both. Thus, researchers are often left with no alternative but to try to establish causality on the basis of observational or non-experimental data. Relying on observation data to infer causal effects, however, is fraught with risk. The crucial one is that a third variable may be associated with both the intervention and outcome of interest, a problem typically referred to as confounding.   

            Social work research isn’t the only area which confronts this problem; economics, sociology, and other disciplines do so as well. One which does so and which in many respects is similar to social work is epidemiology. What epidemiology has done to address this issue is adopt a mathematical way of representing causal effects which is based on the work of the computer scientist Judea Pearl.

            Pearl and others developed an approach for representing causal relationships referred to as causal directed acyclic graphs (DAGs). DAGs allow researchers to encode their beliefs about causal relationships in mathematical objects called graphs. Assuming the relationships in a given graph obtain, there are certain rules which researchers can use to determine whether the causal effect of interest can be identified. The answer to the identification question also provides guidance regarding which variable (or variables) need to be included in a statistical model to estimate a causal effect of interest.

            Another use of DAGs is that once a researcher uses one to encode their causal assumptions, certain conditional independence relationships between variables are implied. If these conditional independence relationships obtain in a given data set, then the DAG is considered to be consistent with those data, in the sense that the causal processes encoded in the DAG could have generated those data. If the implied conditional independencies aren’t actually found in the data set, then the causal processes encoded in the DAG couldn’t have generated those data. In short, this second use of DAGs allows researchers to determine if their causal assumptions are supported by the data.

            This workshop will provide an overview of causal DAGs, including details regarding the two uses of causal DAGs referred to above. DAGs serve other purposes, but I think these two are particularly useful for social work researchers. Once the basic concepts are illustrated, I’ll demonstrate their use on a real data set.  

            A word of warning is in order. Unlike many other quantitative innovations, DAGs aren’t really a new statistical technique or set of techniques for estimating relationships between variables, at least that’s not how they’re typically used in epidemiology. Instead, the innovation DAGs provide is a rigorous way of thinking about and offering guidance regarding which relationships should be estimated in the first place.