Interpreting Multiple-Regression Coefficients

What a Multiple-Regression Slope Describes

A multiple-regression equation includes several regressors at once. Its slopes are therefore read differently from slopes in a regression with just one regressor: each slope describes how the fitted outcome changes when one regressor changes and the other included regressors are held fixed. This is a precise statement about the equation, not automatically a statement about what would happen if someone intervened on that regressor.

The distinction matters because two quantities that sound similar can be different: the unadjusted association between an outcome and one regressor, and the association represented by that regressor's slope in a multiple regression. A numerical example will show how the two can differ even when all calculations are straightforward.

Definition: For an observed sample with outcome \(y_i\) and regressors \(x_{i1},\ldots,x_{ik}\), the ordinary least squares (OLS) regression with an intercept chooses coefficients to minimize the sum of squared residuals. Its fitted value for observation \(i\) is \(\hat y_i=\hat\beta_0+\hat\beta_1x_{i1}+\cdots+\hat\beta_kx_{ik}\). The OLS slope \(\hat\beta_j\) is the change in this fitted value for a one-unit increase in \(x_j\), with all other included regressors held fixed.

For a unique set of OLS coefficients, the sample design matrix must have full column rank: no included column can be an exact linear combination of the others. This condition ensures that different coefficient vectors do not produce the same fitted values. It is not necessary to assume normally distributed errors, or any other stochastic error model, to interpret a fitted-value difference algebraically.

The Change-in-Fitted-Value Calculation

The meaning of a slope follows directly from the fitted equation. Compare two sets of regressor values that differ only in \(x_j\): increase \(x_j\) by \(\Delta\), and leave the other included regressors unchanged. The two fitted values differ by the following amount.

\begin{aligned} \hat y_{\mathrm{new}}-\hat y_{\mathrm{old}} &=\left(\hat\beta_0+\sum_{\ell\ne j}\hat\beta_\ell x_\ell +\hat\beta_j(x_j+\Delta)\right) -\left(\hat\beta_0+\sum_{\ell\ne j}\hat\beta_\ell x_\ell +\hat\beta_jx_j\right)\\ &=\hat\beta_j\Delta. \end{aligned}

The intercept and the terms for the other regressors cancel because their values are unchanged. In particular, a one-unit increase (\(\Delta=1\)) changes the fitted outcome by \(\hat\beta_j\) outcome units. A two-unit increase changes it by \(2\hat\beta_j\), as long as the same linear equation is being used.

The phrase “holding the other included regressors fixed” is essential. It describes a comparison within the regression equation: the value of the outcome predicted by the equation at two regressor combinations. It does not say that the other regressors are fixed in the observed population, or that this comparison is necessarily possible for every combination of values.

Worked Example: Reading Two Slopes in Context

Suppose a hypothetical OLS regression predicts a student's exam score, measured in points, from hours of study and hours of sleep the night before the exam: \(\hat y=42+3.2\,\text{Study}+1.1\,\text{Sleep}\). Interpret both slopes and calculate the fitted-score difference for a specified change.

The study coefficient is \(3.2\) points per study hour. For students with the same recorded sleep hours, a one-hour increase in study hours corresponds to a fitted exam score \(3.2\) points higher. This is a comparison of fitted scores, not proof that an extra study hour causes a \(3.2\)-point increase.

The sleep coefficient is \(1.1\) points per sleep hour. For students with the same study hours, a one-hour increase in sleep corresponds to a fitted exam score \(1.1\) points higher.

For example, compare 5 and 7 study hours while keeping sleep fixed at 8 hours. The fitted-score difference is \(\hat\beta_{\text{Study}}\Delta=3.2(7-5)=3.2(2)=6.4\) points. Direct substitution gives the same result: at 5 study hours the fitted score is \(42+3.2(5)+1.1(8)=66.8\); at 7 hours it is \(42+3.2(7)+1.1(8)=73.2\); and \(73.2-66.8=6.4\).

Partial and Unadjusted Associations Are Different

The multiple-regression slope for \(x_j\) is often called a partial slope: it describes the fitted-outcome change associated with \(x_j\) while the other included regressors are fixed. By contrast, a simple regression of \(y\) on \(x_j\) alone gives an unadjusted slope. That slope describes the fitted change as \(x_j\) changes without accounting for the other regressors in a multiple-regression equation.

The two slopes need not agree. If another regressor varies along with \(x_j\) and is related to the outcome in the data, the simple regression does not hold that regressor fixed. The calculation below demonstrates the difference directly; it does not require treating either slope as causal.

Worked Example: Calculating Partial and Unadjusted Slopes

Consider these hypothetical observations on an outcome \(y\), a regressor \(x\), and another included regressor \(z\). The units are arbitrary, and the data are constructed for illustration.

Observation	\(x\)	\(z\)	\(y\)
1	0	0	10
2	1	2	18
3	2	1	17
4	3	3	25

First consider the multiple-regression equation \(\hat y=10+2x+3z\). Substituting each row's values gives \(10, 18, 17,\) and \(25\), exactly the four observed outcomes. Thus this equation has zero residual sum of squares, the smallest possible value. The regressors are not exactly linearly dependent: the \(z\) values \(0,2,1,3\) do not follow a straight-line pattern as \(x\) takes values \(0,1,2,3\). The design has full column rank, so the OLS coefficients are unique. In this regression, the partial slope for \(x\) is \(2\): for a one-unit increase in \(x\), the fitted \(y\) increases by 2 units when \(z\) is fixed.

Now calculate the simple-regression slope of \(y\) on \(x\). The means are \(\bar{x}=(0+1+2+3)/4=1.5\) and \(\bar{y}=(10+18+17+25)/4=17.5\). The simple-regression slope is the sum of the cross-products of deviations divided by the sum of squared deviations in \(x\):

\begin{aligned} \hat\beta_{\text{simple}} &=\frac{\sum_i(x_i-\bar{x})(y_i-\bar{y})} {\sum_i(x_i-\bar{x})^2}\\ &=\frac{(-1.5)(-7.5)+(-0.5)(0.5)+(0.5)(-0.5)+(1.5)(7.5)} {(-1.5)^2+(-0.5)^2+(0.5)^2+(1.5)^2}\\ &=\frac{11.25-0.25-0.25+11.25}{2.25+0.25+0.25+2.25} =\frac{22}{5}=4.4. \end{aligned}

The unadjusted slope is \(4.4\), whereas the partial slope for \(x\) in the regression that includes \(z\) is \(2\). In the simple regression, the fitted change reflects the observed pairing of \(x\) and \(y\) without holding \(z\) fixed. In the multiple regression, the coefficient for \(x\) describes a change in the fitted equation at the same value of \(z\). The arithmetic shows that these are different summaries of the data.

Units and the Scale of a Regressor

A coefficient's units are the outcome units divided by the units used to measure its regressor. A coefficient of 2 points per hour means something different from a coefficient of 2 points per 100 hours. Always check how each variable is recorded before describing a one-unit change.

Worked Example: Interpreting a Rescaled Regressor

A hypothetical regression predicts monthly store sales in dollars using advertising expenditure \(A\), measured in thousands of dollars, and price \(P\), measured in dollars: \(\hat y=12{,}000+850A-400P\). Interpret the advertising coefficient and calculate the fitted change when advertising expenditure increases from \(A=6\) to \(A=8\), holding price at \(P=20\).

The advertising slope is 850 dollars of fitted monthly sales per one-unit increase in \(A\). Since \(A\) is measured in thousands of dollars, one unit means an additional \(1{,}000\) dollars of advertising expenditure. The fitted sales difference is \(\hat\beta_A\Delta A=850(8-6)=850(2)=1{,}700\) dollars, with price held at 20 dollars.

Checking by substitution, the fitted sales at \(A=6\), \(P=20\) are \(12{,}000+850(6)-400(20)=9{,}100\) dollars. At \(A=8\), \(P=20\), they are \(12{,}000+850(8)-400(20)=10{,}800\) dollars. Their difference is \(10{,}800-9{,}100=1{,}700\) dollars. The price slope is \(-400\) dollars of fitted monthly sales per one-dollar increase in price, with advertising expenditure held fixed.

What the Coefficient Does—and Does Not—Claim

An OLS slope is a feature of a fitted equation for the chosen sample and included regressors. Its algebraic interpretation is exact: change one regressor by a stated amount, keep the others fixed, and the fitted outcome changes by the slope times that amount. This interpretation does not depend on whether the regression is a good description of every observation.

The same algebra does not establish a causal effect. For a causal interpretation, additional assumptions about how the data were generated and how the regressors relate to unobserved determinants of the outcome would be needed. A fitted comparison between two regressor combinations is not, by itself, evidence that changing a regressor would produce that outcome change in the real world.

Also check whether the “holding fixed” comparison is supported by the data. If observations with different values of \(x_j\) do not occur at similar values of the other regressors, the equation still defines a fitted difference, but that comparison may depend on extrapolating beyond combinations well represented in the sample. The coefficient's arithmetic meaning remains clear; its empirical relevance may be limited.

Common Mistakes

Leaving out what is held fixed. Saying “the \(x\) slope is the change in \(y\)” is incomplete in a multiple regression. State that it is the change in fitted \(y\) for a one-unit change in \(x\), with the other included regressors fixed.
Calling a fitted change a causal effect. The regression equation describes fitted values. A causal claim requires additional justification beyond reading the coefficient.
Confusing adjusted and unadjusted slopes. The simple-regression slope does not hold other regressors fixed. It can differ from the corresponding partial slope in a multiple regression.
Ignoring measurement units. If a regressor is measured in thousands of dollars, a one-unit increase means an increase of \(1{,}000\) dollars, not one dollar.
Assuming every fixed-regressor comparison is observed. A coefficient defines the equation's fitted change, but the data may contain little or no support for some combinations of regressor values.

Key takeaway: In a multiple regression, an OLS slope is the change in the fitted outcome for a one-unit increase in its regressor, with the other included regressors held fixed. State the outcome units, regressor units, and what is held fixed; do not confuse this partial fitted relationship with an unadjusted association or a causal effect.

Check Your Understanding

For each question, give the interpretation in context and show any requested calculation.

A fitted equation is \(\hat y=5+2x-3z\). What is the fitted-outcome change for a three-unit increase in \(x\) with \(z\) fixed?
In a model predicting annual earnings in dollars, a slope is 1,200 for training measured in hours. State the fitted-outcome interpretation of a one-unit increase in training, holding the other included regressors fixed.
Why can the slope from a simple regression of \(y\) on \(x\) differ from the slope for \(x\) in a multiple regression that also includes \(z\)?
A regressor is recorded in thousands of dollars and has a coefficient of \(-250\). What fitted-outcome change corresponds to a \(4{,}000\)-dollar increase in that regressor, with the other included regressors fixed?
Does the algebraic interpretation of an OLS slope alone establish that changing the regressor causes the stated change in the outcome? Explain briefly.