Reporting an Adjusted Association With Appropriate Limits

Why Reporting Needs More Than a Coefficient

A regression coefficient can be numerically correct and still be easy to misread. A statement such as “the adjusted coefficient was −1.8” leaves out the outcome, the exposure scale, the covariates, and the uncertainty. It may also invite a causal interpretation that the analysis does not support. A useful results statement makes the comparison understandable and identifies its limits.

This tutorial focuses on how to report an adjusted association from a multiple linear regression. The goal is not to repeat a covariate-selection analysis, but to say clearly what was estimated, what variables were adjusted for, how uncertain the estimate is, and what conclusions should not be drawn from it.

Definition: An adjusted association is the relationship between an exposure and an outcome estimated by a model that also includes specified covariates. In a multiple linear regression, the exposure coefficient describes the modeled difference in mean outcome associated with a one-unit exposure difference, with the included covariates held fixed. It is a model-based conditional association; by itself, it is not a causal effect.

Model and Assumptions Behind the Statement

Let \(Y\) be a quantitative outcome, \(X\) the exposure of interest, and \(Z_1,\ldots,Z_k\) the covariates in the adjustment set. A multiple linear regression for the conditional mean can be written as

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

Here, \(\beta_1\) is the population model coefficient for the exposure. Its estimate, \(\hat{\beta}_1\), is reported in the outcome units per one unit of \(X\), unless the exposure has been coded differently. For a categorical exposure, the coefficient is a modeled difference relative to the stated reference category. The covariates define the adjustment; they should be named in the report rather than hidden behind a phrase such as “adjusted analysis.”

The interpretation depends on the model being a reasonable description of the data. For conventional model-based confidence intervals in this setting, the analysis also relies on independent observations and an appropriate model for the residual variability; normally distributed residuals are typically assumed for exact small-sample t-based inference. The fitted mean should represent the relevant exposure and covariate relationships adequately. These assumptions concern the statistical model and its uncertainty. They do not, by themselves, make an observational comparison causal.

A confidence interval is a way to report uncertainty about the coefficient. A 95% confidence-interval procedure is designed so that, over repeated samples under its assumptions, 95% of intervals constructed by that procedure would contain the population coefficient. It is not correct to describe a particular computed interval as having a 95% probability of containing the fixed coefficient.

What a Complete Results Statement Contains

A concise report usually identifies the population or analysis group, the outcome and its units, the exposure contrast, the covariates included, and the estimate with an uncertainty interval. It should also say whether the estimate is an association or a causal effect. For an observational analysis, “adjusted association” is usually the appropriate description unless a separate causal design and its assumptions justify stronger language.

For example, a reusable structure is:

Reporting template: “Among [analysis group], each [exposure contrast] higher in [exposure] was associated with an estimated [increase/decrease] of [estimate] [outcome units] in [outcome] (95% CI [lower, upper]), after adjustment for [named covariates]. This is an adjusted association, not by itself evidence that changing the exposure would cause the outcome to change.”

Use the units that make the comparison interpretable. If a one-unit change is too small or unusual, report a meaningful contrast and scale both the estimate and its interval accordingly. Retain enough digits to represent the calculation, but avoid false precision; three or four significant figures are generally sufficient.

Worked Example: Reporting a Change of Two Exposure Units

Suppose a hypothetical observational analysis models systolic blood pressure in millimeters of mercury (mmHg) against weekly activity in hours, adjusting for age in years and baseline body mass index in kg/m². The model’s estimate for one additional activity hour per week is \(-1.80\) mmHg, with a reported 95% confidence interval from \(-2.70\) to \(-0.90\) mmHg. Assume observations are independent, the fitted relationships and residual checks are considered adequate, and the interval procedure is appropriate for this analysis.

A two-hour contrast is twice the one-hour contrast under the model’s linear exposure term. Thus the estimated difference is

2(-1.80)=-3.60\text{ mmHg}.

Scaling the interval endpoints by the same positive factor gives

[2(-2.70),\,2(-0.90)]=[-5.40,\,-1.80]\text{ mmHg}.

A suitable report is: “Among the adults in this hypothetical analysis, two additional hours of weekly activity were associated with an estimated 3.60 mmHg lower systolic blood pressure (95% CI: 5.40 mmHg lower to 1.80 mmHg lower), after adjustment for age and baseline body mass index.” This describes the fitted conditional association. Because the analysis is observational, it does not establish that increasing activity would cause blood pressure to fall.

Why the Coefficient Has an Adjusted Interpretation

The coefficient interpretation follows directly from the model’s mean equation. Hold the covariates fixed and compare two exposure values that differ by one unit. Subtract the conditional mean at \(X=x\) from the conditional mean at \(X=x+1\):

\begin{aligned} &E(Y\mid X=x+1,Z_1,\ldots,Z_k) - E(Y\mid X=x,Z_1,\ldots,Z_k)\\ &= \left[\beta_0+\beta_1(x+1)+\sum_{j=1}^{k}\gamma_jZ_j\right] - \left[\beta_0+\beta_1x+\sum_{j=1}^{k}\gamma_jZ_j\right]\\ &=\beta_1. \end{aligned}

The intercept and covariate terms cancel, leaving the exposure coefficient. Replacing \(\beta_1\) with its estimate gives the fitted adjusted difference. This justification explains both the “holding the included covariates fixed” language and why the coefficient’s units are outcome units per exposure unit. It does not show that the covariates are sufficient to remove confounding or that the model represents a causal comparison.

Worked Example: A Binary Exposure and a Continuous Outcome

In a hypothetical clinic analysis, the outcome is recovery time in days. The exposure is participation in a support program (yes versus no), and “no” is the reference category. The fitted coefficient for participation is \(-4.20\) days, with a 95% confidence interval from \(-7.00\) to \(-1.40\) days. The model adjusts for age, baseline symptom severity, and baseline health status. Suppose the two groups have adequate overlap in these covariates, observations are independent, and model diagnostics do not reveal a serious problem with the mean form or residual variability.

Because the exposure is binary, the coefficient represents the adjusted difference between participation and nonparticipation, with the named covariates held fixed. The negative estimate is a shorter fitted recovery time for participants. The interval endpoints are also in days, so the interval ranges from a difference of 7.00 days shorter to 1.40 days shorter.

A suitable report is: “In this hypothetical clinic analysis, program participation was associated with an estimated 4.20-day shorter recovery time than nonparticipation (95% CI: 7.00 days shorter to 1.40 days shorter), after adjustment for age, baseline symptom severity, and baseline health status.” If participation was not randomly assigned, this adjusted association alone does not show that the program caused faster recovery. Differences between groups not adequately represented by the model could still matter.

Reporting an Interval That Includes Zero

For a difference or slope, zero represents no difference or no linear association on the stated scale. If a confidence interval includes zero, the data and model are compatible with a range of values that includes no association, as well as values on either side of zero. Do not convert that fact into a claim that there is no association. The estimate still gives the fitted direction and size, while the interval shows that the estimate is imprecise enough to include multiple possibilities.

Worked Example: Describing an Uncertain Estimate

A hypothetical analysis relates a measured environmental exposure, in units of 10 micrograms per cubic meter, to a lung-function score in points. The estimated adjusted slope is \(0.12\) points per exposure unit, with a 95% confidence interval from \(-0.08\) to \(0.32\) points. The model adjusts for age, smoking status, and height. Assume the model’s functional form and uncertainty assumptions are considered adequate for this analysis.

The point estimate indicates an estimated increase of \(0.12\) score points for each 10-microgram-per-cubic-meter higher exposure. The interval includes zero because \(-0.08<0<0.32\). It also includes modest negative and positive slopes, so a report should not claim either a definite increase or proof of no association.

A suitable statement is: “Each 10-microgram-per-cubic-meter higher exposure was associated with an estimated 0.12-point higher lung-function score (95% CI: 0.08 points lower to 0.32 points higher), after adjustment for age, smoking status, and height. The interval includes no difference, so the estimate is compatible with both a small negative and a small positive adjusted association.” This is a description of the interval, not a conclusion that the null hypothesis is true.

Limits on Causal Language

Adjustment means that the fitted comparison accounts for the covariates included in the model, according to the model’s structure. It does not mean that every relevant difference between exposure groups has been controlled. A causal interpretation requires more than a coefficient and a confidence interval: the adjustment set must be suitable for the causal question, important common causes must be adequately measured and handled, and the analysis must avoid problems such as conditioning on inappropriate variables or selecting a distorted analysis sample. The earlier material on confounding, mediators, colliders, and selection bias develops these issues.

Be precise about the target of the statement. A regression coefficient generally summarizes a conditional association at fixed values of included covariates, under the specified model. It need not equal a marginal difference averaged over the whole population. A report should not suggest that the estimate applies to groups, exposure ranges, or settings that were not represented or supported by the analysis.

When evidence supports only an association, use terms such as “was associated with,” “the adjusted estimate was,” or “the model estimated.” Avoid “led to,” “reduced,” or “increased” if those phrases assert that the exposure caused the outcome. Naming the covariates makes the adjustment transparent; stating the observational limit keeps the interpretation proportionate to the design.

Common Mistakes

Reporting a number without its scale. “The estimate was −1.8” is incomplete. State the outcome units per exposure unit, or translate the result to a meaningful exposure contrast.
Calling the coefficient an effect without support. Adjustment for several covariates does not automatically establish causation. Describe an observational result as an adjusted association unless causal assumptions and design justify more.
Leaving the adjustment set unnamed. “After adjustment” is not enough for a reader to know what comparison the model made. List the covariates included.
Treating an interval that includes zero as proof of no relationship. The interval indicates which values remain compatible with the data and model; it does not prove the coefficient is zero.
Misstating confidence-interval meaning. A 95% confidence interval is not a statement that the fixed coefficient has a 95% probability of lying in this particular interval.
Using “adjusted” as a guarantee. Adjustment describes inclusion of variables in a model; it does not guarantee that confounding, model misspecification, or selection bias has been eliminated.

Key takeaway: A useful adjusted-results statement names the outcome, exposure contrast, adjustment covariates, estimate, units, and uncertainty interval. It then identifies the result as an adjusted association unless the study design and causal assumptions support a causal claim.

Check Your Understanding

Use the reporting principles in this tutorial to answer the following questions.

A model estimates a slope of \(-0.6\) hours of sleep per additional daily screen-time hour. What additional details are needed before this estimate can be reported clearly?
Why should the covariates in an adjusted association be named in the results statement?
A 95% confidence interval for a difference is \([-2.1, 0.4]\) outcome units. What does its inclusion of zero allow you to say, and what does it not allow you to conclude?
In a hypothetical observational study, a treatment coefficient is negative after adjustment for age and baseline severity. Why is “the treatment reduced the outcome” stronger than the regression result alone supports?
If a coefficient and its confidence interval describe the association per one exposure unit, what must be done before reporting them for a five-unit exposure contrast under a linear model?