Choosing control variables is one of the most consequential decisions
in observational research, and one of the least scrutinized. The common
practice of controlling for any plausible common cause of the treatment
and outcome may introduce bias (Achen 2005; Cinelli et al. 2022).
DAGassist provides a way to systematize adjustment
decisions and report the results.
Regression output can’t tell you which controls are wrong
The consequences of controlling for a variable depend on its causal role:
- Controlling for a confounder removes bias.
- Controlling for a mediator changes the estimand, from a total effect to something closer to a direct effect.
- Controlling for a collider or a descendant of the outcome introduces bias (Montgomery et al. 2018).
These distinctions are not visible in regression output. Consider two
regressions of voter turnout on income using the simulated
turnout_data included with the package. The first controls
for every available variable; the second controls only for the
confounders, age and state.
library(DAGassist)
library(modelsummary)
all_controls <- lm(turnout ~ income + state + age + polint + industry + elect_comp,
data = turnout_data)
confounders <- lm(turnout ~ income + state + age, data = turnout_data)
models <- list(all_controls, confounders)
modelsummary(
models,
coef_map = c("income" = "Income"),
stars = TRUE,
gof_omit = ".*"
)| (1) | (2) | |
|---|---|---|
| + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001 | ||
| Income | 0.281*** | 0.493*** |
| (0.016) | (0.016) | |
Both estimates are precise and highly significant, but they do not estimate the same quantity. Distinguishing between them requires assumptions about how the variables are causally related.
The gap between DAGs and regressions
Directed acyclic graphs (DAGs) provide a systematic framework for
choosing control variables (Pearl 2009; Elwert
2013; Hünermund et al.
2025). Several tools already make it easy to work with DAGs
in R. dagitty provides a syntax for specifying DAGs and
deriving adjustment sets (Textor et al. 2016), while
ggdag provides tools for plotting them.
What is less straightforward is checking whether an estimated regression actually follows from the DAG used to justify it. Reviews of applied research find that researchers’ adjustment decisions do not always correspond to the DAGs they report (Tennant et al. 2021). That discrepancy can also be difficult for readers or reviewers to detect from the information typically reported.
DAGassist is designed to close the gap between
representation and estimation. Given a DAG and a fitted regression, it
identifies the causal role of each control, flags controls that alter
the estimand or induce bias according to the DAG, and re-fits the model
using DAG-implied adjustment sets. It can target a specified estimand
and examine whether the conclusions depend on uncertain assumptions
within the DAG. For the turnout example, a single call identifies the
roles of the included controls:
DAGassist(turnout_dag,
lm(turnout ~ income + state + age + polint + industry + elect_comp,
data = turnout_data),
show = "roles", verbose = FALSE)
#> DAGassist Report:
#>
#> Roles:
#> variable role Exp. Out. conf med col dOut dMed dCol dConfOn dConfOff NCT NCO
#> income exposure x
#> turnout outcome x
#> age confounder x
#> state confounder x
#> polint mediator x
#> elect_comp nco x
#> industry nct x x
#>
#> (!) Bad controls in your formula: {polint}
#>
#> Legend hidden because verbose = FALSE. Re-run with verbose = TRUE to see role definitions.Design principles
Start from the researcher’s model. Pass a regression
formula through lm(), fixest::feols(),
lme4::lmer(), or another estimator with a formula interface
(supported engines).
DAGassist re-fits the model using the same estimator and
options while only changing the adjustment set.
Keep the original specification visible. The original model appears alongside the DAG-derived specifications. The purpose is to show how estimates change under different adjustment decisions, not to replace the researcher’s preferred model.
Make the estimand explicit. Because adjustment
decisions can change the quantity being estimated,
DAGassist distinguishes between the average total effect
and average controlled direct effect (Lundberg et al. 2021; Acharya et al. 2016).
Allow uncertainty about the DAG. Everything follows
from the DAG, so DAGassist checks how the adjustment sets
change if arrows are reversed or missing (Haber et al. 2022).
Make the results easy to report. Results can be exported as LaTeX, Word, Excel, plain text, and dotwhisker plots, making it possible to include the analysis in an appendix or response to reviewers.
Use existing tools where possible.
DAGassist relies on dagitty for graph analysis,
WeightIt and marginaleffects for weighting,
DirectEffects for sequential g-estimation, and
modelsummary for tables. Its role is to connect these tools
to the regression specification being evaluated.
Where DAGassist fits
| Task | Established tools | What DAGassist adds |
|---|---|---|
| Draw and analyze a DAG |
dagitty, ggdag, the DAGitty
web tool |
Accepts dagitty and ggdag
DAGs as input |
| Find adjustment sets | dagitty::adjustmentSets() |
Compares the model’s controls with DAG-implied adjustment sets |
| Fit models |
lm(), fixest,
lme4, and others |
Re-fits the original specification using DAG-derived adjustment sets |
| Target an estimand |
WeightIt, marginaleffects,
DirectEffects
|
Builds the weighting and sequential g-estimation specifications from the DAG |
| Question the DAG | dagitty::localTests() |
Examines how adjustment sets and variable roles change under uncertain or missing edges |
| Unmeasured confounding |
sensemakr (Cinelli and Hazlett 2020)
|
Identifies when a hypothesized unmeasured confounder prevents identification by adjustment |
| Report | modelsummary |
Exports the whole diagnostic, including variable roles and model comparisons, in one call |
What DAGassist does not do
- It doesn’t tell you whether your DAG is right. DAGassist does not learn causal structure from the data. Its conclusions are conditional on the DAG supplied by the researcher. The robustness functions robustness functions can be used to examine whether conclusions depend on uncertain or missing edges.
- It identifies effects by adjustment only. Designs that rely on other sources of identification, such as instrumental variables, the front-door criterion, difference-in-differences, or regression discontinuity, are outside its scope.
Who it’s for
DAGassist is intended for researchers using regression with
observational data who want to check and document their adjustment
decisions. It can also help reviewers and readers evaluate how a paper’s
causal assumptions inform its empirical specification. For instructors,
the package includes known-answer datasets (turnout_data
and toy_data) that illustrate the consequences of good and
bad controls concretely.
To try it, see Get started.