The Table 2 Fallacy: Why Not Every Number in a Regression Is a Causal Effect

A multivariable regression is built to answer one causal question at a time, so only one number in the table, the coefficient for the primary exposure, is designed to carry a causal reading. The other adjusted coefficients beside it are there to clean up that one estimate, not to be read as effects in their own right. Treating each row as its own causal finding is the Table 2 fallacy, named for the table where these coefficients usually appear. A variable chosen to remove bias from your exposure can be the wrong thing to adjust for when you read its own coefficient.

This piece is educational and not medical advice; for decisions about your own care, please talk with your own clinician. I meet the problem from both directions. My doctoral research at the Lund University Diabetes Centre is on the genetics of type 2 diabetes, where a model is packed with correlated variables and the discipline is knowing which single coefficient the design was built to interpret.

What the fallacy looks like in practice

Picture a study of whether a treatment lowers the risk of an outcome. The authors fit a model with the treatment plus a familiar list of covariates: age, sex, body mass index, smoking, blood pressure, a marker of kidney function. The treatment coefficient is the point of the exercise, and the covariates were added to hold those factors constant so the estimate is not distorted by them. So far, so good.

The fallacy begins when a reader scans down the same column and reports, with equal confidence, that smoking raises the outcome by this much and higher blood pressure by that much, each read as a clean causal effect. That model was never tuned to answer those questions. The set of variables that correctly isolates the treatment effect is, in general, the wrong set for isolating the effect of smoking, or of blood pressure, or of any other covariate in the list. Each would need its own adjustment set, almost never the one printed in the table.

Why one adjustment set cannot serve every variable

A variable plays different roles depending on which effect you are trying to estimate. The same covariate can be a confounder for one exposure and a mediator for another, and those two roles call for opposite handling.

A confounder is a common cause of both the exposure and the outcome. You adjust for it to remove a spurious path, and that is exactly what good models do for the primary exposure. A mediator sits on the causal path between exposure and outcome, one of the steps through which the exposure works. Adjusting for a mediator does not clean up the estimate; it removes part of the very effect you are trying to measure, and can bias it in either direction.

Here is where the table quietly misleads. Suppose blood pressure is a confounder for the treatment, worth adjusting away. The same variable may sit on the causal path between another covariate, say kidney function, and the outcome. For the treatment's sake you correctly held blood pressure fixed, but in doing so you adjusted away part of kidney function's effect, so its coefficient no longer means what a naive reading assumes. One decision that was right for the exposure was wrong for a neighbor in the same table.

A third role makes things worse. Conditioning on a collider, a variable that is a common effect of two others, can open a spurious path that was not there before. A set chosen to close confounding paths for the exposure can open a collider path for a secondary variable, manufacturing an association out of the adjustment itself.

The unmeasured confounding you never budgeted for

Even setting mediators and colliders aside, there is a plainer reason to distrust the secondary rows. When you design a study around one exposure, you invest in measuring its confounders well, and you may have no comparable list for the covariates. A variable added mainly to sharpen the exposure estimate might have its own uncontrolled confounders that nobody measured, because controlling them was never the goal. Its coefficient then carries whatever residual bias those factors introduce, dressed in the same tidy format as the exposure estimate.

How to read a regression table defensively

You can protect yourself without redoing the analysis. A few habits do the work.

Find the one question the model was built to answer

Ask what single exposure the study set out to estimate. That is the coefficient the adjustment set was designed for, and usually the only one you should read causally. Methods sections and analysis plans typically name it. Everything else in the table is machinery.

Treat covariate coefficients as adjustments, not answers

When a covariate is associated with the outcome in an adjusted model, read it as a statistical association conditional on everything else in the equation, not as that covariate's causal effect. If a secondary coefficient surprises you or points the wrong way, that is often the signature of adjustment doing its job for a different variable, not a discovery about this one.

Ask whether the same set could possibly be right twice

For any covariate you are tempted to interpret, ask whether the printed adjustment set is plausibly correct for it too. Would you need to add a confounder the authors left out, or drop a mediator they kept? If the set was chosen for the exposure alone, the covariate row is not yours to interpret.

Why careful authors present it this way anyway

None of this means a full regression table is a mistake. Reporting all coefficients is standard and expected, and it lets readers see how the model was specified. The failure is not in printing the numbers; it is in reading them all as if each were the study's headline result. The clearest papers say plainly which estimate is the target and treat the rest as covariates, sometimes labeling secondary associations as exploratory rather than confirmatory. When every row is instead narrated as a causal effect, the table has outrun what the design can support.

The same instinct runs through the trials I have helped run. With EASY Diabetes, an AI clinical decision-support system for type 2 diabetes that I co-developed, our evaluation was EASY-1, a registered randomized controlled trial (NCT03258268), and randomization is what lets a single comparison carry a clean causal reading. In observational models you rebuild that discipline one coefficient at a time, claiming it for the variable you designed around, not for the whole column.