Overview
A common tension in empirical work involves covariates that are affected by the treatment (aka “bad controls”), where there is an argument for (at least in some sense) trying to control for them, but also good reasons to avoid them. In Caetano, Callaway, Payne, and Sant’Anna (2026), in the context of difference-in-differences identification strategies, we provide two approaches to dealing with bad controls that respect the bad control being affected by the treatment while playing a genuine role as a covariate too. The badcontrols package implements these approaches.
Approaches
Our paper considers two alternative approaches to identifying the untreated potential version of the bad control:
Condition on the pre-treatment value of the bad control. This approach leads to a version of Callaway and Sant’Anna (2021) that involves conditioning on the pre-treatment value of the bad control. It is also simple and can be implemented with standard tools.
Assume covariate unconfoundedness for the bad control. This generalizes the first approach by modeling the bad control’s untreated evolution using additional covariates. This approach leads to a more complicated estimator, but the assumptions may be more credible in some applications.
For both approaches, badcontrols provides imputation, parametric doubly robust, and machine-learning estimators. The main function is didbc(), which follows a syntax similar to did::att_gt() and ptetools::pte_default().
Additional Resources
Paper: Difference-in-Differences with “Bad Controls”, by Caetano, Callaway, Payne, and Sant’Anna.
Installation
The development version can be installed from GitHub:
install.packages("remotes")
remotes::install_github("hugosantanna/badcontrols")Basic example
The package includes a simulation with a treatment-affected covariate X and exogenous covariate Z. The demo here is based on the covariate unconfoundedness assumption (approach 2 above), where we assume that unconfoundedness holds for the bad control after conditioning on the exogenous covariate Z, the pre-treatment value of the bad control, and the lagged outcome. The data that we generate below is a panel with 2000 units and 4 periods.
library(badcontrols)
sim <- simulate_bad_controls(n = 2000, T_max = 4)
head(sim$data) id period G D Y X Z W
1 1 1 0 0 -1.709723760 -1.05953298 -1.2070657 -1.3337493
2 1 2 0 0 -1.564713922 -1.76561193 -1.2070657 -1.3337493
3 1 3 0 0 -2.423375081 -1.63998935 -1.2070657 -1.3337493
4 1 4 0 0 -1.494701781 -1.16835526 -1.2070657 -1.3337493
5 2 1 2 0 -0.003018587 -0.04756332 0.2774292 -0.1592186
6 2 2 2 1 1.833864636 0.85624078 0.2774292 -0.1592186
res <- didbc(
yname = "Y",
gname = "G",
tname = "period",
idname = "id",
data = sim$data,
bad_control_formula = ~X,
xformula = ~Z,
bad_control_cov_formula = ~Y,
est_method = "dr_ml",
nuisance_method = "parametric",
bstrap = FALSE
)
extract_att(res)The same interface can be used with est_method = "imputation" or with nuisance_method = "ml" for cross-fitted machine-learning nuisance estimates.