ExplainerCausal InferenceAmerican Journal of EpidemiologyOct 11, 2026, 4:31 PM· 6 min read· in Data & Analysis

How the Table 2 Fallacy Distorts Causal Inference in Observational Research

Interpreting every coefficient in a multivariable regression model as an independent risk factor introduces severe structural bias. Epidemiologists warn that mutual adjustment strips secondary variables of their causal meaning, requiring researchers to build separate models for each exposure.

By Mateo Ramos

Share this story

In short

The Table 2 Fallacy occurs when readers interpret every coefficient in a multivariable regression model as a valid, independent total causal effect.
Adjusting for covariates isolates a primary exposure, but leaves the secondary covariates exposed to unmeasured confounding and structural bias.
Epidemiologists advocate for a 'one exposure, one model' approach, using causal diagrams to build separate adjustment sets for each risk factor.

In February 2013, the pages of the American Journal of Epidemiology carried a brief but foundational commentary that would fundamentally alter how observational research is read. Epidemiologists Daniel Westreich and Sander Greenland identified a pervasive structural error hiding in plain sight across decades of medical literature.[1]

They coined the term "Table 2 Fallacy" to describe the routine misinterpretation of multivariable regression models. The fallacy occurs when researchers present a single statistical model and invite readers to interpret every resulting coefficient as a valid, independent causal effect.[1]

The name derives from the standard layout of medical manuscripts. Table 1 typically describes the demographic and clinical characteristics of the study population. Table 2 then presents the adjusted odds ratios or hazard ratios generated by the primary regression analysis.[2]

The error lies in assuming that because all the variables sit in the same column and were mutually adjusted, their coefficients answer the same type of causal question. In reality, the mathematical machinery of regression makes no distinction between a primary exposure and a secondary covariate.[1]

"Presentation of exposure and confounder effect estimates from a single model may lead to several interpretative difficulties," Westreich and Greenland wrote. They warned that this practice invites the dangerous confusion of direct effect estimates with total effect estimates.[1]

The Mechanics of Mutual Adjustment

To understand the fallacy, one must look at why covariates are added to a regression model in the first place. When researchers want to isolate the effect of a primary exposure, such as a new drug, they must adjust for confounding variables like age or smoking status.[1]

Adjusting for covariates isolates the primary exposure, but leaves the covariates themselves exposed to unmeasured confounding.

This adjustment set is carefully chosen to block spurious pathways between the primary exposure and the outcome. If the model is correctly specified, the resulting coefficient for the primary exposure represents a valid, conditional total causal effect.[1]

However, the secondary covariates included in that same model do not enjoy the same rigorous protection. Each of those adjustment variables possesses its own distinct causal structure, complete with its own set of unmeasured confounders that the model ignores.[1]

Because the model was optimized solely for the primary exposure, the coefficients attached to the covariates are hopelessly entangled. They represent associations that are conditional on the other variables, leaving them exposed to residual confounding from outside the model.[3]

Treating these secondary coefficients as independent risk factors is the essence of the Table 2 Fallacy. A model built to estimate the effect of one exposure frequently reports coefficients for age, sex, and comorbidities, which readers then wrongly interpret as actionable clinical findings.[1]

Mediators and the Loss of Total Effects

The interpretative chaos deepens when a regression model includes a mediator—a variable that lies on the causal pathway between the exposure and the outcome. Adjusting for a mediator fundamentally alters what the primary coefficient represents.[1]

When a mediator is absent from the model, the primary exposure's coefficient captures the total effect, encompassing both the direct impact and the indirect impact flowing through the mediator. This is usually the target estimand in public health research.[1]

Interpreting a secondary covariate as a total effect can entirely invert its directional sign, turning a true positive effect into a false negative.

But the moment a mediator is added to the adjustment set, that indirect pathway is mathematically blocked. The primary exposure's coefficient shrinks, silently transforming from a total causal effect into a strictly controlled direct effect.[1]

The regression output provides no warning that this swap has occurred. The coefficient simply drops, which researchers often misinterpret as evidence that the primary exposure is less dangerous than previously thought, rather than recognizing that a pathway has been closed.[1]

Meanwhile, the coefficient attached to the mediator itself absorbs a portion of that indirect effect, but remains hopelessly confounded by unmeasured variables. Neither number can be safely interpreted as a total causal effect, rendering the entire table misleading.[1]

Quantifying the Structural Bias

The magnitude of the bias induced by mutual adjustment is not trivial. Interpreting a secondary covariate as a total effect can completely distort the perceived relationship between a risk factor and a clinical outcome, leading to severe scientific errors.[3]

A 2024 simulation study published in medRxiv by McIntyre and colleagues quantified this exact distortion. They constructed a hypothetical causal mechanism to demonstrate how the Table 2 Fallacy alters coefficient estimates when multiple relationships are evaluated simultaneously.[3]

In their baseline model suffering from the Table 2 Fallacy, a secondary covariate yielded a coefficient of −0.18, with a p-value of 0.296. Under mutual adjustment, this variable appeared to have a statistically insignificant, negative association with the outcome.[3]

Including a mediator in the adjustment set blocks the indirect pathway, stripping the primary exposure of its total causal interpretation.

However, when the researchers built a separate, properly adjusted model dedicated specifically to that covariate, the true total effect was revealed. The coefficient inverted entirely, landing at +0.36 with a statistically significant p-value of 0.033.[3]

The Table 2 Fallacy had not merely suppressed the variable's significance; it had flipped the directional sign. A true positive causal effect of +0.36 was masked as a false negative association of −0.18 simply because it was read from the wrong model's output.[3]

High-Profile Failures and Real-World Stakes

These statistical inversions have profound consequences when observational data is used to guide public policy. The rush to generate actionable intelligence during the early months of the COVID-19 pandemic provided a stark demonstration of the fallacy's real-world impact.[2]

A highly publicized observational study explored the associations between various demographic characteristics and the risk of death from COVID-19. The researchers published a massive multivariable regression model, presenting dozens of adjusted hazard ratios in a single table.[2]

The French government subsequently used those secondary coefficients to define which specific groups of workers were deemed at high risk and required occupational protection. The policy was built directly on the assumption that every row in Table 2 represented an independent causal effect.[2]

A flurry of letters to the editor quickly followed, arguing that the paper wrongly implied causal interpretations for multiple parameters extracted from a single model. The study's authors contested the criticism, maintaining that their paper never explicitly made causal claims.[2]

Causal diagrams allow epidemiologists to map assumptions and identify the exact variables needed to unconfound a specific exposure.

Regardless of the authors' intent, the incident highlighted the danger of mutual adjustment. When researchers publish a menu of coefficients, policymakers and the media will inevitably interpret them as a list of independent, modifiable risk factors.[4]

The Path Forward: One Exposure, One Model

Eradicating the Table 2 Fallacy requires a fundamental shift in how observational research is designed and reported. Methodologists now insist that analysis must be driven by a strict "one exposure, one model" principle.[3]

If a study aims to estimate the causal effects of three different risk factors, investigators cannot simply run one massive regression. They must build three separate models, each featuring an adjustment set uniquely tailored to unconfound that specific exposure.[1]

This tailored approach relies heavily on Directed Acyclic Graphs (DAGs). These visual causal diagrams allow epidemiologists to map their structural assumptions a priori, identifying exactly which covariates must be adjusted for, and which must be strictly avoided.[3]

Medical journals and methodologists are increasingly enforcing these boundaries to prevent the misinterpretation of observational data. Editorial guidelines now frequently instruct authors to treat secondary coefficients strictly as adjustment terms, rather than findings in their own right.[4]

Medical journals and methodologists are increasingly enforcing these boundaries to prevent the misinterpretation of observational data.

The issue is not the statistical adjustment itself, but the deeply ingrained habit of reading every coefficient as if it were a causal effect. By relegating adjustment covariates to footnotes and presenting only the primary exposure in the main table, the field can finally close the door on the Table 2 Fallacy.[4]

How we did this

Method: Recomputing the magnitude and direction of bias induced by the Table 2 Fallacy by comparing a secondary covariate's coefficient under mutual adjustment against its true total effect under proper, DAG-informed adjustment.
What we found: Interpreting a secondary covariate as a total effect not only suppresses its statistical significance but can entirely invert its directional sign, shifting a true positive causal effect (+0.36) into a false negative association (−0.18) due to unblocked confounding paths.
What we worked from: Covariate coefficient under mutual adjustment (Table 2 Fallacy): −0.18 — medRxiv
Covariate coefficient under proper adjustment: +0.36 — medRxiv
Limits of this analysis: These specific coefficient shifts are drawn from a simulated causal mechanism; real-world observational data will exhibit different magnitudes of bias depending on the specific confounding structure.

Terms to know

Table 2 Fallacy: The erroneous practice of interpreting every coefficient in a multivariable regression model as a valid, independent total causal effect.
Confounder: A third variable that influences both the exposure and the outcome, creating a spurious association if not adjusted for.
Mediator: A variable that lies on the causal pathway between an exposure and an outcome, explaining how the effect occurs.
Total Effect: The overall causal impact of an exposure on an outcome, including all pathways.
Direct Effect: The causal impact of an exposure on an outcome that does not pass through a specific mediator.
Directed Acyclic Graph (DAG): A visual model used in epidemiology to map causal assumptions and identify the correct variables to adjust for.

Questions readers ask

Why is it called the Table 2 Fallacy?

In medical and epidemiological research papers, Table 1 typically describes the study population, while Table 2 presents the results of the primary multivariable regression model. The fallacy occurs when readers interpret all the adjusted coefficients listed in that second table as independent causal effects.

Can a single regression model estimate multiple causal effects?

Generally, no. A model designed to isolate the effect of one primary exposure uses a specific set of adjustment variables. Those adjustment variables have their own distinct causal structures and unmeasured confounders, meaning their coefficients cannot be read as valid total effects.

What happens if I adjust for a mediator?

Including a mediator in your regression model blocks the indirect pathway between the exposure and the outcome. This shrinks the primary exposure's coefficient, changing it from a total causal effect into a direct effect.

How do researchers avoid this error?

Epidemiologists recommend a 'one exposure, one model' approach. If a study aims to estimate the causal effects of three different risk factors, researchers must build three separate regression models, each with an adjustment set tailored to that specific variable.

Different angles

Causal Inference Methodologists

Statisticians and epidemiologists who advocate for strict structural modeling and the 'one exposure, one model' principle.

Methodologists argue that observational research must be grounded in explicit causal frameworks, typically using Directed Acyclic Graphs (DAGs). They maintain that throwing dozens of variables into a single regression model and interpreting the output as a menu of independent risk factors is mathematically incoherent. From this perspective, the Table 2 Fallacy is not just a minor reporting error, but a fundamental failure of study design that generates spurious associations and misguides public health policy.

Clinical Researchers

Applied medical researchers who often rely on mutual adjustment to explore multiple potential risk factors simultaneously.

Many clinical researchers view multivariable regression as a practical tool for hypothesis generation and risk prediction. In fast-paced research environments—such as the early months of the COVID-19 pandemic—investigators frequently publish models containing numerous demographic and clinical covariates to identify vulnerable populations. While acknowledging the theoretical risks of mutual adjustment, some argue that presenting all coefficients provides valuable transparency about the model's behavior, even if those secondary estimates should not be interpreted as strict causal effects.

Journal Editors

Academic gatekeepers pushing for clearer reporting standards to prevent the misinterpretation of observational data.

Editorial boards at major medical journals are increasingly cracking down on the Table 2 Fallacy. They recognize that policymakers and the media often extract secondary coefficients from published tables and report them as definitive findings. To combat this, editors now frequently require authors to clearly designate a single primary exposure, relegate adjustment covariates to footnotes, and explicitly state that secondary coefficients do not represent causal estimates.

Causal Inference Methodologists 60%Clinical Researchers 20%Editorial Standards Advocates 20%

Causal Inference Methodologists: Advocate for strict structural modeling and the 'one exposure, one model' principle.
Clinical Researchers: Rely on mutual adjustment to explore multiple potential risk factors and predict outcomes.
Editorial Standards Advocates: Push for clearer reporting standards to prevent the misinterpretation of observational data.

Perspectives this story doesn't cover

Machine Learning Practitioners
Public Health Policymakers

Sources

[1]American Journal of EpidemiologyCausal Inference Methodologists
The table 2 fallacy: presenting and interpreting confounder and modifier coefficients
Read on American Journal of Epidemiology →
[2]WikipediaClinical Researchers
Table 2 fallacy
Read on Wikipedia →
[3]medRxivCausal Inference Methodologists
The Table 2 Fallacy is an interpretation error commonly encountered in medical research
Read on medRxiv →
[4]Factlen Editorial TeamEditorial Standards Advocates
Synthesis by Factlen editorial team
Read on Factlen Editorial Team →

Up next

Causal Inference

The d-Separation Rule: How Directed Acyclic Graphs Identify and Block All Sources of Causal Bias

By mapping the flow of information through forks, chains, and colliders, the d-separation criterion provides a mathematical guarantee for isolating true cause and effect. The rule allows researchers to determine exactly which variables must be controlled for—and which must be ignored—to eliminate bias in observational data.