Introduction
A regression line describes how a fitted response changes with one predictor. With two quantitative predictors, the fitted values lie on a plane instead. Reading that plane means understanding two directions of change: one associated with the first predictor while the second is held fixed, and one associated with the second predictor while the first is held fixed.
These directions are represented by the two slopes in the fitted equation. Each slope has a sign and units, and each describes an adjusted association: the fitted relationship for one predictor at a fixed value of the other. That interpretation is about the model’s predicted response, not a guarantee that changing a predictor causes an outcome to change.
The Two-Predictor Fitted Plane
Let \(Y\) be a quantitative response and let \(X_1\) and \(X_2\) be two quantitative predictors. A standard additive linear model writes the response as a plane plus an error:
Here, \(\beta_0,\beta_1,\beta_2\) are population model coefficients, and \(\varepsilon\) is the error. The model assumes that the conditional mean response is linear and additive in the predictors:
Equivalently, the model’s error has conditional mean zero, \(E(\varepsilon\mid X_1,X_2)=0\). In this additive model, the slope for either predictor does not vary with the value of the other predictor. The fitted equation, calculated from data, is
The lowercase \(x_1\) and \(x_2\) are specified predictor values, \(\hat{y}\) is the fitted or predicted response at those values, and \(b_0,b_1,b_2\) are estimated coefficients. The plane is the collection of fitted values produced by this equation. The equation alone does not require normally distributed errors; inference about coefficients may require additional assumptions, but is not needed to read the directions of this fitted plane.
Reading the Slopes as Directions
To read \(b_1\), imagine moving in the direction that changes \(x_1\) while leaving \(x_2\) unchanged. A positive \(b_1\) means the fitted response rises in that direction; a negative \(b_1\) means it falls; and a zero \(b_1\) means the fitted plane is level in that direction. The same sign rules apply to \(b_2\), with \(x_1\) held fixed instead.
“Held fixed” is essential. The slope \(b_1\) is not simply the overall pattern between \(X_1\) and \(Y\) when \(X_2\) is allowed to vary. It describes the fitted relationship between \(X_1\) and the response among predictor settings with the same \(X_2\) value. This is why the slope is called adjusted for the other predictor.
The intercept \(b_0\) is the fitted response when both predictors equal zero. That may be a useful reference point if zero is meaningful and relevant, but it may have little practical interpretation if zero is impossible or far outside the observed predictor values. The slopes can still be meaningful even when the intercept is not.
Worked Example: Study Hours and Prior Score
Suppose a hypothetical fitted equation predicts a quiz score in points from weekly study hours and a prior assessment score in points:
Here \(x_1\) is study hours per week, \(x_2\) is prior assessment score in points, and \(\hat{y}\) is the fitted quiz score in points. The equation is assumed to describe an additive plane over the predictor values being considered.
Study-hours direction. The fitted slope is \(b_1=2.4\). Its units are quiz-score points per study hour per week. At a fixed prior score, the fitted quiz score rises by 2.4 points for each additional weekly study hour. The held-fixed condition matters: this interpretation compares directions on the plane where prior score does not change.
Prior-score direction. The fitted slope is \(b_2=0.30\), measured in quiz-score points per prior-score point. At a fixed number of weekly study hours, the fitted quiz score rises by 0.30 points per additional prior assessment point.
Intercept. The fitted intercept is 18 points. It is the fitted quiz score at zero study hours and zero prior score. If those values are outside the range of the data or are not realistic, the intercept is a mathematical reference rather than a useful description of a typical student.
Both slopes are positive, so the plane rises in each predictor direction when the other predictor is fixed. These are descriptions of fitted associations. The equation by itself does not establish that increasing study time or a prior score causes quiz scores to rise.
Why “Holding Fixed” Produces the Slope
The adjusted interpretation follows directly from the fitted equation. Fix \(x_2\) at a value \(c\). The equation, considered only as \(x_1\) varies, becomes
This is a straight line in \(x_1\), with slope \(b_1\). The other predictor changes the line’s vertical position through \(b_2c\), but not its slope. Similarly, fixing \(x_1\) gives a straight line in \(x_2\) with slope \(b_2\). This explains both the meaning of “holding fixed” and why the additive plane has a constant slope in each direction.
The justification depends on the form of the fitted equation: there is a separate term for each predictor and no term that makes one predictor’s slope depend on the other. Thus, a fitted two-predictor plane should not be interpreted as allowing the slope to change from one value of the other predictor to another.
Worked Example: A Negative Adjusted Slope
A hypothetical fitted plane predicts delivery time in minutes from route distance in kilometers and the number of traffic lights on the route:
Here \(x_1\) is distance in kilometers, and \(x_2\) is the number of traffic lights. The fitted response is in minutes. Suppose this plane is used only for predictor combinations in the range represented by the data.
Distance. The coefficient \(b_1=4.1\) has units of minutes per kilometer. For a fixed number of traffic lights, the fitted delivery time increases by 4.1 minutes per additional kilometer. This reads the positive direction of the plane along the distance axis.
Traffic lights. The coefficient \(b_2=-0.8\) has units of minutes per traffic light. At a fixed route distance, the fitted delivery time decreases by 0.8 minutes per additional traffic light. This negative adjusted association may seem surprising, but its interpretation is determined by the fitted equation and the fixed-distance condition; it is not a claim that adding a traffic light makes a delivery faster.
The intercept, 12 minutes, is the fitted time at zero distance and zero traffic lights. Since a route of zero distance may not be part of the data or a meaningful delivery, that intercept need not have a practical interpretation. The two directional slopes remain readable without treating the intercept as realistic.
Units and Zero Slopes
Always attach the response and predictor units to a slope. If a response is measured in dollars and a predictor in hours, the slope is in dollars per hour. A coefficient is not simply “dollars” or “hours”; it relates the units of the response to those of one predictor. The two slopes in a plane usually have different units because they correspond to different predictors.
A zero coefficient has a precise, limited meaning: in the fitted plane, there is no slope in that predictor’s direction while the other predictor is held fixed. It does not say that the outcome is always the same, that the other predictor has no association, or that the predictor is unimportant in every possible model. Nor does a zero estimated slope by itself settle whether a population coefficient is zero; that is an inferential question.
Worked Example: A Level Direction in the Plane
Suppose a hypothetical model predicts monthly energy use in kilowatt-hours from home floor area in square meters and the number of occupants:
Here \(x_1\) is floor area and \(x_2\) is the number of occupants. The fitted response is monthly energy use in kilowatt-hours. The hypothetical fitted equation has \(b_1=1.6\) kilowatt-hours per square meter and \(b_2=0\) kilowatt-hours per occupant.
Floor-area direction. At a fixed number of occupants, the fitted energy use rises by 1.6 kilowatt-hours per additional square meter.
Occupant direction. At a fixed floor area, the fitted plane has zero slope for the number of occupants. In this fitted equation, changing the occupant count does not change the fitted value when floor area is fixed. This describes the plane as estimated; it does not establish that occupancy has no relationship with energy use in the population or in a different model.
The intercept of 95 kilowatt-hours is the fitted response at zero floor area and zero occupants. Since an occupied home cannot have zero floor area, it is unlikely to be a meaningful real-world prediction. The slope interpretations do not depend on giving the intercept a realistic interpretation.
Common Mistakes and Practical Limits
A fitted plane summarizes the pattern represented by its equation. The adjusted slopes are conditional associations in that model, not causal effects. If the data are observational, other explanations for an association may remain. Even in a designed study, a causal conclusion requires a suitable design and supporting assumptions beyond merely fitting a plane.
The directions are also most useful within the range of predictor values represented by the data. A plane is mathematically defined for any \(x_1,x_2\), but interpreting its slope far beyond observed values can be misleading if the relationship does not continue there. In addition, the constant-slope interpretation belongs to the additive plane shown here; it should not be assumed for every possible regression model.
- Wrong: “The coefficient of \(X_1\) is the change in \(Y\) for a one-unit increase in \(X_1\), whatever happens to \(X_2\).” Correction: State that \(X_2\) is held fixed.
- Wrong: “A positive adjusted slope proves that increasing the predictor causes the response to increase.” Correction: The fitted slope describes an adjusted association, not causation by itself.
- Wrong: “A zero slope means the predictor is unrelated to the response in every sense.” Correction: It means only that the fitted plane is level in that predictor’s direction while the other predictor is fixed.
- Wrong: “The intercept must be a sensible prediction.” Correction: It refers to the response at zero for both predictors, which may not be a relevant setting.
Check Your Understanding
For each question, state the fixed-predictor condition explicitly.
- A fitted plane is \(\hat{y}=7-1.2x_1+3.5x_2\), where \(Y\) is measured in kilograms, \(X_1\) in centimeters, and \(X_2\) in years. Interpret both slopes with units.
- In the equation \(\hat{y}=40+0x_1+2x_2\), what does the zero coefficient say about the fitted plane, and what does it not establish?
- Why is the coefficient of \(X_1\) an adjusted slope rather than a description of every observed change in \(Y\) as \(X_1\) changes?
- What does the intercept represent in a two-predictor fitted plane? Give one reason it might not be practically meaningful.
- Does a positive adjusted slope alone show that changing a predictor causes the response to change? Explain briefly.