Your Regression Table Is Not a List of Treatment Targets

Why This Matters

Open almost any observational paper and scroll to the adjusted regression model (frequently labelled as Table 2). You know the table. There is one row for the relationship the study was actually designed to investigate, perhaps smoking and periodontitis. Beneath it is a list of other variables that the model adjusted for: age, sex, diabetes, toothbrushing frequency, and income. Each comes with its own estimate and confidence interval. Every row looks equally important. Same font. Same columns. Same two decimal places. By the Discussion section, the authors may be describing irregular dental attendance as an “independent risk” and suggesting that changing attendance patterns could reduce disease.

That sentence might sound harmless. It is not.

This is known as the Table 2 fallacy, a term introduced by Westreich and Greenland in 2013. The issue is not that statistical adjustment was not done correctly. Adjustment is the reason the model was fitted in the first place. The issue is treating every adjusted coefficient as though it represents a separate causal finding. These variables were included in the model to perform a specific job: to make the estimate for the main exposure more trustworthy. They were not necessarily included so that their own effects on the said outcome could be estimated. We wrote this commentary because this is one of the most common inferential errors we see in observational research.

What Was Done

We created a simple worked example, starting with a directed acyclic graph, or DAG. A DAG is a diagram showing the causal relationships we assume exist between variables.

Our diagram contained three variables:

• an exposure, X
• an outcome, Y
• a confounder, Z, which causes both the exposure and the outcome

In the paper, we used rheumatoid arthritis and cardiovascular risk, with smoking as the confounder.

A version for oral health is easy to imagine:

• X: periodontitis
• Y: cardiovascular events
• Z: smoking

We simulated data for 5,000 people, allowing us to compare the regression results with the true effects (let’s pretend we know).

What We Showed

The coefficient for the main exposure did what it was supposed to do. The model estimated an effect of 0.51 when the true effect was 0.50. But the smoking coefficient told a different story. The adjusted model produced a smoking coefficient of 0.39. This was close to smoking’s direct effect of 0.40, but smoking’s total effect was actually 0.79 because part of its effect travelled through the main exposure (refer to the Supplementary if you want to see the details, or just Figure 1 for simplicity). By adjusting for the exposure, the model blocked that pathway. The smoking row therefore captured only part of smoking’s effect.

This is the key lesson: the same variable can be a confounder in one question and a mediator in another. Its role depends on what you are trying to estimate. One adjustment set also cannot answer every question in the table. It may be appropriate for estimating the effect of periodontitis, but not the effects of smoking, diabetes, income or dental attendance! Each variable has its own causal structure and may need a different model.

Interaction terms can also be misleading. For example, periodontitis may seem more strongly linked to an outcome in smokers. This is called effect modification. However, it does not prove that smoking causes the stronger effect or that quitting smoking would remove the extra risk.

Terms such as “independent association,” “potential risk factor,” and “predisposing factor” can still imply causation. Even when the results are described as associations, causal claims may sneak back into the Discussion section through theorised biological mechanisms or clinical recommendations.

What This Means for Practice

Researchers should decide whether a study is descriptive, predictive, or causal before analysing the data before them. For studies with causal designs, focus on one clear exposure outcome question and interpret only the effect the model has been designed to estimate. As a reviewer or just a casual reader, when you read your next manuscript, treat these other coefficients as supporting information, not as proof of causes or clinical actions.

Publication

Tay JRH, Bashir NZ, Cooray U. Interpreting multivariable regression coefficients in observational clinical research: The Table 2 fallacy. Ann Acad Med Singap. July 2026. doi:10.47102/annals-acadmedsg.2026420.