Common DAG Mistakes

A Directed Acyclic Graph (DAG) is a map of your causal assumptions that shows which variables influence which others. The most common mistake is adjusting for variables that look like confounders but actually introduce bias—particularly colliders, which are variables caused by two other variables. When you control for a collider, you create a false association between its causes, leading to collider bias. Another frequent error is controlling for all pre-exposure variables, which can accidentally adjust for colliders. The correct approach is to first identify all variables that are causes of the exposure, outcome, or selection into your study based on subject matter knowledge—then use the DAG to determine which variables block biasing paths without opening new ones. Try this: Before including a variable in your adjustment set, ask whether it could be a common effect of two other causes in your system.

DAG mistakes typically arise from three sources: misunderstanding causal structure, relying on data instead of assumptions, and using oversimplified rules for confounder selection.

The most dangerous mistake is conditioning on colliders. A collider is a variable with two or more arrows pointing into it—it is caused by multiple other variables. When you control for a collider through stratification, matching, or regression adjustment, you open a backdoor path between its causes, creating a spurious association where none exists. This is called M-bias when the collider sits on a path between exposure and outcome via two unmeasured variables. Even if a collider occurs before your exposure in time, adjusting for it can still introduce bias.

A second major error is controlling for all pre-exposure variables—a rule sometimes taught as "adjust for everything measured before treatment." This approach fails because it can inadvertently adjust for colliders. Instead, use the disjunctive cause criterion: control for variables that are causes of the exposure, causes of the outcome, or causes of both—but exclude variables known to be instrumental variables (causes of exposure with no other connection to the outcome), as these can amplify bias from unmeasured confounders.

A third mistake is drawing DAGs based on statistical associations rather than causal assumptions. Arrows in DAGs represent causal relationships that would exist if you intervened, not correlations in your data. Two variables can be statistically independent even when a causal arrow exists between them if effects cancel out in the population. Conversely, associations in data can be misleading guides for causal structure.

Finally, many DAGs omit important nodes like selection into the study or unmeasured variables. Including a selection node clarifies how people entered your analysis and whether conditioning on being included (such as restricting to survivors) introduces bias. Unmeasured variables should be drawn explicitly to make hidden assumptions visible.

Research on DAG pitfalls has identified systematic errors that recur across applied research. Suzuki et al. (2019) document ten common pitfalls, including the failure to distinguish between nodes as random variables versus their realized values, misunderstanding that omitted arrows represent strong assumptions of no effect, and neglecting that DAGs describe confounding in expectation rather than realized confounding from random allocation [1]. Greenland (2003) demonstrated that collider-stratification bias can be as severe as classical confounding, yet is often overlooked because traditional confounder selection criteria (associated with exposure, associated with outcome, not on the causal pathway) do not protect against adjusting for colliders [2]. VanderWeele (2019) formalized the disjunctive cause criterion for confounder selection, showing that controlling for causes of exposure or outcome (excluding instrumental variables) provides protection against both M-bias and insufficient adjustment, while the common-cause criterion is too conservative and the pre-treatment criterion is too liberal [3]. Tennant et al. (2021) conducted a review of studies using DAGs and found widespread variation in reporting practices, with many studies failing to report key information such as unmeasured variables, selection nodes, or minimally sufficient adjustment sets, limiting the transparency and utility of the diagrams [4].

DAGs themselves have inherent limitations that can be mistaken for user error. DAGs are nonparametric—they indicate the presence or absence of causal effects but not their magnitude, functional form, or whether effects vary across individuals. They also cannot distinguish between confounding in expectation (systematic bias) versus realized confounding (chance imbalances from randomization or sampling). DAGs require complete knowledge of the causal structure, which is rarely available; incorrect assumptions encoded in the DAG will lead to incorrect conclusions about confounding control. Some biases, such as selection bias from conditioning on a common effect, cannot be clearly distinguished from confounding bias even in DAG notation. Additionally, DAGs assume static causal relationships and require separate nodes for each time point when modeling longitudinal processes with feedback or time-varying exposures.