Describing Adjusted Associations in Plain Language

From a Coefficient to a Clear Comparison

A regression coefficient is a number, but a useful interpretation is a sentence about a comparison. To describe an adjusted association clearly, name the outcome, say how the focal predictor differs, and state which other predictors are held constant. This makes clear what the coefficient compares—and what it does not establish.

The goal is not to list every feature of a regression model. It is to give a concise statement that preserves the meaning of the estimate, includes units, and does not imply that changing a predictor would necessarily change the outcome.

Definition: An adjusted association describes the model-based relationship between an outcome and a focal predictor while other specified predictors are held fixed. For a numerical focal predictor, it describes the fitted outcome difference for a stated increase in that predictor at the same values of the other predictors. For a group predictor, it describes the fitted outcome difference between the named groups at the same values of the other predictors. The adjustment set is the collection of predictors held fixed.

The Model Behind the Wording

For a numerical outcome \(Y\), a focal numerical predictor \(X\), and other predictors \(Z_1,\ldots,Z_k\), consider an additive linear mean model:

E(Y\mid X,Z_1,\ldots,Z_k) = \beta_0+\beta_1X+\beta_2Z_1+\cdots+\beta_{k+1}Z_k.

Here \(E(Y\mid X,Z_1,\ldots,Z_k)\) is the modelled mean outcome at the stated predictor values. The coefficient \(\beta_1\) describes the change in that mean for a one-unit increase in \(X\), with the \(Z\) predictors held fixed. A fitted regression reports an estimate of this coefficient; write that estimate as \(b_1\). The model’s additive form means the fitted change per unit of \(X\) is the same at each fixed combination of the other predictors.

This interpretation concerns the modelled association. A fitted model by itself does not show that changing \(X\) would cause a change in \(Y\). Nor does “held fixed” mean that the data necessarily contain two otherwise identical individuals. It names the comparison described by the model. Such a comparison is most informative for predictor values represented in the data; it should not be presented as well supported far beyond the observed range.

Plain-language template: “For a [specified increase] in [focal predictor], the fitted mean of [outcome] is [estimated amount] [higher/lower], comparing cases with the same [adjustment predictors].” For a group predictor, name both groups and say which group has the higher or lower fitted mean. Include the outcome’s units, and do not use causal wording unless a separate study design and assumptions justify it.

Why the Other Predictors Disappear From the Difference

The reason the adjustment predictors are named as fixed is visible by comparing two fitted means. For values \(x\) and \(x+\Delta x\), keep every \(Z\) value unchanged. Subtracting the fitted mean at \(x\) from the fitted mean at \(x+\Delta x\) gives:

\begin{aligned} &\left[\beta_0+\beta_1(x+\Delta x)+\beta_2Z_1+\cdots+\beta_{k+1}Z_k\right]\\ &\quad-\left[\beta_0+\beta_1x+\beta_2Z_1+\cdots+\beta_{k+1}Z_k\right]\\ &=\beta_1\Delta x. \end{aligned}

The intercept and the terms for the fixed predictors cancel. For a one-unit increase, \(\Delta x=1\), so the fitted difference is \(\beta_1\). For an increase of \(d\) units, it is \(d\beta_1\). This is why a good description states both the size of the predictor change and what was held fixed. In an application, the estimated fitted difference uses \(b_1\) in place of \(\beta_1\).

Worked Example: Study Time and Quiz Score

Suppose a hypothetical additive regression estimates quiz score in points from weekly study time in hours, baseline stress score, and course load. The estimated coefficient for study time is \(b_1=0.42\) points per hour. Describe the adjusted association for a one-hour increase in weekly study time.

The outcome is quiz score, measured in points. The focal predictor is weekly study time, measured in hours. The adjustment predictors are baseline stress score and course load. For a one-hour increase, the estimated fitted difference is:

0.42\ \frac{\text{points}}{\text{hour}}\times 1\ \text{hour} =0.42\ \text{points}.

A suitable description is: At the same baseline stress score and course load, students with one additional hour of weekly study time have a fitted mean quiz score 0.42 points higher. This states the outcome, the predictor change, the comparison, and the adjustment set. It describes an adjusted association in this hypothetical model; it does not establish that adding study time would cause a student’s score to rise.

Describing a Group Comparison

A focal predictor can instead represent groups. In that case, the sentence should name the two groups and the direction of the difference. The coefficient’s sign and units are not enough on their own: readers need to know which group is being compared with which. The comparison also needs to name the other predictors held constant.

For example, suppose a fitted model reports a coefficient for “program” of \(3.2\) points, where the reported contrast is program participants minus usual-care participants. The coefficient represents a fitted mean difference of 3.2 points between those groups at the same stated adjustment values. It does not mean that every participant’s score is 3.2 points higher, nor does adjustment alone prove that the program caused the difference.

Worked Example: Program Group and Follow-Up Score

In a hypothetical evaluation, the outcome is a follow-up assessment score in points. A fitted model includes program group, baseline assessment score, and age. The reported coefficient for the contrast “program minus usual care” is \(b=3.2\) points. Describe this adjusted group comparison.

The comparison is program participants versus usual-care participants. The adjustment predictors are baseline assessment score and age. The contrast is already expressed as program minus usual care, so its positive sign indicates a higher fitted mean for the program group at the same values of those predictors:

\text{fitted mean difference} = 3.2\ \text{points} \quad (\text{program}-\text{usual care}).

A suitable description is: At the same baseline assessment score and age, the fitted mean follow-up score is 3.2 points higher for the program group than for the usual-care group. The statement is about fitted means, not a claim that each program participant scored more highly. Because the coefficient alone does not establish a causal effect, “the program increased scores by 3.2 points” would overstate what this model-based comparison shows.

Make Units and Predictor Changes Explicit

A coefficient’s meaning depends on the units in which its predictor was recorded. A coefficient per one unit may be numerically small or large for reasons that have nothing to do with the importance of the association. When a more natural comparison uses several units, multiply the coefficient by that number and state the change explicitly.

Worked Example: Medication Dose and Blood Pressure

Suppose a hypothetical model estimates systolic blood pressure in millimetres of mercury (mmHg) from medication dose in milligrams, age, and baseline systolic blood pressure. The estimated dose coefficient is \(b=-0.36\) mmHg per milligram. Describe the adjusted association for a 10 mg difference in dose.

The focal predictor change is 10 mg, so the estimated fitted difference is:

-0.36\ \frac{\text{mmHg}}{\text{mg}}\times 10\ \text{mg} = -3.6\ \text{mmHg}.

The negative sign means the fitted mean is lower at the higher dose, with age and baseline systolic blood pressure held fixed. A suitable description is: At the same age and baseline systolic blood pressure, a 10 mg higher dose is associated with a fitted mean systolic blood pressure 3.6 mmHg lower. This is an adjusted association from the hypothetical model, not evidence by itself that increasing the dose would lower blood pressure by that amount.

The units cancel correctly: the coefficient is measured in mmHg per mg, and multiplying by a 10 mg difference gives a fitted outcome difference in mmHg. Leaving out “per 10 mg” would make the interpretation ambiguous.

A Practical Checklist for the Sentence

Before reporting an adjusted coefficient, check that a reader can answer these questions from the sentence:

What is the outcome? Name it and include its measurement units.
What is the focal comparison? Give the predictor increase and its units, or name both groups.
Which direction and amount? State whether the fitted outcome is higher or lower, and by how much.
What is held constant? Name the adjustment predictors rather than saying only “adjusted.”
Is the claim supported? Describe a fitted association, not an unsupported individual-level or causal conclusion.

“Adjusted for several factors” is often too vague, especially when the specific factors are important to understanding the comparison. Name the variables when possible. Also avoid saying that the comparison is “for people who are otherwise identical”: the model fixes specified predictors, not every characteristic that might differ between cases.

Common Mistake: Turning an Association Into an Effect

A common error is to read an adjusted coefficient as the effect of deliberately changing the focal predictor. For example, “one extra hour of study raises quiz scores by 0.42 points” uses causal language. The coefficient supports a more limited statement: at the same stated adjustment values, the fitted mean score is 0.42 points higher for a one-hour difference in study time. Adjustment describes which predictors are held fixed in the model; it does not, by itself, establish that all relevant alternative explanations have been ruled out.

Another error is to report only the number: “the adjusted coefficient is 0.42.” Without the outcome, predictor, units, comparison, and adjustment set, readers cannot tell what the number means. Translate the estimate into the actual comparison before interpreting its importance.

Key takeaway: A clear adjusted-association statement names the outcome, the focal predictor change or group contrast, the direction and size of the fitted difference, and the predictors held constant. Keep the wording associational unless there is separate justification for a causal interpretation.

Check Your Understanding

For each item, practise stating the comparison precisely and avoiding claims that the coefficient alone cannot support.

A fitted model estimates weight in kilograms from daily activity in hours, age, and height. The activity coefficient is \(0.8\) kg per hour. Write an adjusted-association statement for a one-hour difference in activity.
A model reports a group contrast of “Group A minus Group B” equal to \(-2.5\) points, adjusting for baseline score. Which group has the higher fitted mean at the same baseline score, and by how many points?
A predictor coefficient is \(1.4\) outcome units per year. What fitted difference corresponds to a three-year increase when the other predictors are fixed?
Why is “the treatment caused the outcome to improve by 3 points” not justified solely by an adjusted coefficient of 3 points?
What is missing from the statement “the adjusted coefficient is \(-0.36\)” if it is meant to describe a dose association with blood pressure?