Interpreting Conditional Effects in a Multiple Regression Output

Reading a Multiple-Regression Output

A multiple-regression table can show a positive coefficient for a predictor even when its unadjusted association with the response is weak or negative. There is no contradiction: a multiple-regression slope describes an association conditional on the other predictors in the model, whereas an unadjusted slope summarizes a separate regression using just one predictor. Correct interpretation requires identifying which quantity the output reports, stating what is held fixed, and reading the coefficient's uncertainty in the right units.

Consider the additive multiple linear regression model

Y_i=\beta_0+\beta_1X_{i1}+\cdots+\beta_pX_{ip}+\varepsilon_i.

Here $Y_i$ is the response for observation $i$, $X_{ij}$ is predictor $j$, $\beta_0,\ldots,\beta_p$ are unknown coefficients, and $\varepsilon_i$ is the error. The fitted model replaces the unknown coefficients by estimates, usually written $\hat{\beta}_0,\ldots,\hat{\beta}_p$. Unless stated otherwise, this tutorial considers a model with no interaction terms.

Definition: In an additive multiple linear regression, the coefficient $\beta_j$ is the change in the model's conditional mean response associated with a one-unit increase in $X_j$, holding all other included predictors fixed. The intercept $\beta_0$ is the model's mean response when every predictor equals zero.

This interpretation concerns the fitted model's conditional mean, not necessarily a causal effect. It also describes a one-unit change in the predictor's recorded scale. A coefficient of 0.8 per score point is not a coefficient of 0.8 per ten score points; the latter change corresponds to $10(0.8)=8$ response units.

Conditional and Unadjusted Associations

A multiple-regression slope compares observations at the same values of the other included predictors, as represented by the model. An unadjusted slope instead comes from regressing $Y$ on one predictor alone. When other predictors are associated with both that predictor and the response, the two slopes can differ in size or even sign. The difference alone does not establish which model is correct, nor does it show that changing a predictor would cause a response change.

The distinction can be made precise using the linear projection of $Y$ on a single predictor $X$. Its slope is the coefficient that minimizes the population mean squared error among straight-line predictions based on $X$. If $\operatorname{Var}(X)>0$, that slope is $\operatorname{Cov}(X,Y)/\operatorname{Var}(X)$. This definition does not require the conditional mean $E[Y\mid X]$ to be a straight line.

b_{\text{proj}}=\frac{\operatorname{Cov}(X,Y)}{\operatorname{Var}(X)}, \qquad a_{\text{proj}}=E[Y]-b_{\text{proj}}E[X].

The intercept follows because the linear projection has mean $a_{\text{proj}}+b_{\text{proj}}E[X]$, which must equal $E[Y]$. The slope follows from the projection's normal equation: the prediction error must have zero covariance with $X$. Thus $\operatorname{Cov}(X,Y)-b_{\text{proj}}\operatorname{Var}(X)=0$, giving the displayed formula.

In a sample, the ordinary least-squares slope from regressing $Y$ on $X$ alone is the corresponding sample covariance divided by the sample variance of $X$, provided the sample variance is positive. It is a sample estimate of the unadjusted projection slope. Neither the sample slope nor the population projection slope automatically equals $E[Y\mid X=x]$ at each $x$. That conditional-mean interpretation requires an additional linearity condition.

Worked Example: Interpreting an Adjusted Output in Context

A hypothetical insurer fits a model for annual expected claim cost, measured in thousands of dollars. Its predictors are $A$, a risk-score point total, and $B$, a safe-driving score in points. The fitted equation is $\widehat{Y}=12.5+0.80A-1.20B$. Interpret the slopes and intercept, and calculate the fitted response for $A=4$ and $B=6$.

Method. Read each coefficient as a conditional model association, keep the response and predictor units explicit, then substitute the stated values.

Risk-score slope. At the same safe-driving score, a one-point increase in $A$ is associated with an increase of $0.80$ thousand dollars in fitted annual expected claim cost, or $\$800$. This is an adjusted association because $B$ is held fixed.

Safe-driving slope. At the same risk score, a one-point increase in $B$ is associated with a decrease of $1.20$ thousand dollars, or $\$1{,}200$, in fitted annual expected claim cost.

Intercept. At $A=0$ and $B=0$, the fitted response is $12.5$ thousand dollars, or $\$12{,}500$. This is the model's zero-predictor prediction. If these values are outside the range observed in the data, the intercept may have little practical interpretation.

Fitted value. Substitute $A=4$ and $B=6$:

\hat{Y}=12.5+0.80(4)-1.20(6) =12.5+3.2-7.2 =8.5.

The fitted annual expected claim cost is $8.5$ thousand dollars, or $\$8{,}500$. This is a model prediction at the stated predictor values, not a claim that changing either score would cause the corresponding change in claim cost.

Common mistake avoided. Calling $0.80$ the unadjusted association between risk score and claim cost drops the condition “at the same safe-driving score.” The coefficient is conditional on $B$ being included and held fixed.

Why the Unadjusted Slope Can Differ

The covariance formula shows how an omitted predictor can contribute to an unadjusted slope. Suppose the conditional mean has the additive form $E[Y\mid X,Z]=\beta_0+\beta_X X+\beta_Z Z$, and suppose the error has mean zero and is uncorrelated with $X$. Taking covariances with $X$ gives

\operatorname{Cov}(X,Y) =\beta_X\operatorname{Var}(X)+\beta_Z\operatorname{Cov}(X,Z).

After dividing by $\operatorname{Var}(X)$, the unadjusted projection slope is

b_{\text{proj}} =\beta_X+\beta_Z\frac{\operatorname{Cov}(X,Z)}{\operatorname{Var}(X)}.

The second term is the contribution associated with the linear relationship between $X$ and the omitted predictor $Z$. If $X$ and $Z$ are uncorrelated, this term is zero under the stated conditions. Otherwise, the unadjusted slope generally differs from the conditional coefficient $\beta_X$.

Worked Example: Deriving an Unadjusted Projection From Moments

In a hypothetical study, $X$ is a monitoring score and $Z$ is a workload index. Suppose $E[X]=5$, $E[Z]=3$, $\operatorname{Var}(X)=4$, and $\operatorname{Cov}(X,Z)=3$. Assume the conditional mean model is $E[Y\mid X,Z]=86+2X+3Z$, with an error uncorrelated with $X$ and mean zero. Find the unadjusted linear projection of $Y$ on $X$, then compare its fitted value at $X=6$ with the multiple-regression conditional mean at $X=6,Z=3$.

Method. Use the covariance formula for the unadjusted projection slope, then use the predictor means to obtain its intercept.

Step 1: Find the overall mean of $Y$. Taking expectations in the conditional mean model gives

$$ E[Y]=86+2E[X]+3E[Z] =86+2(5)+3(3)=105. $$

Step 2: Find the unadjusted slope. The covariance contribution from $X$ is $2\operatorname{Var}(X)=2(4)=8$. The contribution from $Z$ is $3\operatorname{Cov}(X,Z)=3(3)=9$. The error contributes zero by assumption. Therefore,

\operatorname{Cov}(X,Y)=8+9=17, \qquad b_{\text{proj}}=\frac{17}{4}=4.25.

Step 3: Find the unadjusted intercept.

a_{\text{proj}}=E[Y]-b_{\text{proj}}E[X] =105-4.25(5)=105-21.25=83.75.

Thus the unadjusted linear projection is $83.75+4.25X$. Its fitted value at $X=6$ is $83.75+4.25(6)=83.75+25.5=109.25$. At $X=6,Z=3$, the multiple-regression conditional mean is $86+2(6)+3(3)=86+12+9=107$. These are different predictions because they represent different quantities: an unadjusted projection using $X$ alone and a conditional mean at specified values of both predictors.

The supplied moments determine the unadjusted projection, but they do not, by themselves, determine $E[Y\mid X=6]$. In particular, they do not specify $E[Z\mid X=6]$. If one additionally assumes that $E[Z\mid X]$ is linear in $X$, then its slope is $\operatorname{Cov}(X,Z)/\operatorname{Var}(X)=3/4$, and its intercept is $3-(3/4)(5)=-0.75$. Under that extra condition,

E[Y\mid X] =86+2X+3(-0.75+0.75X) =83.75+4.25X.

Only with this additional linearity condition does the unadjusted projection line equal the conditional mean at each $X$. At $X=6$, it would then give $109.25$ as that conditional mean. Without the condition, that interpretation is not justified.

Common mistake avoided. The projection $83.75+4.25X$ must not be described as $E[Y\mid X]$ merely because it is a straight line. The moments establish the projection; they do not establish the shape of the conditional mean.

Interpreting Coefficient Standard Errors

A coefficient estimate is subject to sampling variation: a different sample could produce a different estimated slope. Its standard error estimates the standard deviation of that sampling variation under the specified regression model and variance assumptions. It is expressed in the same units as the coefficient. A slope standard error therefore has units of response per predictor unit; it is not the standard deviation of individual responses around the regression line.

Under the classical linear model with a correctly specified linear conditional mean, independent errors with mean zero and constant variance, and a full-rank design matrix, the conventional least-squares standard errors quantify estimated coefficient uncertainty. Normal errors additionally give exact $t$-based inference in finite samples. If errors are dependent or heteroscedastic, the usual standard errors may not be appropriate; a suitable alternative variance estimator or model may be needed.

For a coefficient estimate $\hat{\beta}_j$, its standard error $SE(\hat{\beta}_j)$ can be used to form an interval when the relevant assumptions and degrees of freedom are specified. Under normal errors, a two-sided $100(1-\alpha)\%$ interval is

\hat{\beta}_j \ \pm\ t_{1-\alpha/2,\,n-p-1}\,SE(\hat{\beta}_j),

where $n$ is the sample size, $p$ is the number of predictors, and $t_{1-\alpha/2,\,n-p-1}$ is the appropriate $t$ critical value with $n-p-1$ residual degrees of freedom. This interval describes uncertainty about the coefficient, not the range of individual outcomes.

Worked Example: Quantifying Uncertainty in a Conditional Slope

A hypothetical regression of monthly processing hours on two predictors uses $n=28$ observations. The estimated slope for case count is $0.42$ hours per case, with standard error $0.15$. The model includes an intercept and both predictors. Under the normal-error model, use $t_{0.975,25}=2.060$ to form a 95% interval and interpret it.

Method. Use the coefficient estimate plus or minus the stated critical value times its standard error. The residual degrees of freedom are $28-2-1=25$.

Step 1: Calculate the margin of error.

$$ 2.060(0.15)=0.309. $$

Step 2: Calculate both endpoints.

0.42-0.309=0.111, \qquad 0.42+0.309=0.729.

The 95% interval is $0.111$ to $0.729$ monthly processing hours per case. In context, it describes uncertainty in the conditional case-count slope, holding the other included predictor fixed, under the stated model assumptions. The endpoints are rounded to three decimal places.

A one-case increase corresponds to an estimated $0.42$-hour difference in fitted monthly processing hours, conditional on the other predictor. The interval is for that slope. An interval for a multi-case change or a prediction for a new observation requires additional information; in particular, the standard error of a linear combination of coefficients depends on their covariances.

Common mistake avoided. The standard error $0.15$ is not a statement that individual processing times typically differ from the fitted values by $0.15$ hours. It quantifies uncertainty in the estimated slope.

Reading the Output Carefully

Before interpreting a coefficient, check what the response and predictors measure, their units, which other variables are included, and whether the model contains interactions. For the additive model here, “holding other predictors fixed” is the essential condition. If an interaction is present, a slope may depend on another predictor's value, so a single unconditional statement about that slope is not sufficient.

Also check whether the predictor values used to describe an intercept or fitted value are represented in the data. A zero-valued predictor may be impossible or far outside the observed range. The intercept remains part of the fitted equation, but its practical interpretation may be limited. Finally, coefficients describe model-based associations; observational data and a regression table alone do not establish causation.

Key takeaway: A multiple-regression coefficient is a conditional slope, not an unadjusted slope. An unadjusted regression is a separate linear projection, and its moments do not generally determine the conditional mean at every predictor value. A standard error measures estimated coefficient uncertainty under specified assumptions, not outcome scatter.

Check Your Understanding

For each problem, identify the quantity being interpreted and keep its units and conditioning explicit.

A fitted model is $\hat{Y}=7+0.5X-2Z$, where $Y$ is in hundreds of dollars. Interpret the slope for $X$, and find the fitted response at $X=4,Z=1$. (Answer: Holding $Z$ fixed, one unit higher in $X$ is associated with $0.5$ hundred dollars, or $\$50$, higher fitted $Y$; at $X=4,Z=1$, $\hat{Y}=7+2-2=7$, or $\$700$.)
Suppose $E[X]=2$, $E[Y]=10$, $\operatorname{Var}(X)=3$, and $\operatorname{Cov}(X,Y)=6$. Find the unadjusted linear projection. (Answer: Slope $=6/3=2$; intercept $=10-2(2)=6$; projection $6+2X$.)
In a model $E[Y\mid X,Z]=20+X+4Z$, suppose $\operatorname{Var}(X)=5$, $\operatorname{Cov}(X,Z)=-1$, and the error is uncorrelated with $X$. What is the unadjusted projection slope of $Y$ on $X$? (Answer: $1+4(-1/5)=0.2$.)
A coefficient estimate is $-0.30$, its standard error is $0.08$, and the supplied interval multiplier is $2.1$. Calculate the corresponding interval using the multiplier. (Answer: Margin $=2.1(0.08)=0.168$; endpoints are $-0.468$ and $-0.132$.)
For the conditional-mean model in question 3, are its stated moments enough to find $E[Y\mid X=4]$? State what is missing. (Answer: No. The value also requires $E[Z\mid X=4]$; the moments alone do not specify that conditional expectation.)

What You'll Learn