Multiple Regression: Adding Predictors to Your Model

Multiple regression extends simple linear regression to more than one predictor. Instead of predicting an outcome from a single variable, you estimate how several predictors relate to the outcome at once, and, crucially, what each one contributes while the others are held constant. That last phrase is the whole reason multiple regression exists, and it's where most of the interpretation work lives.

This guide assumes you're comfortable with one predictor. If you're not, start with the guide to simple linear regression, since everything here builds on the slope, intercept, residuals, and R squared it covers. Below: the multiple regression equation, how to interpret a coefficient with other variables controlled, why adjusted R squared replaces R squared, how to add categorical predictors, how to spot multicollinearity, and how to report it all.

From One Predictor to Several

Simple regression fits a line: one predictor, one slope. Multiple regression fits a plane or a higher-dimensional surface: several predictors, each with its own slope, sharing one intercept. The equation just grows a term for each predictor.

ŷ = b ₀ + b ₁x ₁ + b ₂x ₂ + ... + b _kx _k

The reason to do this is rarely just better prediction. It's that real outcomes have several causes, and you usually want to know the effect of one predictor after accounting for the others. Does advertising still predict sales once you account for how many salespeople a store has? Does a treatment still predict recovery once you control for age? Multiple regression is the tool for that kind of question, and the answer lives in what each coefficient means.

The Key Idea: Holding Other Variables Constant

In simple regression, a slope is the change in y per one-unit change in x. In multiple regression, each slope is a partial effect: the change in y per one-unit increase in that predictor, with every other predictor held constant. This is the single most important idea in the whole method, and the phrase "holding other variables constant" (or "controlling for") should appear whenever you interpret a coefficient.

A worked example makes it concrete. Suppose you're predicting monthly sales (in thousands of dollars) from advertising spend (in thousands) and the number of sales staff. Fit advertising on its own first, then add staff.

Look at what happened to advertising's coefficient. On its own, each additional thousand dollars of advertising was associated with 3.67 thousand dollars more in sales. Once staff enters the model, that drops to 2.35. The reason is that stores which advertise more also tend to employ more staff, so the advertising-only model was quietly crediting advertising with some of what staff was actually doing. Controlling for staff strips that out. The 2.35 is advertising's effect at a fixed level of staffing, which is a more honest estimate of advertising's own contribution.

This is why a coefficient can change, shrink, or even flip sign when you add a predictor. The number isn't wrong in either model; it's answering a different question. The simple-regression slope is the total association; the multiple-regression slope is the association net of the other predictors.

The Full Fitted Model

Here's the complete two-predictor model for that sales example, with all the pieces a write-up needs.

ŷ = 12.19 + 2.35(advertising) + 5.05(staff)

Read the coefficients in the same partial way throughout. Holding staff constant, each additional thousand dollars of advertising predicts 2.35 thousand dollars more in sales. Holding advertising constant, each additional staff member predicts 5.05 thousand dollars more. The intercept, 12.19, is the predicted sales when both predictors are zero, which here is an extrapolation outside the data and not worth interpreting substantively.

Both predictors are statistically significant in this model: advertising at t(7) = 9.67, p < .001, and staff at t(7) = 5.82, p < .001. Each significance test asks the same question simple regression asks, whether that predictor's partial slope differs from zero, using the hypothesis-testing logic of assuming no effect and measuring how surprising the result is.

Why Adjusted R Squared Replaces R Squared

In simple regression, R squared is the proportion of variance in the outcome explained by the model. That definition carries over, but multiple regression introduces a trap: R squared can only go up when you add a predictor, never down, even if the predictor is pure noise. Add enough junk predictors and R squared creeps toward 1 while the model gets worse, not better.

Adjusted R squared fixes this by penalizing the model for each predictor it includes. It rises only when a new predictor improves the fit by more than you'd expect from chance, and it can fall when a predictor doesn't earn its place. In the sales example, adding staff moved R squared from 0.976 to 0.996 and adjusted R squared from 0.973 to 0.995. Both rose, which tells you staff was a genuinely useful predictor rather than noise.

• Report R squared to describe how much variance the final model explains.
• Use adjusted R squared to compare models with different numbers of predictors, since it's the one that doesn't reward mere complexity.

Adding Categorical Predictors

Predictors don't have to be continuous. You can include categorical variables such as region, treatment group, or education level through dummy coding: each category becomes a 0/1 indicator, with one category left out as the reference. Each dummy's coefficient is then the difference between that category and the reference group, holding the other predictors constant.

For example, with a three-region predictor (North, South, West) and North as the reference, the model includes a South indicator and a West indicator. The South coefficient is the average difference in the outcome between South and North, controlling for everything else in the model. The mechanics of choosing reference categories and interpreting dummy coefficients are covered in the guide to the independent variable, which walks through dummy coding with a real dataset.

Multicollinearity: The Main New Risk

The biggest problem unique to multiple regression is multicollinearity: when two or more predictors are strongly correlated with each other. When predictors overlap heavily, the model struggles to separate their individual effects, because the data can't tell which of the correlated predictors deserves the credit. The coefficients become unstable, their standard errors inflate, and small changes in the data can swing them wildly.

In the sales example, advertising and staff correlate at 0.93, which is high. That's why advertising's coefficient shifted so much between models, and it's a warning sign worth checking directly rather than eyeballing. The standard diagnostic is the variance inflation factor (VIF), computed for each predictor.

• VIF near 1 means the predictor is essentially uncorrelated with the others. No problem.
• VIF above 5 is a common flag for moderate multicollinearity worth investigating. The sales predictors sit around 7.
• VIF above 10 is widely treated as serious multicollinearity that needs action.

When multicollinearity is severe, the options include dropping one of the redundant predictors, combining them into a single index, collecting more data, or accepting that the individual coefficients can't be cleanly separated and interpreting them jointly. What you should not do is report tightly correlated predictors' individual coefficients as if each were a clean, independent effect.

Assumptions: What Changes from Simple Regression

The four assumptions of simple linear regression, linearity, independence, normality of residuals, and equal variance, all carry over to multiple regression unchanged. You still check residual plots, still want a shapeless band of residuals against predicted values, still care most about residual normality in small samples.

Multiple regression adds one consideration on top: the absence of severe multicollinearity among the predictors, checked with VIF as above. It also makes the residual checks a little more involved, since you plot residuals against each predictor as well as against the fitted values, looking for any leftover pattern a predictor failed to capture. The visualization guide covers the residual and scatter plots this requires.

Reading Multiple Regression Output

Software returns a coefficient table plus overall model statistics. The table has one row per predictor; the model statistics describe the fit as a whole.

The standardized beta is worth a note. Because the raw coefficients are in each predictor's own units (dollars, staff count, years), you can't compare their sizes directly. The standardized betas put every predictor on a common scale, so a larger absolute beta means a stronger relative contribution. That comparison is one of the things reviewers most often want from a multiple regression.

Reporting Multiple Regression in APA Format

Report the overall model first, then the individual predictors. A complete write-up of the sales example reads:

A multiple regression predicted monthly sales from advertising spend and sales staff. The model was significant, F(2, 7) = 833.68, p < .001, explaining 99.6% of the variance ( R² = .996, adjusted R² = .995). Both predictors contributed: advertising, b = 2.35, t(7) = 9.67, p < .001; staff, b = 5.05, t(7) = 5.82, p < .001.

The conventions match simple regression: give F with its two degrees of freedom for the model, report each predictor's unstandardized coefficient with its own t test, italicize statistical symbols, and interpret each coefficient in plain language with the "holding constant" qualifier in the surrounding text. For a model with several predictors, a coefficient table is usually clearer than prose, with columns for b, its standard error, the standardized beta, t, and p.

Common Mistakes

• Interpreting a coefficient without "holding constant." Every multiple-regression slope is a partial effect. Dropping the qualifier turns a correct statement into a misleading one.
• Comparing models with R squared instead of adjusted R squared. R squared always rises with more predictors. Only adjusted R squared tells you whether the additions earned their place.
• Ignoring multicollinearity. Strongly correlated predictors produce unstable coefficients. Check VIF, and don't report overlapping predictors' individual effects as if each were clean.
• Reading a nonsignificant predictor as proof of no effect. A predictor can be nonsignificant because it truly doesn't matter, or because it overlaps with another predictor, or because the sample is too small. Nonsignificance is not proof of zero effect.
• Throwing in every available variable. More predictors is not better. Each one costs a degree of freedom and risks overfitting, especially in small samples. Choose predictors from theory, not convenience.
• Overfitting with too few observations. A rough guideline is at least 10 to 15 observations per predictor. Fitting many predictors on a small sample produces a model that fits the sample and fails on new data.
• Claiming causation. Controlling for other variables strengthens an argument but does not, by itself, establish causation. The language stays associational unless the design supports more.

Where to Go Next

Multiple regression handles a continuous outcome. When your outcome is binary, such as pass or fail, purchased or not, the appropriate model is logistic regression, which applies the same predictor logic to a probability. When you need to compare group means rather than model a continuous predictor, analysis of variance is the natural tool.

For the foundations underneath this article, see the guides to simple linear regression and what the correlation coefficient measures. For the predictor terminology and dummy coding, see the guide to the dependent and independent variables.

When your analysis is written up, the results section is where a subject-matter editor earns their place, because a reviewer who spots a coefficient interpreted without its controls or a model compared on the wrong statistic will question the rest. Editor World's academic editing, dissertation editing, and journal article editing services include review of statistical reporting, and you choose your own editor by field so the person reading your regression output knows your discipline's conventions.

Frequently Asked Questions

Page last reviewed: July 2026. Content reviewed by Editor World editorial staff. Editor World, founded in 2010 by Patti Fisher, PhD, graduate of The Ohio State University, provides professional editing and proofreading services for academic researchers, doctoral candidates, faculty, business professionals, and authors worldwide. 100% human editing, no AI at any stage. BBB A+ accredited since 2010 with 5.0/5 Google Reviews and 5.0/5 Facebook Reviews. More than 100 million words edited for over 8,000 clients in 65+ countries. Multiple Gold and Bronze Stevie Award winner. Native English editors from the United States, the United Kingdom, and Canada. Less than 5% of applicants are accepted to the editor panel. Recommended by the Boston University Economics Department, University of San Diego, University of Michigan, UCLA, University of Missouri, and more.