EpidemiologyMatters Library · Articles
Home Library Articles Causal inference
Causal inference · Issue 01

The Table 2 fallacy, or why the second row of every regression almost never means what we want it to mean

A single multivariable regression looks like it gives you many causal effects for the price of one. It does not. Here is the bias, drawn out — and a way to read tables that makes it visible.

You run a regression. You list your exposure of interest, you list your covariates, you fit the model, and out comes a tidy table with a coefficient for every variable. The exposure has its row. Age has a row. Sex, smoking status, comorbidities — each gets a row, each gets a number, each gets a confidence interval and a star or two.

The temptation, almost universal, is to read the whole column as if every row were the causal effect of that variable on the outcome. The exposure was adjusted for the covariates, so why would the covariates not be adjusted for the exposure? Why would they be different in kind?

This is the Table 2 fallacy, named by Westreich and Greenland in 2013 for the place it most often hides1: in the second numbered table of a paper, where the “fully adjusted” model lives. The argument is short, and once you have seen it you will see it everywhere.

The setup

Suppose we want to estimate the effect of HIV infection on the risk of stroke, using observational cohort data. We know — or assume — that smoking is associated with both (it makes HIV-positive people more likely to be smokers, and it raises stroke risk independently). We adjust for smoking. We also adjust for age, because older participants are both more likely to have been HIV-positive long enough for sequelae and more likely to stroke.

The DAG below is the canonical Westreich–Greenland example. HIV is the exposure. Stroke is the outcome. Smoking sits between HIV (which, plausibly, raises smoking via the social patterning of HIV in the cohort) and stroke. Age is a confounder of the HIV → stroke effect, and also of smoking → stroke.

Figure 1 · interactive
Exposure Outcome Adjusted (conditioned on) Unmeasured
Click a covariate to toggle whether it is adjusted for. Watch how the open and blocked paths between HIV and stroke change. Smoking is a mediator of the HIV → stroke effect; adjusting for it removes part of the very effect we want to measure.
Try this → Toggle smoking off (unadjusted), then back on (adjusted). Notice that adjusting closes the path through smoking — but that path was part of the total effect of HIV on stroke. The “adjusted” HIV coefficient now estimates only the direct effect of HIV on stroke, not the total. That is fine if a direct effect is what you wanted. It is a problem if you write up the number and call it “the effect of HIV on stroke.”

Two coefficients, two estimands

Now look at the smoking row of the same regression. The smoking coefficient is the effect of smoking on stroke, conditional on HIV and age. But the right adjustment set for the smoking → stroke effect is not the same as the right adjustment set for the HIV → stroke effect. For smoking, HIV is a cause of the exposure (smoking), not a confounder; conditioning on HIV is unnecessary at best, and — if HIV affects stroke through unmeasured pathways shared with smoking — actively introduces collider bias2.

So the same model, in the same column of the same table, is producing one number that is (let’s say) a reasonable estimate of one effect, and another number that is an estimate of nothing in particular. The reader has no way to tell them apart.

Figure 2 · interactive
True HIV → stroke effect (βHIV)0.40
HIV → smoking pathway strength0.30
Unmeasured U (HIV–stroke shared cause)0.00
Variable β̂ 95% CI Reads as…
Adjust the sliders and watch the smoking row drift. The HIV row is reasonably stable when unmeasured U is zero; the smoking and age rows are not what they look like, even when the model is correct for HIV.
Try this → Push Unmeasured U up. Now even the HIV row is biased — because we omitted a confounder. None of the rows are guaranteed to be unbiased simultaneously. That is the point.

What the textbook gets right, and what it can’t

Most epidemiology textbooks teach you to draw a DAG, identify confounders for the exposure–outcome relationship, and adjust for them. We do this in Epidemiology Matters. What is harder to convey on a page — and what the Table 2 fallacy turns on — is that the right adjustment set is a property of an exposure–outcome pair, not of the data. A regression conditions on every covariate at once. Each coefficient inherits an adjustment set whether or not that set was chosen for it.

The standard fix, when authors think about it at all, is a footnote: “Coefficients for covariates should not be interpreted causally.” Westreich and Greenland call this the table’s fig leaf. Readers ignore it. Reviewers ignore it. The number is on the page, and the number gets cited.

Walk through the bias

The scrollytelling figure below builds the bias one step at a time. Scroll the panel on the right; the DAG on the left updates to show the path being created or blocked at each step.

Step 1 · The question

We want the total effect of HIV on stroke. That is one well-defined quantity: how much would population stroke risk change if HIV prevalence changed?

Step 2 · Identify confounders

Age confounds the HIV → stroke relationship. We adjust. Good.

Step 3 · Add a mediator by mistake

Smoking lies on a path from HIV to stroke. Adjusting for it removes part of the HIV effect we wanted to estimate. The HIV coefficient is no longer the total effect — it is a controlled direct effect.

Step 4 · Read the smoking row

The smoking coefficient is the effect of smoking on stroke, conditional on HIV and age. But HIV is upstream of smoking. Adjusting for it doesn’t help; it can hurt.

Step 5 · The fallacy

Reporting both rows as if they were causal effects is the Table 2 fallacy. They are answers to different questions, fit by the same model, presented as if they were equivalent.

Step 6 · The fix

State your estimand. Pick the adjustment set for that estimand. If you want effects for several exposures, run several models — or fit one model and report only the row you can defend.

What to do instead

The remedy is unglamorous and well known: be explicit about the estimand, draw the DAG, and pick the minimum sufficient adjustment set for the specific exposure–outcome pair you care about. If you want effects for multiple exposures, fit multiple models — one per estimand — each with its own adjustment set. Or fit one model and report only the row whose interpretation you can actually defend, with the rest as nuisance parameters explicitly labeled as such3.

This is more work than printing one fully-adjusted table. It also turns out, in practice, to clarify what the analysis is for. Many studies that present six causal-looking coefficients are really designed for one. The fallacy hides that fact from the reader and, often, from the author.

Citation

Cite this article Abba-Aji M. The Table 2 fallacy: why secondary regression coefficients almost never mean what we want them to mean. Epidemiology Matters Journal. 2026;1(1). https://epidemiologymatters.com/journal/table-2-fallacy/

References & footnotes

  1. Westreich D, Greenland S. The Table 2 fallacy: presenting and interpreting confounder and modifier coefficients. American Journal of Epidemiology. 2013;177(4):292–298. academic.oup.com/aje/article/177/4/292
  2. Greenland S, Pearl J, Robins JM. Causal diagrams for epidemiologic research. Epidemiology. 1999;10(1):37–48.
  3. Bandoli G, Palmsten K, Forbess Smith CJ, Chambers CD. A review of systemic corticosteroid use in pregnancy and the risk of select pregnancy and birth outcomes. Rheumatic Disease Clinics. 2017;43(3):489–502. pmc.ncbi.nlm.nih.gov/PMC6103824
  4. Textor J, van der Zander B, Gilthorpe MS, Liśkiewicz M, Ellison GTH. Robust causal inference using directed acyclic graphs: the R package dagitty. International Journal of Epidemiology. 2016;45(6):1887–1894. dagitty.net
  5. Hernán MA, Robins JM. Causal Inference: What If. Boca Raton: Chapman & Hall/CRC; 2020.

Exercises

Three short questions to work through before your next regression. Click to reveal a model answer.

Hypertension is plausibly a mediator of the activity → mortality effect (activity lowers blood pressure, which lowers mortality). Adjusting for it removes part of the effect of activity, so the activity coefficient is a controlled direct effect, not a total effect. The hypertension coefficient, meanwhile, is conditioned on activity — which is upstream of hypertension — and so its adjustment set is wrong for a hypertension → mortality estimand. Both rows look comparable in the table; neither answers the question the abstract claims to answer.

(1) State the primary estimand and report only the coefficient that estimates it; label the rest explicitly as nuisance parameters not to be interpreted causally. (2) If the authors want effects for several exposures, ask them to draw a DAG, identify the adjustment set for each exposure–outcome pair, and fit a separate model per estimand.

Essentially never, in practice. The minimum sufficient adjustment set would have to be identical for every variable in the model, which requires a very particular DAG (for example, all covariates being independent direct causes of the outcome with no shared causes among them). When this holds it should be argued for explicitly. Westreich and Greenland’s point is that authors almost never check, and readers almost never demand it.