Can multiple R be negative?

Its conventional definition is nonnegative; individual regression coefficients and pairwise correlations can still be negative.

Does R = 0.8 mean 80% of variance is explained?

No. In the usual least-squares setting, R squared is 0.64, corresponding to 64% of centered variation accounted for in sample.

Does a higher training R guarantee a better model?

No. Additional predictors can improve training fit while worsening generalization or increasing complexity.

AI Glossary / Statistics and probability / Intermediate

Coefficient of multiple correlation

Definition

The coefficient of multiple correlation, usually written R, measures the strength of the linear relationship between one outcome and its best linear prediction from a set of predictors. In ordinary least-squares regression fitted with an intercept, it is the nonnegative square root of the in-sample coefficient of determination R squared. It summarizes the predictors jointly rather than averaging their separate correlations with the outcome.

What it means

Several predictors can carry overlapping or complementary information. A model combines them into fitted values, and multiple correlation describes how those fitted values align with the observed outcome. Unlike a signed correlation between two variables, multiple R is conventionally nonnegative. It does not assign one overall positive or negative direction to a model containing several effects. Its square, not R itself, gives the familiar proportion of centered outcome variation accounted for by the fitted least-squares model under the usual conditions.

How it works

Fit the linear regression with an intercept on the chosen observations. Calculate the total sum of squared deviations of the outcome from its mean and the residual sum of squares. Then R squared = 1 - residual sum of squares/total sum of squares, and R is its nonnegative square root. For nonconstant fitted values, R also equals the sample correlation between observed and fitted outcomes. With a nonsingular predictor correlation matrix R_xx and outcome-correlation vector c, R squared can be written c-transpose R_xx-inverse c. These identities depend on the stated least-squares setting; they should not be transferred indiscriminately to penalized fits, models without intercepts or test-set scores.

Three practical examples

Forecast development: summarize the joint in-sample linear association between demand and several measured predictors.
Research reporting: distinguish the overall association of a predictor set from individual regression coefficients and partial correlations.
Model comparison: examine how fit changes when a predictor is added, while separately assessing overfitting and held-out performance.

Think of it this way

Think of several clues used together to reconstruct a picture. Two clues may reveal much of the same area, while another fills a different gap. Multiple correlation measures how closely the combined reconstruction follows the observed picture; adding individual clue scores would double-count shared information.

Worked example

Suppose two standardized predictors have correlations 0.6 and 0.5 with an outcome and correlation 0.3 with each other. For this valid correlation structure, R squared = [0.6 squared + 0.5 squared - 2(0.3)(0.6)(0.5)]/[1 - 0.3 squared] = 0.43/0.91 = 0.4725 approximately. Multiple R is the square root, about 0.6874. Thus the linear predictor accounts for about 47.25% of outcome variance in this specified correlation setting, not 68.74%. Adding the individual squared correlations would give 0.61, which incorrectly ignores their shared information. The result says nothing by itself about causal effects, the cost of prediction errors or performance after deployment. If these correlations are sample estimates, uncertainty and possible overfitting also matter.

Illustrative case study

Fictional case: Maple Planning fits a demand model using recent orders, promotions and calendar indicators. Managers initially describe multiple R as the percentage of demand predicted correctly. The analyst explains that neither R nor R squared is a classification accuracy measure, and supplies error measures in demand units alongside the regression summary. A separate later-period evaluation tests whether the relationship generalizes. The team also inspects whether promotions and recent orders carry redundant information. This invented case illustrates interpretation and validation, not a measured result for an actual forecasting product.

Common misconceptions and mistakes

Multiple R is not the mean of predictor-outcome correlations and cannot identify which predictor matters most. A high value does not prove causation, absence of bias or good calibration. Adding predictors to an ordinary least-squares fit on the same observations cannot lower the training R squared, even if the new predictors are useless outside the sample. Adjusted R squared is a different statistic and can be negative. Test-set R squared can also be negative and is not generally the square of the observed-predicted correlation. A constant outcome makes the usual centered variance ratio undefined.

Business use

Businesses can use multiple correlation as a compact description of linear fit, provided it accompanies measures that answer operational questions. Report prediction errors in meaningful units, evaluate on appropriately separated data and examine performance where decisions matter. Compare models on the same outcome and observations when interpreting changes in training R. Missing-data handling can otherwise change the sample and invalidate a simple comparison. For AI systems with nonlinear predictions, do not assume that a correlation between predictions and observations inherits every least-squares regression identity.

Check your understanding

Why does the worked example not add 0.36 and 0.25 to obtain the joint explained fraction?

Answer: The predictors are correlated, so their relationships with the outcome overlap. The multiple-correlation formula accounts for the predictor correlation of 0.3 and yields approximately 0.4725 rather than 0.61.

Related terms

Coefficient of determination
Partial correlation
Multicollinearity
Overfitting
Validation set
Multiple linear regression
Pearson correlation
semipartial correlation
adjusted R squared
residual sum of squares

Learn next

distinguish R, R squared and adjusted R squared, then derive the least-squares decomposition with an intercept. Examine how correlated predictors share explanatory information and why individual coefficients can be unstable despite strong overall fit. Finally, compare training summaries with held-out error and calibration measures that reflect the intended decision.

Sources

Penn State STAT 462, The Multiple Linear Regression Model: https://online.stat.psu.edu/stat462/node/131/ Penn State STAT 501, Multiple Linear Regression: https://online.stat.psu.edu/stat501/Lesson05 Penn State STAT 800, Regression: https://online.stat.psu.edu/stat800/Lesson08

Previous: Coefficient of determination Next: Coefficient of variation