Reading Coefficients as Adjusted Effects

What a Multiple-Regression Slope Describes

In a multiple-regression model, a quantitative predictor’s coefficient describes its relationship with the response conditional on the model’s other predictors. For example, a coefficient for study time in a model that also includes sleep time compares students with the same sleep time but different study times. This is why the coefficient is often described as an adjusted slope.

“Adjusted” has a specific, limited meaning: the comparison keeps the other predictors included in the model fixed. It does not mean that every possible difference between the individuals has been controlled, and it does not by itself establish a cause-and-effect relationship.

Definition: In a multiple-regression model with quantitative predictors, the coefficient for a predictor is its partial regression coefficient. It describes the change in the model’s expected response associated with a one-unit increase in that predictor, while the model’s other predictors are held fixed. Its units are response units per unit of that predictor.

The Model and the Meaning of “Held Fixed”

For a response \(Y\) and \(k\) quantitative predictors \(X_1,\ldots,X_k\), consider the additive conditional-mean model

E(Y\mid X_1=x_1,\ldots,X_k=x_k) = \beta_0+\beta_1x_1+\cdots+\beta_kx_k.

Here \(E(Y\mid X_1=x_1,\ldots,X_k=x_k)\) is the expected response among cases with the stated predictor values. The coefficient \(\beta_j\) is the model’s slope for \(X_j\); the sample fit estimates it with a coefficient such as \(b_j\). The fitted model uses estimated coefficients to describe or estimate the conditional mean.

This interpretation assumes the conditional mean is adequately represented by the stated additive linear model over the predictor values being compared. In particular, the model assigns the same slope to \(X_j\) at different values of the other predictors. If the comparison requires predictor combinations that are absent or poorly represented in the data, the coefficient’s mathematical interpretation still follows from the model, but that comparison may be weakly supported by observed data.

To interpret \(\beta_j\), compare two cases whose values of every predictor other than \(X_j\) are the same. If \(X_j\) increases by \(\Delta x_j\), the model’s expected response changes by \(\beta_j\Delta x_j\). For an increase of one unit, \(\Delta x_j=1\), so the change is \(\beta_j\). This is a difference in expected responses, not a promise that each individual response changes by that amount.

Why the Slope Is an Adjusted Comparison

The meaning follows directly by subtracting two conditional means. Suppose the predictor values differ only in \(X_j\): the first case has \(X_j=x_j\), and the second has \(X_j=x_j+\Delta x_j\). Every other predictor value is the same in both cases. Their model-based expected-response difference is

\begin{aligned} &E(Y\mid X_j=x_j+\Delta x_j,\text{ other predictors fixed})\\ &\quad-E(Y\mid X_j=x_j,\text{ other predictors fixed})\\ &= \left(\beta_0+\sum_{\ell\ne j}\beta_\ell x_\ell+ \beta_j(x_j+\Delta x_j)\right) - \left(\beta_0+\sum_{\ell\ne j}\beta_\ell x_\ell+ \beta_jx_j\right)\\ &=\beta_j\Delta x_j. \end{aligned}

The intercept and the terms for the other predictors cancel because those quantities are held fixed. When \(\Delta x_j=1\), the expected-response difference is \(\beta_j\). The fitted coefficient \(b_j\) gives the corresponding estimated difference, in the same units.

This cancellation is the key to reading the coefficient correctly. The slope is not simply the change in response seen when \(X_j\) increases in a data set where other predictors may change too. It is the model’s comparison for a change in \(X_j\) with the other included predictors unchanged.

Worked Example: Study Time Adjusted for Sleep

Suppose a hypothetical fitted model relates quiz score, in points, to study time, in hours, and sleep time, in hours:

\widehat{E(Y\mid X_1,X_2)} = 42+4.2X_1+1.1X_2,

where \(X_1\) is study time and \(X_2\) is sleep time. Interpret the study-time coefficient and calculate the estimated difference for a 2.5-hour increase in study time when sleep time is unchanged.

The coefficient of study time is \(4.2\), with units of quiz-score points per study hour. Under this model, for students with the same sleep time, one additional hour of study is associated with an estimated increase of \(4.2\) points in expected quiz score.

For a 2.5-hour increase, the model-based difference is

4.2\ \frac{\text{points}}{\text{hour}}\times 2.5\ \text{hours} =10.5\ \text{points}.

Thus, at the same sleep time, the fitted model estimates a 10.5-point difference in expected quiz score between study times that differ by 2.5 hours. This is an average model-based comparison, not a guarantee of a 10.5-point increase for every student.

Units, Direction, and Size

A coefficient’s sign gives the direction of the model-based relationship, and its magnitude gives the expected response difference per one predictor unit. Always name both variables and state the units. If \(b_j\) is negative, a one-unit increase in \(X_j\), with the other predictors fixed, corresponds to a decrease of \(|b_j|\) units in the expected response. A positive coefficient describes an increase under the same comparison.

The size of a coefficient depends on the units used to measure its predictor. Measuring distance in meters rather than kilometers changes the numerical slope, but not the underlying fitted relationship. To interpret a change that is not one unit, multiply the coefficient by the size of the change, keeping its direction. The result has the response’s units.

Worked Example: A Negative Slope in a Health Model

A hypothetical model estimates weekly exercise minutes from age, measured in years, and resting heart rate, measured in beats per minute. The estimated coefficient for age is \(-0.80\) minutes per week per year. Interpret it for a comparison of people whose ages differ by 5 years but whose resting heart rates are the same.

For a one-year increase in age at the same resting heart rate, the fitted model associates age with a decrease of \(0.80\) minutes in expected weekly exercise time. For a 5-year increase, the estimated difference is

(-0.80)\ \frac{\text{minutes per week}}{\text{year}} \times 5\ \text{years} =-4.0\ \text{minutes per week}.

In the model, the person who is 5 years older has an expected weekly exercise time 4.0 minutes lower, when resting heart rate is held fixed. This interpretation concerns the conditional mean described by the model. It does not say that every older person exercises less, nor that age itself caused the difference.

Adjusted Does Not Mean Causal

A multiple-regression coefficient is an adjusted association within the model: it compares expected responses at different values of one predictor while holding the other included predictors fixed. Whether that association can be interpreted causally depends on how the data were collected and on additional assumptions about the relationship between predictors and response. A regression calculation alone does not establish those assumptions.

The word “adjusted” also does not mean “adjusted for everything.” A model can hold fixed only predictors that are included and represented appropriately. Other relevant differences may remain. In addition, a comparison at fixed values is most credible when observations with those combinations of predictor values are reasonably represented in the data. A model-based contrast far outside the observed combinations can be an extrapolation rather than a well-supported comparison.

A partial slope may also differ from the association seen between the response and that predictor alone. The simple association can reflect changes in other predictors that tend to occur alongside the predictor of interest. The multiple-regression slope instead describes the model’s comparison at fixed values of the other included predictors. These answer different questions; neither coefficient should be substituted for the other without considering the question being asked.

Worked Example: Changing Predictor Units

A hypothetical model for travel time estimates a coefficient of \(0.30\) minutes per kilometer for distance, while holding road grade fixed. A colleague wants to express the same relationship per 100 meters. Convert the coefficient and interpret it.

Since \(100\) meters is \(0.1\) kilometer, the estimated change in expected travel time for a \(0.1\)-kilometer increase is

0.30\ \frac{\text{minutes}}{\text{kilometer}} \times 0.1\ \text{kilometer} =0.030\ \text{minutes}.

The equivalent coefficient per 100 meters is therefore \(0.030\) minutes per 100 meters, or \(1.8\) seconds per 100 meters because \(0.030\times60=1.8\). Holding road grade fixed, each additional 100 meters is associated in the model with an increase of \(1.8\) seconds in expected travel time. The unit conversion changes the coefficient’s numerical value, not the modeled relationship.

Common Mistakes

Leaving out what is held fixed. Saying “one more hour of study adds 4.2 quiz points” can sound unconditional. Say that the model estimates a 4.2-point increase in expected score per additional study hour, at the same values of the other included predictors.
Treating an expected change as an individual guarantee. A slope describes a difference in conditional means. Individual responses can vary around those means.
Calling an adjusted association a causal effect. Holding included predictors fixed in a regression model does not, on its own, prove that changing the predictor would cause the response to change.
Ignoring units or direction. A coefficient is measured in response units per predictor unit. For a negative coefficient, a positive increase in the predictor corresponds to a decrease in expected response.
Assuming every fixed-predictor comparison is supported by data. The model defines the comparison algebraically, but combinations far from those represented in the data can make the interpretation depend heavily on extrapolation.

Key takeaway: In an additive multiple-regression model, a quantitative predictor’s coefficient is the model-based change in expected response per one-unit increase in that predictor, with the other included predictors held fixed. State the units and the fixed-predictor comparison, and do not infer causation from the coefficient alone.

Check Your Understanding

For each question, focus on the conditional comparison described by the model.

A fitted model estimates a coefficient of \(2.4\) centimeters per month for a quantitative predictor. Interpret the coefficient while holding the model’s other predictors fixed.
If that predictor increases by 3 months, what model-based expected-response difference does the coefficient imply? Include the units and direction.
A coefficient is \(-1.6\) response units per predictor unit. What does the model imply for a 2-unit increase in the predictor when the other predictors stay fixed?
Explain why an adjusted regression slope is not automatically a causal effect.
Why might an interpretation for a combination of predictor values rarely observed in the data be less credible, even if the regression equation gives a numerical result?