Estimate Adjusted Exposure Associations With Regression

Why Adjust an Outcome Regression?

In an observational study, people with different exposure histories may also differ in characteristics related to the outcome. A regression of outcome on exposure alone can therefore mix the exposure association with differences in those characteristics. An outcome regression includes selected covariates alongside exposure so that the exposure coefficient describes a comparison at the same modeled covariate values.

That adjusted coefficient is a conditional association: it compares expected outcomes among individuals with the same included covariates. Adjustment does not, by itself, establish that the association is causal, nor does the coefficient automatically describe an average contrast for a target population. Both the model and the assumptions behind the causal question matter.

Definition: Let \(Y\) be a quantitative outcome, \(A\) a binary exposure coded 0 or 1, and \(C\) the set of measured covariates included in the model. An adjusted outcome regression models the conditional mean \(E(Y\mid A,C)\) using both \(A\) and \(C\). The exposure coefficient describes the modeled difference in mean outcome between \(A=1\) and \(A=0\), holding the included covariates fixed.

The Linear Outcome Model and Its Coefficient

For a continuous outcome, a commonly used model is an additive linear regression. For simplicity, first take \(C\) to be one quantitative covariate:

Y_i=\beta_0+\beta_A A_i+\beta_C C_i+\varepsilon_i, \qquad E(\varepsilon_i\mid A_i,C_i)=0.

Here, \(\beta_0\) is the modeled mean outcome when \(A=0\) and \(C=0\); \(\beta_C\) is the modeled mean change in outcome for a one-unit increase in \(C\), holding exposure fixed; and \(\beta_A\) is the modeled mean difference in outcome for \(A=1\) versus \(A=0\), holding \(C\) fixed. If \(C\) is categorical, it is represented by indicator variables, with each coefficient interpreted relative to its reference category. Additional covariates are handled in the same way.

The key interpretation follows by comparing the two modeled means at the same covariate value:

E(Y\mid A=1,C=c)-E(Y\mid A=0,C=c) =(\beta_0+\beta_A+\beta_Cc)-(\beta_0+\beta_Cc) =\beta_A.

Thus, in this model, the exposure coefficient is a common conditional mean difference: it is the same for every \(c\). This is a consequence of the model’s additive form, not a general feature of all relationships between exposure and outcome. It assumes that the exposure contrast does not vary with the included covariates. A model that imposes this common contrast may be a poor description if the data show otherwise.

Model assumptions: For the coefficient to represent the stated conditional mean comparison, the conditional mean must be adequately described by the chosen regression form, including its covariates and functional relationships. The zero conditional mean condition \(E(\varepsilon\mid A,C)=0\) is required for the model’s errors to have mean zero at each exposure and covariate profile. Usual regression standard errors additionally rely on assumptions such as independent observations and, for the conventional homoskedastic standard error, constant error variance. Normal errors support exact small-sample \(t\)-based inference. These statistical assumptions do not alone establish a causal interpretation.

From an Adjusted Association to a Causal Question

A regression coefficient is first a feature of a fitted statistical model. To interpret it as a causal effect, the study question and design must support that interpretation as well. For example, the adjusted covariates should be appropriate pre-exposure confounders rather than variables that introduce bias when conditioned on. The causal interpretation also requires consistency, conditional exchangeability given the measured covariates, and positivity in the target population. In plain terms, the observed outcome under a person’s actual exposure must correspond to the relevant potential outcome; there must be no unmeasured confounding after conditioning on the chosen covariates; and each relevant covariate profile must have a possibility of receiving either exposure.

Even with those identification assumptions, the outcome model must be suitably specified for the intended estimand. A regression coefficient conditional on covariates and an average contrast across a population answer different questions. The conditional coefficient does not average over the covariate distribution by itself. In the special case of a correctly specified additive linear model with a common exposure contrast, the conditional contrast is constant across covariate profiles, so averaging that contrast yields the same numerical contrast. That equality depends on the model and target; it should not be assumed automatically.

Worked Example: Crude and Adjusted Differences

Consider a hypothetical dataset about an exposure and a continuous health score, measured in points. A binary pre-exposure risk category \(C\) is a confounder: it is related to exposure status and to the outcome. The table gives the number of people and the mean score in each exposure-by-category group.

Risk category \(C\)	Exposure \(A\)	Number of people	Mean score (points)
0	0	80	10
0	1	20	14
1	0	20	20
1	1	80	24

The crude mean for the exposed group is

\bar{Y}_{A=1} =\frac{20(14)+80(24)}{20+80} =\frac{2200}{100} =22\text{ points}.

The crude mean for the unexposed group is

\bar{Y}_{A=0} =\frac{80(10)+20(20)}{80+20} =\frac{1200}{100} =12\text{ points}.

So the crude difference is \(22-12=10\) points. But within each risk category the mean difference is \(14-10=4\) points for category 0 and \(24-20=4\) points for category 1. In the additive regression with an indicator for \(C\), the cell means are represented by the fitted equation \(E(Y\mid A,C)=10+4A+10C\): it predicts 10, 14, 20, and 24 points for the four cells. Thus the adjusted exposure coefficient is 4 points, not the crude 10-point difference. The exposure groups have different mixes of risk categories, so the crude contrast also reflects that imbalance.

All four exposure-by-category combinations are represented, so there is observed overlap for these two categories in this hypothetical dataset. The calculation establishes the model-based adjusted association. A causal reading would still require that \(C\) and the other included variables are appropriate, that no relevant confounding remains, and that the model and causal assumptions hold. The group summaries alone cannot verify those assumptions.

Uncertainty and Contextual Interpretation

A fitted exposure coefficient is an estimate from a sample. A confidence interval describes uncertainty in that estimate under the regression and sampling assumptions; it does not account automatically for model misspecification, unmeasured confounding, or a mismatch between the study sample and target population. A \(p\)-value for \(H_0:\beta_A=0\) measures how unusual a statistic at least as extreme as the observed one would be if that model-based null hypothesis and the inferential assumptions held. It is not the probability that the null hypothesis is true.

Worked Example: Interpreting an Adjusted Coefficient and Interval

Suppose a hypothetical regression estimates the association between a binary exposure and a continuous symptom score measured in points. The model adjusts for age, baseline symptom score, and a pre-exposure risk category. A report gives \(\hat{\beta}_A=-2.40\) points, \(SE=0.90\) points, and \(df=196\). Assume independent observations, a suitable linear conditional mean, and constant error variance; suppose the normal-error condition is reasonable for the \(t\)-based interval.

The test statistic for \(H_0:\beta_A=0\) against a two-sided alternative is

t=\frac{\hat{\beta}_A-0}{SE} =\frac{-2.40}{0.90} =-2.67.

With 196 degrees of freedom, the two-sided \(p\)-value is approximately 0.0083. At \(\alpha=0.05\), this is evidence against a zero conditional association under the stated model and inferential assumptions.

Using \(t^*=1.972\) for a 95% confidence interval with 196 degrees of freedom gives

\hat{\beta}_A\pm t^*SE =-2.40\pm1.972(0.90) =-2.40\pm1.775 =(-4.175,-0.625)\text{ points}.

The estimated adjusted association is a 2.40-point lower mean symptom score for the exposed group compared with the unexposed group at the same modeled age, baseline score, and risk category. The interval is consistent with a conditional mean difference from about 4.18 points lower to 0.625 points lower. It is not a 95% probability statement about the fixed coefficient, and it does not by itself show that exposure caused the difference.

Conditional Does Not Mean Population-Average

The phrase “adjusted for age and baseline score” does not turn a regression coefficient into an average effect for all people in a population. The coefficient compares exposure levels while those covariates are held fixed in the model. A population-average contrast instead concerns outcomes averaged over a specified population’s covariate distribution. The two summaries may coincide under a suitable model, such as a correctly specified additive linear model with a common exposure contrast, but they need not coincide in general.

The distinction matters for communication. “At the same modeled covariate values, the fitted mean score is 2.4 points lower for the exposed group” describes a conditional association. “Exposure lowers the population’s average score by 2.4 points” is a population-level causal claim and requires additional justification. A single regression coefficient cannot supply that justification merely because covariates appear on the right-hand side of the equation.

Worked Example: State the Level of the Conclusion

A hypothetical study models annual hospital visits as a continuous outcome, in visits per person, using exposure status and age as predictors. The fitted exposure coefficient is 0.60 visits per person. Assume the conditional mean is adequately modeled as linear in age and additive in exposure, and assume the usual sampling conditions for the reported estimate are reasonable.

Conditional interpretation. At the same age, the fitted mean number of annual visits is 0.60 higher for exposed than unexposed people. The units are visits per person per year, and the statement concerns the model’s conditional mean.

What the coefficient alone does not say. It does not establish that the population-average number of visits would be 0.60 higher if everyone were exposed than if everyone were unexposed. Such a causal population comparison needs a target population and defensible causal assumptions. If the conditional contrast truly is constant at every relevant age and the model is correct for the target, averaging the same 0.60 difference across ages gives 0.60. If those conditions do not hold, the coefficient need not equal the population-average contrast.

Common Mistakes and a Practical Check

Before interpreting an adjusted exposure coefficient, check the model’s outcome scale, the exposure coding, the covariates and their timing, the form of the conditional mean, and the uncertainty calculation. In particular, a coefficient’s sign depends on which exposure group is coded 1, and its units are the outcome units—not a percentage or a causal effect unless the model and assumptions support those interpretations.

Mistake: “The regression adjusted for confounders, so this is the causal effect.” Adjustment only controls measured variables represented appropriately in the model. Causal interpretation also requires the identification assumptions, including no unmeasured confounding given the covariates.
Mistake: “The adjusted coefficient is the population-average effect.” The coefficient is conditional on the included covariates. A population average is a different target unless conditions make the conditional contrast constant and the model appropriate.
Mistake: “A small \(p\)-value proves the model is correct.” The test addresses a model-based null under its assumptions; it does not test whether confounding is absent or whether the functional form is suitable.
Mistake: “The coefficient is a change in outcome for each additional unit of exposure.” For binary exposure coded 0/1, it is a modeled difference between those two exposure levels, in outcome units, conditional on the covariates.

Key takeaway: An adjusted linear regression coefficient for a binary exposure is a conditional mean difference at fixed modeled covariate values. It is an association unless causal assumptions justify more, and it is not automatically a population-average effect.

Check Your Understanding

Use the model and distinctions in this tutorial to answer the following.

In a model of blood pressure (mmHg) on a binary exposure and age, what does the exposure coefficient compare?
A fitted exposure coefficient is 3.2 with standard error 1.1. What is the test statistic for \(H_0:\beta_A=0\), and what does that statistic measure?
Why can an adjusted coefficient differ from the crude difference in outcome means?
Name two assumptions needed beyond fitting a regression before interpreting its exposure coefficient causally.
Under what kind of model condition can a conditional exposure contrast equal a population-average contrast?