Why Adjusting for Balanced Covariates Diverges Conditional and Marginal Odds Ratios

In a standard linear regression, adding a balanced variable—like a patient's baseline age in a perfectly randomized trial—leaves the estimated treatment effect unchanged while shrinking the margin of error. But when researchers measure outcomes using odds ratios through a logistic model, that exact same balanced variable forces the treatment estimate to change.

This divergence, known as noncollapsibility, occurs because the nonlinear mathematics of the logit link mean that the average of individual odds does not equal the odds of the average population. It is a mathematical artifact that fundamentally alters what a clinical trial is measuring.

For decades, statisticians have relied on a conventional wisdom established in linear models, where adjusting for prognostic covariates improves precision without biasing the treatment effect. This principle is so deeply embedded in trial design that regulatory agencies routinely recommend it to boost statistical power.[4][5]

Yet, when the outcome is binary and measured via odds ratios, the rules of linear algebra break down. The adjusted and unadjusted models are no longer estimating the same parameter, forcing researchers to choose between a marginal effect that reflects the population and a conditional effect that reflects the individual.[6]

The Mathematics of Noncollapsibility

The root of the divergence lies in the logit link function used in logistic regression. Unlike a risk difference, which scales linearly, an odds ratio operates on a multiplicative scale that compresses probabilities as they approach zero or one.

When a prognostic covariate is omitted from a linear model, the unmeasured variation simply gets absorbed into the residual error term. The treatment effect estimate remains unbiased, though the confidence intervals become wider due to the increased residual noise.[2]

In logistic regression, however, there is no separate residual error term, because the variance is entirely determined by the mean probability. If a prognostic covariate is omitted, the model averages the probabilities across the unmeasured subgroups before converting them to odds.[8]

Because the relationship between probability and odds is nonlinear, the odds of the average probability are always closer to 1.0 than the average of the individual odds. This mathematical certainty means the unadjusted marginal odds ratio is systematically attenuated toward the null compared to the adjusted conditional odds ratio.[6]

In a landmark 1984 paper in Biometrika, researchers demonstrated that certain nonlinear regression models lead to biased estimates of treatment effect if needed covariates are omitted, even in perfectly randomized experiments. The bias they described was not confounding, but the inherent divergence between marginal and conditional parameters.[1]

Conditional Versus Marginal Effects

The choice between an adjusted and unadjusted logistic regression is not merely a choice of statistical efficiency; it is a choice of estimand. The unadjusted model produces a marginal effect, which describes the average treatment effect across the entire trial population.[3]

The adjusted model produces a conditional effect, which describes the treatment effect for a specific individual, holding their baseline covariates constant. In linear models, the marginal and conditional effects are identical, but in logistic models, they are mathematically distinct.[6]

A 2021 analysis in the Biometrical Journal highlighted this distinction, noting that unadjusted comparisons deliver marginal measures of treatment effect but tend to be inefficient. Conversely, adjusted analyses deliver conditional measures whose magnitude depends entirely on precisely which covariates are included in the model.[6]

This creates a paradox for trialists. Adding more prognostic covariates to a logistic regression pushes the conditional odds ratio further from the null, making the treatment appear more effective for the individual, even though the population-level impact remains unchanged.[8]

Consequently, comparing odds ratios across different studies becomes impossible if the studies adjusted for different sets of covariates. An odds ratio adjusted for age and sex cannot be directly compared to an odds ratio adjusted only for age, because they are estimating fundamentally different parameters.[6]

The Precision Paradox in Logistic Models

The second major divergence involves statistical precision. In linear regression, adjusting for a highly prognostic covariate reduces the residual variance, which in turn reduces the standard error of the treatment effect estimate, providing the primary motivation for covariate adjustment.[2]

Logistic regression behaves entirely differently. A 1991 paper in the International Statistical Review proved that adjusting for a non-confounding predictive covariate in a logistic model actually results in a loss of precision for the coefficient estimator, increasing its variance.[2]

In classic linear regression, the adjustment for a non-confounding predictive covariate results in improved precision, whereas such adjustment in logistic regression results in a loss of precision. This finding directly contradicts the intuition most researchers carry over from linear models.[2]

However, this loss of precision in the coefficient does not necessarily translate to a loss of statistical power. Because the point estimate of the conditional odds ratio moves further from the null at a faster rate than the standard error increases, the resulting hypothesis test is often more powerful.[2]

Thus, researchers adjusting for covariates in logistic regression find themselves in a counterintuitive position: their point estimate is larger, their standard error is wider, yet their p-value is smaller. The mechanics of the power gain are entirely different from the variance reduction seen in linear models.[2]

Regulatory Stances on Covariate Adjustment

Despite the mathematical complexities of noncollapsibility, major regulatory bodies strongly encourage covariate adjustment in randomized trials to maximize statistical power. The U.S. Food and Drug Administration finalized its guidance on the topic in 2024.[4]

Covariate adjustment leads to efficiency gains when the covariates are prognostic for the outcome of interest in the trial, the FDA guidance states. Therefore, the agency recommends that sponsors adjust for covariates that are anticipated to be most strongly associated with the outcome of interest.[4]

The European Medicines Agency issued similar guidelines in 2015, recommending adjustment for baseline covariates to improve the precision of treatment effect estimates. Both agencies recognize that adjusting for highly prognostic variables like baseline disease severity can significantly reduce the required sample size.[5]

However, the regulatory guidance primarily focuses on the efficiency gains for hypothesis testing, often glossing over the fact that the estimand itself changes when logistic regression is used. This leaves trial statisticians to navigate the tension between maximizing power and maintaining a clear population-level interpretation.[7]

To resolve this, some statisticians advocate for standardization techniques. By fitting a conditional logistic model to maximize power, and then marginalizing the predicted probabilities over the observed covariate distribution, researchers can recover the marginal risk difference or risk ratio while retaining the efficiency gains.[3][7]

Quantifying the Divergence in Practice

The magnitude of the divergence between marginal and conditional odds ratios depends entirely on how strongly the adjusted covariates predict the outcome. If the covariates are weak predictors, the noncollapsibility effect is negligible, and the two estimates will be nearly identical.[8]

But when covariates are highly prognostic, the divergence can be substantial. In a re-analysis of the GUSTO-I trial for acute myocardial infarction, adjusting for age shifted the treatment coefficient from negative 0.159 to negative 0.188, an 18 percent increase in the estimated effect.[1][8]

A 2021 study in BMC Medical Research Methodology quantified this bias, demonstrating that noncollapsibility plays a major role in the discrepancy between adjusted and unadjusted estimates. The researchers found that the larger the effect of the covariate, the greater the attenuation of the marginal estimate.[8]

This attenuation is not a flaw in the study design or a result of confounding; it is a strict mathematical property of the logit link. Even in a trial with an infinite sample size and perfect randomization, the marginal odds ratio will always be closer to 1.0 than the conditional odds ratio.[6]

For clinicians interpreting trial results, this distinction is critical. A conditional odds ratio tells a doctor how a treatment will affect a specific patient with a specific set of baseline characteristics, while a marginal odds ratio tells a health system how the treatment will affect the population as a whole.[9]

Practical Implications for Trial Design

As clinical trials increasingly rely on complex covariate adjustments, understanding noncollapsibility becomes essential. The more prognostic information a model contains, the more the conditional estimand drifts from the marginal one, complicating cross-trial comparisons.[6]

Trialists must pre-specify exactly which estimand they are targeting. A 2022 guide published in Trials outlined methods for planning covariate adjustment, noting that if the goal is to estimate a population-level effect, researchers must use methods that explicitly target the marginal parameter.[7]

A 1999 paper in Statistical Science explored confounding and collapsibility, emphasizing that researchers can use targeted maximum likelihood estimation or other g-computation methods to adjust for covariates during the analysis phase while still outputting a marginal risk difference.[3]

Ultimately, the assumption that linear regression rules apply universally to all statistical models is a dangerous oversimplification. Recognizing the unique behavior of the logit link ensures that researchers do not mistake a mathematical artifact for a biological breakthrough.[2][9]