How Specification Changes Alter the Target Association

Why the Same Coefficient Can Mean Something Different

A regression coefficient is not a description of an association in isolation. Its meaning depends on the model terms around it and on the comparison it implies. A coefficient of 2, for example, might describe a constant change in the fitted outcome per unit of a predictor, a slope only for a reference group, or a conditional slope holding another variable fixed. When a specification changes, a useful question is not only “Did the coefficient move?” but also “What comparison does each model now describe?”

The clearest way to answer that question is to compare matched predictions: predictions for two predictor settings that differ in the feature of interest but agree on other settings. Their difference states the association described by the model in context. This approach works whether the change involves adding curvature, an interaction, or an adjustment term.

Write \(\hat m(\mathbf{x})\) for a model’s fitted mean outcome at predictor settings \(\mathbf{x}\). For two settings that differ in predictor \(x\), a matched contrast is \( \hat m(x_1,\mathbf{z})-\hat m(x_0,\mathbf{z}) \), where \(\mathbf{z}\) denotes the other settings held fixed. It is measured in outcome units. A coefficient may equal this contrast, a contrast divided by a predictor change, or a slope at a particular setting; it need not represent one constant comparison in every model.

Curvature Changes a Constant Slope into a Setting-Specific One

In a linear mean model, the fitted mean changes at a constant rate as \(x\) changes, provided the other terms are held fixed. Adding a quadratic term allows that rate to vary with \(x\). Consider the two models

In the first model, \(\hat\beta_1\) is the predicted outcome change per one-unit increase in \(x\). In the quadratic model, it is the slope at \(x=0\), not a constant slope over the full range. At a general value \(x\), the slope is \(\hat\beta_1+2\hat\beta_2x\). For two values \(x_0\) and \(x_1\), the fitted change is

The change per unit between those two values is therefore \(\hat\beta_1+\hat\beta_2(x_1+x_0)\), which depends on the comparison’s endpoints. The derivative gives the local slope; a finite matched contrast gives the predicted change over a specific interval. These are related, but they are not interchangeable.

The coefficient \(\hat\beta_1\) in a quadratic model is also sensitive to where \(x=0\) is placed. If zero is far outside the data range, its “slope at zero” may be of little practical interest. Centering \(x\) at a meaningful value changes the interpretation of the linear coefficient, although it does not change the fitted curve when the model includes both the linear and quadratic terms.

Interactions Make Slopes Conditional on Another Predictor

An interaction allows the association between \(x\) and the outcome to vary with another predictor. With a binary group indicator \(G\), where \(G=0\) is the reference group and \(G=1\) is the comparison group, a model with an interaction is

For \(G=0\), the fitted slope in \(x\) is \(\hat\beta_1\). For \(G=1\), it is \(\hat\beta_1+\hat\beta_3\). The coefficient \(\hat\beta_3\) is the difference between those slopes. Meanwhile, \(\hat\beta_2\) is the group difference at \(x=0\); it is not necessarily the group difference at other values of \(x\).

With an interaction, a group comparison also depends on the chosen value of \(x\). In this example, the fitted group difference at dose \(x\) is \(8+2x\). It is 16 points at \(x=4\) and 20 points at \(x=6\). Naming the comparison value is therefore part of stating what the model says.

Adjustment Terms Define What Is Held Fixed

Adding an adjustment predictor changes the comparison represented by a coefficient. In a model such as \( \hat m(x,z)=\hat\beta_0+\hat\beta_1x+\hat\beta_2z \), the coefficient \(\hat\beta_1\) describes the fitted change per unit of \(x\) when \(z\) is held fixed. A model without \(z\) does not make that same conditional comparison. Its coefficient summarizes the fitted association between \(x\) and the outcome without explicitly holding \(z\) fixed.

The two coefficients can differ because the models use different information to describe the association. Such a change does not, by itself, show that either model is correct or that the association is causal. Interpretation also depends on the question, the predictor values that are meaningfully comparable, and whether the model’s mean structure is suitable for the data.

A Practical Rule for Comparing Specifications

When a specification changes, first state the target comparison in words, then calculate it from each model’s predictions using the same settings wherever possible. This makes the comparison explicit even when individual coefficients no longer align.

For comparisons across fitted models, use the same outcome definition, observations, predictor units, and predictor settings when possible. If the models use different samples or units, a coefficient difference may reflect those changes as well as the changed terms. Matched predictions do not make two models equivalent; they make the comparison being drawn easier to see.

Common Mistake: Treating a Coefficient as a Universal Effect

• Wrong reasoning: “The coefficient for \(x\) is 2, so increasing \(x\) by one unit always increases the outcome by 2.”
• Correction: Check the model. With curvature, the slope depends on \(x\). With an interaction, it depends on the interacting predictor. With adjustment terms, it is conditional on the stated values of those predictors.
• Wrong reasoning: “The coefficient changed after adjustment, so the adjusted model proves the true effect.”
• Correction: A coefficient change shows that the fitted association depends on the specification. It does not alone identify which model is appropriate or establish causation.