Interpreting Coefficients in Outcome-Relevant Units

From a Coefficient to a Contextual Statement

A regression coefficient is a number on a particular scale. By itself, “the slope is 1.7” says little: the reader needs to know what outcome is changing, what predictor is changing, and the units of both. A useful interpretation supplies that information and describes the change in the expected outcome, not a guaranteed change for every individual.

The basic rule is straightforward: in a linear model without interactions involving a predictor, its coefficient describes the expected outcome change associated with a one-unit increase in that predictor, with other predictors held fixed. “One unit” means one unit in the way the predictor was recorded. If the predictor is measured in days, it means one day; if it is measured in kilometres, it means one kilometre.

Definition: In an additive linear regression model, the coefficient of a quantitative predictor is the change in the model’s expected outcome for a one-unit increase in that predictor, with all other predictors held fixed. Its units are outcome units per predictor unit. The coefficient describes a conditional association under the stated model; it does not, by itself, establish a causal effect.

The Model Gives the Interpretation

Consider the multiple linear regression model

Y_i=\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}+\varepsilon_i, \qquad E(\varepsilon_i\mid x_{i1},\ldots,x_{ip})=0.

Here \(Y_i\) is the outcome for observation \(i\), the \(x_{ij}\) are predictors, and \(\beta_j\) is the population coefficient for predictor \(j\). The zero conditional-mean assumption makes the model’s conditional expected outcome equal to its linear predictor. In a fitted model, \(b_j\) denotes the estimated coefficient corresponding to \(\beta_j\).

For this interpretation, the model’s conditional mean must be linear and additive in the predictors shown: it contains no interaction terms that change the slope for \(x_j\). The algebraic interpretation also compares outcomes at the same values of the other predictors. In applications, such comparisons should be supported by the data; a fitted equation does not make every combination of predictor values informative or realistic.

To see why the units follow the rule, compare the expected outcomes at two predictor settings that differ by one unit in \(x_j\), while every other predictor stays fixed:

E(Y\mid x_j+1,\text{ other predictors}) - E(Y\mid x_j,\text{ other predictors}) =\beta_j.

The estimated version replaces \(\beta_j\) with \(b_j\). If the outcome is measured in centimetres and \(x_j\) in days, then \(b_j\) is measured in centimetres per day. If \(b_j\) is negative, the model predicts a lower expected outcome at the larger predictor value; if it is positive, the expected outcome is higher. The coefficient is not a percentage change unless the model and variables are specifically defined on a scale that supports that interpretation.

Worked Example: A Slope for Plant Growth

Suppose a hypothetical greenhouse analysis fits a simple linear regression for plant height \(Y\), in centimetres, using time under a specified light condition \(x\), in days. The fitted model is \(\widehat{Y}=12.4+1.7x\). Assume the conditional mean is adequately represented by a straight line over the observed days, and there are no other predictors or interactions in this model. Interpret the slope.

The estimated slope is \(b_1=1.7\), with units of centimetres per day. For a one-day increase in time under the light condition, the fitted model’s expected plant height increases by

Calculation: \(1.7\ \text{cm/day}\times 1\ \text{day}=1.7\ \text{cm}\).

In context, plants observed one day further along this time scale are predicted to have an average height 1.7 cm greater, according to the fitted model. This describes an expected difference, not a claim that each plant grows exactly 1.7 cm in a day. Because the data are hypothetical and observational, the slope alone also does not show that additional time under the light caused the difference.

Coefficients in Multiple Regression

In multiple regression, a coefficient is interpreted conditional on the other predictors in the model. This qualification is not an optional phrase: it says which comparison the coefficient represents. For example, a distance coefficient adjusted for congestion compares observations with the same modeled congestion value, rather than comparing any two trips regardless of congestion.

The coefficient’s units still come from the outcome and the predictor whose coefficient is being interpreted. Including other predictors changes the comparison represented by the coefficient, but does not change that units calculation. The fitted model’s additive form is important: absent interactions, the expected one-unit change is the same at all fixed values of the other predictors.

Worked Example: Distance Adjusted for Congestion

Suppose a hypothetical model predicts a trip’s delay \(Y\), in minutes, from distance \(x_1\), in kilometres, and a congestion score \(x_2\), in points:

\widehat{Y}=3.0+2.4x_1+1.1x_2.

Assume the model’s conditional mean is adequately represented by this additive linear form over the predictor values being compared. The distance coefficient is \(b_1=2.4\ \text{minutes per kilometre}\). To interpret it, hold the congestion score fixed and increase distance by one kilometre.

Substitution: the fitted expected-delay difference is \(2.4(1)+1.1(0)=2.4\) minutes. The distance term changes by 2.4 minutes; the congestion term does not change because the score is held fixed.

Thus, for trips with the same modeled congestion score, a one-kilometre greater distance is associated with a predicted delay that is 2.4 minutes higher on average. This is an adjusted association, not necessarily a causal effect. The interpretation is tied to the model and the observed comparisons; it should not be extended to distances or predictor combinations unsupported by the data.

Indicator Predictors and Interactions

A categorical predictor with two groups is often represented by an indicator \(D\), coded 0 for one group and 1 for the other. In an additive model without interactions involving \(D\), its coefficient is the expected outcome difference between the \(D=1\) and \(D=0\) groups, conditional on the other predictors. The direction depends on the coding: reversing which group is coded 1 reverses the sign of the coefficient.

That interpretation changes when the model includes an interaction. In a model with a product term \(DZ\), the group difference depends on \(Z\); the coefficient of \(D\) alone gives the difference only when \(Z=0\). This is a direct consequence of subtracting the model predictions for the two indicator values.

Worked Example: A Group Difference That Depends on Severity

Suppose a hypothetical model predicts follow-up pain score \(Y\), in points, from a treatment indicator \(D\) and baseline severity \(Z\), measured in points:

\widehat{Y}=8.0-1.2D+0.5Z+0.15DZ.

Here \(D=1\) denotes the treatment group and \(D=0\) the comparison group. Assume this interaction model is the stipulated form for the conditional mean. To find the fitted group difference at baseline severity \(Z\), subtract the prediction for \(D=0\) from the prediction for \(D=1\):

\begin{aligned} \widehat{Y}(D=1)-\widehat{Y}(D=0) &=(8.0-1.2+0.5Z+0.15Z)-(8.0+0.5Z)\\ &=-1.2+0.15Z. \end{aligned}

The coefficient \(-1.2\) has units of pain-score points per one-unit change in \(D\). Since \(D\) changes from 0 to 1, the coefficient gives the predicted treatment-group difference at \(Z=0\), not at every severity level. The interaction coefficient \(0.15\) has units of pain-score points per severity point and makes the group difference vary with baseline severity.

For example, at \(Z=4\) points, the fitted difference is \(-1.2+0.15(4)=-1.2+0.6=-0.6\) pain-score points. At that severity value, the treatment group’s predicted score is 0.6 points lower. It would be incorrect to report \(-1.2\) points as the group difference at all severity values. These fitted contrasts describe model-based associations and, without an appropriate causal design and assumptions, do not establish a treatment effect.

Check the Units and the Comparison

A coefficient interpretation is most useful when it answers four questions: what outcome is being described, what predictor changes, by how much, and what is held fixed? For a quantitative predictor in an additive model, the change is one recorded predictor unit. For an indicator, it is the change from the coded reference group to the coded comparison group. If an interaction is present, state the value of the interacting predictor or express how the coefficient-defined difference depends on it.

The intercept is different from a slope. It is the model’s expected outcome when every quantitative predictor equals zero and every indicator is at its coded zero value. That may or may not describe a meaningful case in the study. A slope interpretation should not be confused with the intercept, nor should a coefficient’s numerical size be interpreted without its units.

Key takeaway: Report a regression coefficient in context as an expected outcome change, with outcome units per predictor unit and the relevant conditions stated. For an additive model, hold the other predictors fixed. With an interaction, derive the conditional change from the full model rather than treating one coefficient as a universal effect.

Common Mistakes

Leaving out units. “The coefficient is 2.4” is incomplete. Say “2.4 minutes per kilometre” and name the outcome and predictor.
Describing an individual’s guaranteed change. A slope concerns the model’s expected outcome, not the exact outcome for every observation.
Forgetting what is held fixed. In multiple regression, describe the coefficient as conditional on the other included predictors.
Calling an association a cause. A fitted coefficient alone does not establish that changing the predictor would cause the outcome to change.
Ignoring an interaction. In a model containing \(DZ\), the coefficient of \(D\) is the group difference only at \(Z=0\); at other values, use the full conditional difference.

Check Your Understanding

Answer each question in context, including units where possible.

A fitted model predicts test score points from study time in hours, and the slope is 3.2. What is the expected score difference associated with one additional hour, under an additive linear model?
A model predicts delivery time in minutes from distance in kilometres and number of stops. How should the distance coefficient be described?
An indicator is coded 1 for one clinic and 0 for another in an additive model. What comparison does its coefficient represent?
In a model with terms \(D\), \(Z\), and \(DZ\), why is the coefficient of \(D\) not generally the group difference at every value of \(Z\)?
Why does a positive regression coefficient not, by itself, show that increasing the predictor causes the outcome to increase?