Two Slopes, Two Comparisons
A regression slope can summarize how an outcome varies with a predictor, but its meaning depends on which other predictors are included in the model. A one-predictor regression describes the overall linear association between the outcome and that predictor. A multiple regression describes the association with the other included predictors held fixed. These are different summaries, so their slopes need not match—even when they use the same outcome, predictor, and observations.
For example, a slope from a regression of examination score on study time alone does not make distinctions between students with different prior preparation. A slope from a regression that includes both study time and prior preparation compares students at the same prior-preparation value. The second slope is adjusted for the included predictor; that description does not, by itself, mean that the slope is more accurate or that it represents a causal effect.
These definitions concern fitted linear associations. The coefficients are calculated from data; they are not automatically descriptions of a causal effect in a population. The models also make specific summaries: the one-predictor line represents a straight-line trend, and the additive multiple-regression equation represents a plane with no \(x\)-by-\(z\) interaction. In the latter equation, the fitted change associated with a one-unit increase in \(x\) is the same at each fixed value of \(z\).
What Each Slope Summarizes
Suppose the outcome \(y\) is measured in points and the predictor \(x\) in hours. If the one-predictor fitted slope is \(4.4\), then the fitted outcome differs by \(4.4\) points for a one-hour difference in \(x\), according to that overall straight-line summary. This is not a statement that every pair of people differing by one hour will differ by exactly \(4.4\) points.
Now suppose a multiple regression also includes a predictor \(z\). If the fitted coefficient of \(x\) is \(2.0\), the model summarizes a \(2.0\)-point difference in fitted outcome for a one-hour difference in \(x\), comparing profiles with the same \(z\). It does not compare the two people in every other respect; it holds fixed only the predictor included in this model. Adding or removing predictors changes the model and can change the coefficient being summarized.
| Feature | One-predictor slope | Multiple-regression slope |
|---|---|---|
| Model | \(\hat{y}=b_0+b_1x\) | \(\hat{y}=b_0+b_1x+b_2z\) |
| Comparison summarized | Overall fitted change as \(x\) varies | Fitted change as \(x\) varies at fixed \(z\) |
| Other predictors | Not included in this regression | Included predictors are held fixed in the interpretation |
The word marginal here means that the slope does not condition on another predictor in the regression. It does not mean that the slope is an average of all possible adjusted slopes, nor that it is unimportant. Likewise, adjusted means conditional on the predictors included in the model; it does not mean “corrected” in a universal sense. The choice of model determines which comparison is being described.
Worked Example: Calculating Both Slopes
The following hypothetical data give an outcome \(y\), predictor \(x\), and additional predictor \(z\) for four observations. Calculate the one-predictor slope for \(x\), then compare it with the coefficient of \(x\) in the multiple regression.
| Observation | \(x\) | \(z\) | \(y\) |
|---|---|---|---|
| 1 | 1 | 1 | 15 |
| 2 | 2 | 3 | 23 |
| 3 | 3 | 2 | 22 |
| 4 | 4 | 4 | 30 |
For the one-predictor regression, the least-squares slope is the sum of cross-products of deviations divided by the sum of squared deviations in \(x\):
The means are \(\bar{x}=2.5\) and \(\bar{y}=22.5\). The denominator is \(2.25+0.25+0.25+2.25=5\). The numerator is \(11.25-0.25-0.25+11.25=22\). Therefore,
The one-predictor intercept is \(b_0=\bar{y}-b_1\bar{x}=22.5-(4.4)(2.5)=11.5\). The marginal fitted line is therefore \(\hat{y}=11.5+4.4x\).
For the multiple regression, the data satisfy \(y=10+2x+3z\) in every row. For example, at \(x=2\) and \(z=3\), \(10+2(2)+3(3)=23\), the recorded outcome. The same check gives 15, 22, and 30 for the other rows. Thus the fitted multiple-regression equation is \(\hat{y}=10+2x+3z\), with adjusted coefficient \(2\) for \(x\).
In this hypothetical data set, the marginal slope is \(4.4\) outcome units per \(x\)-unit, whereas the adjusted slope is \(2\) outcome units per \(x\)-unit at fixed \(z\). The difference is \(4.4-2=2.4\) outcome units per \(x\)-unit. The slopes differ because they summarize different fitted comparisons; neither number should be substituted for the other without stating which model it comes from.
Why the Numbers Can Differ
The two models choose their fitted lines or planes by minimizing different sums of squared residuals. A residual is an observed outcome minus its fitted value. In the one-predictor model, the fitted values use \(x\) alone; in the multiple model, they use both \(x\) and \(z\). Consequently, each fit allocates the observed variation differently, and the coefficient of \(x\) can change when \(z\) is included.
This distinction can be seen from the objectives that least squares minimizes. In the one-predictor model, it chooses \(b_0\) and \(b_1\) to minimize the sum of squared differences between observed and fitted outcomes. In the multiple model, it chooses \(b_0\), \(b_1\), and \(b_2\) to minimize the corresponding sum using both predictors:
Because the fitted values and the minimization problems differ, the value of \(b_1\) need not be the same. The difference does not, by itself, establish why the slopes changed, whether the multiple-regression model is preferable, or whether either slope has a causal interpretation. Those judgments require context and additional reasoning.
The slopes can also agree. For instance, if \(x\) and \(z\) have no cross-product association in the observed data, adding \(z\) may leave the fitted coefficient of \(x\) unchanged. Agreement does not make the interpretations identical: the one-predictor slope remains an overall fitted association, while the multiple-regression slope remains a comparison at fixed \(z\).
Worked Example: When the Slopes Agree
Consider four hypothetical observations with \(x\), \(z\), and \(y\) as follows. Here the outcome values follow \(y=5+2x+4z\).
| Observation | \(x\) | \(z\) | \(y\) |
|---|---|---|---|
| 1 | 1 | 1 | 11 |
| 2 | 1 | 3 | 19 |
| 3 | 3 | 1 | 15 |
| 4 | 3 | 3 | 23 |
For the one-predictor slope, \(\bar{x}=2\) and \(\bar{y}=17\). The sum of squared deviations in \(x\) is \((-1)^2+(-1)^2+1^2+1^2=4\). The sum of cross-products is \((-1)(-6)+(-1)(2)+(1)(-2)+(1)(6)=8\). Thus,
The multiple-regression equation fits all four observations exactly: \(11=5+2(1)+4(1)\), \(19=5+2(1)+4(3)\), \(15=5+2(3)+4(1)\), and \(23=5+2(3)+4(3)\). Its adjusted coefficient of \(x\) is also \(2\). Both slopes therefore equal \(2\) outcome units per \(x\)-unit in this data set.
The numerical agreement does not erase the distinction between the models. The first slope summarizes the overall line using \(x\) alone; the second summarizes the fitted change in \(y\) for a one-unit increase in \(x\) at the same value of \(z\).
Interpreting Reported Slopes in Context
Worked Example: Comparing Apartment-Area Slopes
Suppose a hypothetical analysis of 120 apartments reports a marginal slope of \(185\) dollars per month for area, measured in units of 100 square feet. A multiple regression that also includes number of bedrooms reports an area coefficient of \(142.4\) dollars per month per 100 square feet. The outcome in both models is monthly rent, and both use the same observations.
The marginal fitted association is \(185\) dollars per month for each additional 100 square feet, according to the one-predictor line. The adjusted fitted association is \(142.4\) dollars per month for each additional 100 square feet when comparing apartments with the same number of bedrooms, according to the multiple-regression model.
The numerical difference is \(185-142.4=42.6\) dollars per month per 100 square feet. This arithmetic comparison is valid because both slopes use the same predictor units, outcome units, and observations. It does not mean that adding 100 square feet causes rent to rise by either amount. It also does not mean the adjusted slope is automatically the better summary; that depends on the question being asked and the adequacy of the model for that purpose.
Common Mistakes and a Practical Check
- Calling the marginal slope “the effect without adjustment.” A one-predictor slope is an overall fitted association, not automatically a causal effect. Describe it as an association unless the study design and assumptions justify a causal claim.
- Describing an adjusted slope without naming what is fixed. Say “at the same number of bedrooms” when that is the predictor included in the model; do not imply that all other features are held fixed.
- Assuming the adjusted slope must be larger, smaller, or more correct. Its value can differ in either direction, or agree with the marginal slope. The model and the question determine which summary is relevant.
- Comparing numbers with different units or samples as though they were directly comparable. Check that the outcome units, predictor units, and observations match before interpreting the difference between slopes.
Before reporting either coefficient, identify the model that produced it, state the outcome and predictor units, and specify whether the comparison is overall or at fixed values of the included predictors. This makes clear both what the slope says and what it does not say.
Check Your Understanding
Answer each question by distinguishing the model from the interpretation.
- In the model \(\hat{y}=b_0+b_1x\), what kind of association does \(b_1\) summarize?
- In \(\hat{y}=b_0+b_1x+b_2z\), explain in context what the coefficient \(b_1\) compares.
- Can the marginal slope and adjusted slope for the same predictor have different values? Give a reason based on what the two models fit.
- If the two slopes have the same numerical value, do they necessarily summarize the same comparison? Explain.
- A multiple-regression slope is positive. Does that alone show that increasing \(x\) causes \(y\) to increase? Why or why not?