Over-Adjustment Bias
Over-adjustment bias is what happens when a statistical model controls for a variable that lies on the causal path between the thing being studied and the outcome, rather than beside it. A confounder is a common cause of both and must be adjusted for. A mediator is a step on the way from one to the other, and adjusting for it removes part of the effect you set out to measure. The arithmetic is identical, so the output cannot tell you which kind of variable it was handed. The choice is a claim about causal structure, made before the analysis rather than discovered by it.
What it is
Imagine a cohort study reporting that a measured trait predicts living longer. The obvious objection is confounding: perhaps the people who score high on the trait are also richer, better educated and healthier to start with, and it is the money and the health doing the work. The standard answer is to adjust, putting income, education and baseline illness into the model so that the comparison is between people alike on those variables. Adjustment is not optional; an unadjusted association between a psychological trait and death is close to uninterpretable.
The trouble is that the same operation applied to a different kind of variable does the opposite of what is wanted. If the trait leads people to smoke less and exercise more, and smoking less and exercising more is how the trait reaches longevity, then smoking and exercise are steps on the pathway. Putting them in the model asks what the trait would do if it changed nobody's smoking or exercise, which is not the question anyone was asking, and the estimate shrinks towards zero as a result.
What makes this a genuine indeterminacy rather than a simple mistake is that the same covariate can plausibly be both. Someone who is already ill smokes differently and exercises differently, so smoking is partly downstream of baseline health, which is itself a confounder. With one measurement occasion and no randomisation, nothing in the data distinguishes these structures. Two defensible models then produce two defensible numbers, and the published paper is often doing exactly the right thing when it reports both.
The term belongs to epidemiology. Schisterman and colleagues' 2009 paper in Epidemiology is the standard reference for the phrase overadjustment bias, though the underlying problem is much older and has no single originator. The same problem is sometimes written as mediator-confounder ambiguity, which is the description used in our own article on optimism and longevity; the two names pick out the same thing.
There are three honest responses, and all three appear in the literature discussed below. Report every model and say which is which, so that a reader can see the estimate move. Estimate the mediated share directly with a mediation analysis, which replaces a binary choice with a quantity, at the cost of further assumptions about unmeasured confounding of the mediator. Or go outside the observational design altogether, with a genetic instrument or a within-family comparison, where the question of which covariates to include does not arise in the same form.
What is not honest is quoting one number without the model behind it. A percentage from a cohort study is an incomplete citation unless the adjustment set is named, because the same table usually holds a different number a few lines away.
In effect
Two defensible answers in one table, printed by the authors
Lewina Lee and colleagues' 2019 paper reports the same comparison three times for 69,744 women in the Nurses' Health Study. Demographics alone: an 18.6 per cent longer life span in the most optimistic quarter. Plus baseline health conditions and depression: 14.9 per cent. Plus smoking, alcohol, physical activity, diet quality, body mass index and attendance at a screening examination: 8.7 per cent. The authors do not pretend to resolve it. Their own sentence is that while it is possible that health behaviours are confounders in this association, such behaviours may also lie on the pathway linking optimism to longevity. If they are confounders, 8.7 per cent is the answer. If they are the mechanism, 14.9 per cent is. Only one of those figures circulates.
Answering the ambiguity with a quantity instead of a choice
Hayami Koga and colleagues took the better route in their 2022 cohort of 159,255 women. Rather than picking a model, they ran a formal counterfactual mediation decomposition, allowing for exposure-mediator interaction because they found one, and put the share of the association mediated by lifestyle at 24 per cent. Critically, the lifestyle composite was measured three years after the optimism assessment, which fixes the temporal order instead of assuming it. Mediation analysis does not dissolve the problem, since it imports its own assumptions about unmeasured confounding of the mediator, but it states a number where the model comparison forces a verdict.
A loyalty programme measured against the thing it was built to change
The commercial version of this error is routine in product analytics. A retail team wants to know whether a loyalty programme raised revenue per customer, and adds visit frequency to the model as a control, because frequent visitors obviously spend more. But raising visit frequency is how a loyalty programme is supposed to work, so the resulting estimate answers a question nobody asked: what the programme would be worth if it brought nobody back more often. The same structure appears in life insurance underwriting, where smoking status and body mass index are legitimately priced in, and where a trait that reached mortality partly through smoking would show almost no residual value once smoking was in the model. Neither is a statistical failure. Both are a specification choice presented as a measurement.
What it does not say
It does not say that adjusting for health behaviours is wrong. Sometimes it is exactly right, and an unadjusted estimate can be badly confounded. The error is not adjustment, it is silent adjustment.
It does not say that the smaller estimate is the true one. Over-adjustment pushes estimates towards zero, so the behaviour-adjusted figure is a lower bound under one causal story and the correct answer under another. Reporting the smallest available number is no more honest than reporting the largest.
It does not say that mediation analysis settles the matter. A mediated share of 24 per cent rests on the assumption that the mediator itself is not confounded by something unmeasured, which is an assumption and not a finding.
And it does not, by itself, rescue or refute any particular literature. Pointing out that a covariate might be a mediator is not an argument that the association is real, and pointing out that it might be a confounder is not an argument that the association is spurious. It is an argument that the published estimate has a range, and that whoever quotes a single point from that range owes the reader the model it came from.
This page describes a methodological problem rather than a testable effect, which is why it carries no evidence status beyond "not a research claim". One sourcing limit, stated plainly: the Schisterman reference is recorded from a published reference list and has not been read by this publication, and no forename is given for its first author because none is recorded here.
Who developed it
No single originator. Schisterman and colleagues, in Epidemiology in 2009, are the standard reference for the term overadjustment bias in epidemiology; the underlying problem of conditioning on a mediator is older than the name.
Sources
- Lee, L. O., James, P., Zevon, E. S., and colleagues (2019). "Optimism is associated with exceptional longevity in 2 epidemiologic cohorts of men and women." Proceedings of the National Academy of Sciences 116(37), 18357-18362, doi:10.1073/pnas.1900712116. Read in full, including all three model specifications and the sentence quoted above.
- Koga, H. K., Trudel-Fitzgerald, C., Lee, L. O., and colleagues (2022). "Optimism, lifestyle, and longevity in a racially diverse cohort of women." Journal of the American Geriatrics Society 70(10), 2793-2804, doi:10.1111/jgs.17897. Accepted manuscript read in full, including the mediation decomposition.
- Kim, E. S., Hagan, K. A., Grodstein, F., and colleagues (2017). "Optimism and Cause-Specific Mortality: A Prospective Cohort Study." American Journal of Epidemiology 185(1), 21-29, doi:10.1093/aje/kww182. Abstract only. The hazard ratio moves from 0.71 to 0.91 when behaviours, conditions and depression enter one model together, which is why that figure cannot be read as a behaviour estimate.
- Wootton, R. E., Lawn, R. B., Millard, L. A. C., and colleagues (2018). "Evaluation of the causal effects between subjective wellbeing and cardiometabolic health: mendelian randomisation study." BMJ 362, k3788, doi:10.1136/bmj.k3788. Structured abstract and results paragraph only. Included here as the design that sidesteps the adjustment question.
- Schisterman and colleagues (2009), Epidemiology, on overadjustment bias. Recorded from a published reference list as the standard reference for the term. Not read by this publication.