1Once you've drawn your DAG, the next step is to list all of the paths from the proposed cause to the outcome of interest. In this example, it means tracing all of the paths that go from road traveled to happiness.

2Start with road traveled and move your finger along each path until you get to another variable or to the outcome, happiness. When you encounter variables with arrows coming in or out, trace each path to the outcome. The direction of the arrows does not matter for tracing. Just don't trace back to variables you've already visited. In other words, no loops.

3In our example DAG there are five paths. Path one: road traveled leads directly to happiness. Path two: road traveled leads to social relationships, which leads to happiness. Path three: cognition points to both road traveled and happiness. Path four: background points to cognition, and background also points to income, which points to happiness. And path five: road traveled points to active, and happiness points to active as well.

4You can check your work, or skip the manual tracing process entirely, by creating and analyzing your DAG in R. While you can manually enter your graph into R, a shortcut is to create your DAG visually in your browser at DAGitty dot net, and copy and paste the model code into R. The paths function returns the same five paths we traced by hand.

5Next we need to determine which paths are good and which are bad. Good paths identify our research question. Bad paths are the alternate explanations for the causal effect that we need to close. In our example DAG, as in many DAGs, good paths often start with an arrow exiting the proposed cause, and bad paths have arrows entering the proposed cause.

6Here the five paths are labeled good or bad. Notice that paths one and two, the good paths, start with road traveled and flow forward to happiness. Those are the paths that carry the causal effect we're trying to estimate.

7The rest are backdoor paths that we need to close. Paths three and four have an arrow entering road traveled. Path five is a backdoor path as well, and it behaves differently, which we'll come to in a moment. Open backdoor paths bias the causal effect we want to estimate, so we need to close them.

8We can do this by conditioning on variables that confound the causal effect of interest, through techniques like regression. The neat thing is that we do not necessarily need to control for every possible confounder, just a minimum set sufficient to close all backdoor paths. Both DAGitty and the adjustment sets function in R will tell us which variables make up minimally sufficient adjustment sets.

9In this figure, you can see that adjusting for cognition in paths three and four is sufficient to close the open backdoor paths. Cognition is boxed in both of those panels, because one adjustment closes both.

10Path five is also a backdoor path, but it's closed by default, because active is a collider. Look at the two incoming arrows. If we condition on active, let's say by including it as a covariate in our regression, we will inadvertently open this path and introduce bias.

11There is such a thing as a bad covariate, and sometimes less is more when it comes to statistical models. And what if you need to adjust for a variable that's not in your dataset? The short answer is that you might be out of luck. The longer answer is more optimistic, but it probably involves creativity, some assumptions, and math.

12If this DAG is correct and complete, adjusting for cognition on the backdoor path makes the relationship between road traveled and happiness d-separated, or direction separated, from all other nodes. In other words, the causal effect is said to be identified.

13Now, you may have read lots of papers that adjust for confounding, and none of them used a DAG. You're right, and this is worth pausing on. If you've been reading the global health literature, you've seen countless papers that control for age, sex, education, wealth, and other variables without ever drawing a diagram. Are we doing something different here?

14Yes and no. Yes, in the sense that we're being explicit about our causal assumptions. The DAG forces us to think carefully about why we're adjusting for certain variables, and to communicate those reasons transparently. It's a tool for clarity and honesty. And no, in the sense that the underlying statistical machinery is the same. When a paper controls for confounders in a regression model, it's doing the same kind of adjustment we just described. It's simply not showing its work. The authors still had to make decisions about which variables to include and exclude, which means they had a causal model in their heads, whether or not they drew it.

15The difference is transparency. Without a DAG, assumptions often go unstated, but they're still there. Every choice about what to adjust for reflects beliefs about causal structure. DAGs just make those beliefs visible, to you, to your readers, and to your critics.