This academic paper, recently published in the Journal of Causal Inference, tackles one of the "final bosses" of observational data analysis: the simultaneous presence of confounding bias and selection bias (specifically attrition).
The authors introduce a new theoretical framework and a robust estimator that can recover causal truths even when the data-generating process is messy and involves variables affected by the treatment itself.
TL;DR
The researchers introduce the Sequential Adjustment Criteria (SAC), a graphical method to identify causal effects when both confounding and missing data (attrition) are present. They also develop the Targeted Sequential Regression (TSR) estimator, a "multiply robust" tool that remains accurate even if several of your statistical models are wrong, outperforming standard methods in complex scenarios.
What Problem Does This Paper Address?
In an ideal world, we’d have randomized controlled trials with zero dropouts. In the real world, we have observational data where:
- Confounding exists: People who take a treatment are different from those who don't.
- Attrition happens: People drop out of studies or have missing records, often for reasons related to the treatment or the outcome (selection bias).
Standard causal tools often fail when the reason someone "drops out" is a post-exposure variable—something that happens after the treatment. For example, if a medication causes insomnia, and people with insomnia are less likely to show up for a follow-up test, the missingness is "informative." Existing graphical rules often forbid adjusting for these post-treatment variables because doing so can inadvertently open up new pathways for bias (the "collider bias" trap). This paper provides a formal way to navigate this minefield.
Methods
The authors contribute two major pieces of machinery:
- Sequential Adjustment Criteria (SAC): This is a set of graphical conditions based on "m-graphs" (causal diagrams that include missingness indicators). It identifies an s-admissible pair of covariate sets $(W, Z)$. Think of this as a "divide and conquer" strategy: $W$ handles the standard confounding, while $Z$ handles the bias introduced by attrition. Crucially, this allows for "forbidden" variables (descendants of the exposure) to be used in the adjustment process.
- Targeted Sequential Regression (TSR): Building on the Targeted Minimum-Loss Estimation (TMLE) framework, the authors propose a two-stage estimator. While standard estimators are often "doubly robust" (consistent if either the treatment model or the outcome model is right), TSR is multiply robust. It can survive misspecifications in various combinations of the treatment propensity, selection propensity, and outcome regression models.
Datasets
The paper validates these methods using:
- Synthetic Simulations: Two complex causal structures designed to induce "type-m insufficiency," where standard single-regression adjustments are mathematically guaranteed to fail.
- Norwegian National Registries: A real-world dataset of 9,352 Norwegian schoolchildren diagnosed with ADHD. The study linked medication history (from the Prescription Database) with national test scores in numeracy and reading, while accounting for the fact that children with more severe symptoms are more likely to be exempted from testing (attrition).
Key Insights and Ideas
- Breaking the Chronological Chain: Most causal methods insist on only adjusting for variables measured before treatment. This paper shows that in cases of attrition, adjusting for post-treatment variables is not just possible but often necessary, provided you use the sequential regression approach defined by the SAC.
- The "Safety Net" of Multiple Robustness: The TSR estimator provides an extra layer of insurance for researchers. In their simulations, when the selection model was wrong, TSR could still "self-correct" using the outcome regression, and vice versa.
- Practical Utility in ADHD Research: The application showed that ADHD medication has a positive but small effect on academic achievement (roughly a 3-5% increase in scores). This aligns with clinical intuition but provides a much more statistically rigorous estimate by accounting for the children who didn't take the tests.
Key Results
- Simulation Superiority: In scenarios with high missingness (50%) and model misspecification, the TSR estimator maintained low bias and better confidence interval coverage compared to Double Inverse Probability Weighting (DIPW) and naive plug-in regressions.
- ADHD Impact: The TSR estimate revealed an effect of 1.06 raw score points for numeracy (Cohen’s $d = 0.10$) and 0.66 points for reading ($d = 0.08$).
- Robustness to Misspecification: In Setup I, even when the exposure propensity model was intentionally sabotaged, the TSR and DML-based estimators remained unbiased, while the DIPW estimator's bias spiked significantly.
Limitations and Future Work
- The "Oracle" Graph: Like all graphical causal methods, SAC assumes you know the "true" causal graph. If your diagram of how variables relate is wrong, the adjustment sets will be wrong.
- Positivity Assumption: The method requires that every subgroup has at least some probability of being treated and being observed. If certain groups never take the test, no amount of math can recover their causal effect.
- Efficiency Optimization: The authors note that while they've shown how to find an admissible adjustment set, they haven't yet defined how to find the optimal one that minimizes the variance (noise) of the estimate. This remains a promising path for future research.
Summary updated