Strategic Covariate Selection to Avoid Collider Bias

This direction is merged into another direction: collider bias. Its last version stays readable and citable here.

Version 1 of 132 papers34 conceptsfrozen 2026-10-04

In empirical research, there is a persistent intuition that "more control is better." Researchers often believe that by adding a "kitchen sink" of covariates to a regression model, they are purging it of bias and moving closer to a true causal effect.

However, a growing body of research across fields—from econometrics Burke et al. to neuroscience Dick et al.—demonstrates that indiscriminately adding control variables can actually introduce new biases, reverse the direction of observed effects, and obscure the very phenomena researchers aim to study. This problem is fundamentally one of causal structure, where the relationship between variables matters more than the quantity of data.

The Mechanism of Collider Bias

The most counterintuitive way that "controlling" for a variable can fail is through collider bias (also known as Berkson’s paradox). A collider is a variable that is a common effect of two other variables. When we condition on a collider—either by including it in a regression or by selecting a sample based on its value—we inadvertently create a spurious correlation between its causes.

As Joseph Y. Halpern notes in his review of Pearl’s work, this phenomenon means that "indiscriminately controlling for more variables can degrade the validity of causal inference" [Halpern].

Consider the "causal blind spots" identified by Nan van Geloven and colleagues. In some observational datasets, it appears that smoking protects against COVID-19. This is often an artifact of collider bias: if the study only includes hospitalized patients, and both smoking and COVID-19 severity influence the probability of being hospitalized, the act of looking only at hospitalized patients (conditioning on the collider) creates a false negative correlation between smoking and the virus [van Geloven et al.].

The Danger of Post-Treatment Variables

A second class of "bad controls" involves variables that occur after the treatment. Adjusting for these variables can lead to post-treatment bias in two ways:

  1. Blocking the Mechanism: If a variable is a mediator (an intermediate step on the causal path), controlling for it "absorbs" the effect you are trying to measure. Marshall Burke and colleagues point out that in climate-conflict studies, including economic indicators as controls can mask the effect of temperature shocks, because those shocks trigger conflict through their impact on the economy [Burke et al.].
  2. Inducing New Paths: If a post-treatment variable is affected by both the treatment and an unobserved factor that also affects the outcome, conditioning on it opens a "backdoor" path that didn't exist before [Chihara et al.].

Beyond Heuristics: Formal Adjustment Criteria

Historically, researchers relied on simple heuristics, such as "never adjust for post-treatment variables." However, Ilya Shpitser and colleagues have shown that these rules are often too restrictive. They developed a complete adjustment criterion that determines exactly which sets of variables are valid for adjustment based on the underlying causal graph [Shpitser et al.].

In complex scenarios, such as when data suffers from both confounding and attrition (dropout), researchers like Johan de Aguas have introduced Sequential Adjustment Criteria (SAC). This framework recognizes that while post-exposure variables can be dangerous, they may also contain the information necessary to correct for selection bias, requiring a nuanced, multi-step approach to adjustment [de Aguas et al.].

Automated Discovery of Valid Controls

Because researchers rarely know the full causal graph of their domain, new algorithmic methods aim to discover valid adjustment sets directly from data:

  • Invariant Learning: Methods like NICE Shi et al. and RAMEN De Bartolomeis et al. leverage data from multiple environments. They assume that while "bad controls" (like colliders) might show different correlations across settings, the true causal relationship remains invariant. This allows the model to automatically strip away variables that introduce bias [Shi et al., De Bartolomeis et al.].
  • Shadow Variables: The ShadowCatcher framework Li et al. generates "shadow" representations to specifically test for and correct collider bias without requiring manual identification of auxiliary variables [Li et al.].
  • Local Search: Rather than mapping the entire universe of variables, algorithms like LSAS focus on the local "neighborhood" of the treatment and outcome to identify the minimal set of variables needed to block confounding while avoiding M-bias (a specific form of collider bias involving pre-treatment variables) [Li et al.].

Conclusion

The shift from "controlling for everything" to "strategic adjustment" represents a maturation of empirical science. As Matthew James Vowels argues, the "theory crisis" in fields like psychology can only be solved by formalizing causal assumptions [Vowels & Hilpert]. By moving away from statistical significance toward structural validity, researchers can avoid the "garden of forking paths" and ensure that their control variables are clarifying the truth rather than obscuring it.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_strategic_covariate_selection_to_avoid_c_5b126881,
  title = {Strategic Covariate Selection to Avoid Collider Bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/strategic-covariate-selection-to-avoid-collider-bias/rdr_5b126881efa04af7be30ad437bdbb6df},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_strategic_covariate_selection_to_avoid_c_5b126881_v1,
  title = {Strategic Covariate Selection to Avoid Collider Bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/strategic-covariate-selection-to-avoid-collider-bias/rdr_5b126881efa04af7be30ad437bdbb6df/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}