The Table 2 fallacy
Why the covariate rows are not effect estimates
The model's adjustment set was chosen to make the EXPOSURE estimate unconfounded. Nothing guarantees that same set is the right adjustment set for any other variable in the table. A covariate's coefficient may be confounded even when the exposure's is not, and it may be a direct effect where a total effect is what a reader would assume.
The table the fallacy is named after
| Variable | Adjusted odds ratio (95% CI) | P value |
|---|---|---|
| Exposure | 1.50 (1.10–2.10) | .01 |
| Covariate A | 0.90 (0.70–1.20) | .43 |
| Covariate B | 1.20 (0.90–1.60) | .18 |
The direct-versus-total problem
Even setting confounding aside, a covariate coefficient in a model that also contains the exposure is estimating something closer to a DIRECT effect, with the exposure held fixed. A reader looking at a risk factor's row almost always wants its TOTAL effect, which includes whatever it does through the exposure. Those are different numbers and the table does not say which one it is showing.
Which path a covariate row estimates
What to do instead
Present the exposure estimate as the study's answer, and say plainly that the remaining rows are adjustment terms rather than findings. If a covariate's effect is genuinely of interest, it needs its own model with its own adjustment set, chosen for that question.
Why Mallard warns about it
Mallard attaches this warning to Table 2 and to forest plots in every plan that produces one, because the fallacy is easy to commit, invisible in the output, and frequently repeated in published papers.
Key assumptions and design requirements
- The exposure of interest is identified explicitly, so a reader knows which row is the answer.
- Covariate rows are labelled as adjustment terms, not as results.
- Any covariate whose effect genuinely matters gets its own analysis rather than a row in this one.
- Effect modification is reported deliberately, not inferred from a covariate coefficient.
What to watch out for
The covariate rows of an adjusted model are not effect estimates for those covariates. They are adjusted for the wrong set: the model was built to make the EXPOSURE estimate unconfounded, and nothing guarantees the same adjustment set is right for anything else in the table.
Clinical example
A model estimates the effect of an early mobility protocol on length of stay, adjusting for age, frailty and admission source. The protocol row answers the study's question. The frailty row does not tell you the effect of frailty on length of stay: it is adjusted for a set chosen for the protocol, and it is conditional on the protocol as well.
- The table 2 fallacy: presenting and interpreting confounder and modifier coefficients (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kws412
- Principles of confounder selection (opens in a new tab) · Eur J Epidemiol · doi 10.1007/s10654-019-00494-6