Can multiple R be negative?

The conventional coefficient is nonnegative. Individual regression coefficients and ordinary pairwise correlations can be negative.

Is R the same as R squared?

No. R is the nonnegative square root, and the squared quantity has the familiar OLS variance-decomposition interpretation.

Does a high R establish a useful model?

Not alone. Error magnitude, calibration, validation, assumptions and decision requirements also matter.

AI Glossary / Statistics and probability / Intermediate

Multiple correlation coefficient

Definition

The multiple correlation coefficient, conventionally written R, measures how strongly an outcome is linearly associated with its best linear prediction from a set of predictors. In ordinary least-squares regression with an intercept, its sample version is the correlation between observed outcomes and their fitted values, when that correlation is defined. R is nonnegative and ranges from zero to one. Its square, R squared, is the corresponding coefficient of determination. Multiple correlation summarizes the joint linear prediction rather than the separate direction or importance of each predictor.

What it means

Several predictors can jointly relate to an outcome even when no single predictor tells the whole story. Multiple correlation evaluates the fitted combination of those predictors. A high value indicates close linear association between the outcome and its fitted prediction in the evaluated sample, but it does not establish causality or future predictive accuracy. The conventional in-sample OLS value is nonnegative because the fitting procedure chooses the prediction's direction. With one predictor and an intercept, R equals the absolute value of the ordinary Pearson correlation, rather than preserving its sign.

How it works

Fit an ordinary least-squares model with an intercept and the chosen predictors. Calculate fitted values and residuals. For a nonconstant outcome, the in-sample coefficient of determination is R squared = 1 - SSE/SST, where SSE is the residual sum of squares and SST is the sum of squared deviations from the outcome mean. Take the nonnegative square root to obtain R. When fitted values vary, directly correlating them with observed outcomes gives the same result. Distinguish this definition from out-of-sample correlation, adjusted R squared or models without an intercept, where familiar identities and interpretations may not apply.

Three practical examples

Forecast development: summarize how well a linear combination of available predictors fits a numerical outcome in development data.
Explanatory modeling: describe the joint linear association of several measured variables with an outcome, without treating the summary as causal evidence.
Model comparison: inspect changes in fit when a prespecified predictor is added, while also using appropriate complexity and validation checks. The coefficient alone cannot decide whether the added model complexity is useful.

Think of it this way

Think of several instruments playing together to reproduce a melody. Multiple correlation evaluates how closely the combined performance follows the target melody, not how loudly each instrument plays or whether one instrument caused the tune. A strong match during rehearsal also does not guarantee the same performance on a different piece.

Worked example

Use four artificial observations with predictors X1 = (-1,-1,1,1) and X2 = (-1,1,-1,1), and outcome Y = (-2,-2,0,4). Ordinary least squares with an intercept gives fitted values 2X1 + X2 = (-3,-1,1,3). The residuals are (1,-1,-1,1), orthogonal to both predictors and the intercept. The outcome mean is zero, so SST = 4 + 4 + 0 + 16 = 24. SSE = 1 + 1 + 1 + 1 = 4. Therefore R squared = 1 - 4/24 = 5/6, approximately 0.8333, and R = square root of 5/6, approximately 0.9129. This original calculation describes in-sample fit. It does not mean that 91.29% of predictions are correct, nor that 83.33% of the outcome has been causally explained.

Illustrative case study

In a fictional case, Cedar Planning reports a high multiple correlation for a demand model. Managers initially interpret it as evidence that the model will forecast accurately next quarter. Analysts explain that the value summarizes fit to the development sample and add a separate time-based validation. They also inspect residual size in demand units and compare with a simple baseline. The final report retains R as a useful summary but separates in-sample association, prediction error and future performance.

Common misconceptions and mistakes

R does not indicate whether an individual predictor has a positive or negative coefficient. It is not the average of pairwise correlations. Adding predictors to the same in-sample OLS model cannot reduce ordinary R squared, so an increase alone is not proof of improvement. High multiple correlation can coexist with unstable coefficients when predictors are strongly correlated. Out-of-sample R squared can be negative, and taking its square root does not produce the conventional R. If fitted values are constant, their Pearson correlation is undefined even though the variance-explained convention can assign R squared zero.

Business use

Businesses should report R or R squared with the model definition, sample and evaluation context. Include prediction errors in meaningful units, residual diagnostics and independent validation when forecasting is the objective. Use adjusted measures or a suitable model-selection procedure when complexity differs, while recognizing their assumptions. Avoid interpreting a strong fit as evidence that manipulating a predictor will change the outcome. The most useful report connects the fit summary to the model's intended decisions and makes clear what remains untested.

Check your understanding

A fitted OLS model has R squared = 0.81. What is its conventional multiple correlation coefficient?

Answer: R = 0.9, assuming the usual definition and valid model context.

Related terms

Learn next

Derive the OLS sum-of-squares decomposition and compare R with residual error in outcome units. Then examine how adding irrelevant predictors can improve training fit without improving predictions on new data.

Residual

Sources

David C. Howell, University of Vermont, multiple regression supplement and correlation of observed with predicted values: https://www.uvm.edu/~statdhtx/fundamentals9/Supplements/Chapter11R.html Penn State STAT 462, The Multiple Linear Regression Model: https://online.stat.psu.edu/stat462/node/131/

Previous: Multiple comparisons Next: Multiple correspondence analysis