R and R-Squared (R²) are two commonly reported statistics in regression analysis and data science. Although they are closely related, they answer different questions.

R describes the strength and direction of a linear relationship, while R² describes the proportion of variation in an outcome that is accounted for by a regression model.

Understanding the difference between R and R² is important when interpreting regression results, evaluating models, and communicating statistical findings.

In this guide, we will explain what R and R² mean, how they are calculated, how they differ, and how to interpret them in simple and multiple linear regression.

What Is R?

The symbol R can refer to different correlation quantities depending on the context.

For two variables, the Pearson correlation coefficient measures the strength and direction of their linear association.

Its value ranges from:

[
-1 \leq r \leq 1
]

where:

• +1 = perfect positive linear relationship
• 0 = no linear association
• −1 = perfect negative linear relationship

For clarity, lowercase r is commonly used for the sample Pearson correlation, while uppercase R is often used for the multiple correlation coefficient in regression.

Positive Correlation

If:

[
r=0.90
]

there is a strong positive linear association between the two variables.

As one variable tends to increase, the other also tends to increase.

Negative Correlation

If:

[
r=-0.85
]

there is a strong negative linear association.

As one variable increases, the other tends to decrease.

Correlation Near Zero

If:

[
r=0.05
]

there is very little linear association.

However, a correlation close to zero does not necessarily mean that the variables have no relationship. A strong nonlinear relationship can still produce a correlation near zero.

What Is R-Squared (R²)?

R², known as the coefficient of determination, measures the proportion of variation in the dependent variable that is accounted for by the regression model.

For a standard linear regression model with an intercept:

[
R^2 =
1-\frac{SS_{\text{Residual}}}{SS_{\text{Total}}}
]

where:

• (SS_{\text{Residual}}) = residual sum of squares
• (SS_{\text{Total}}) = total sum of squares

In simple linear regression with one predictor:

[
R^2=r^2
]

For example, if:

[
r=0.90
]

then:

[
R^2=(0.90)^2=0.81
]

So the model accounts for 81% of the variation in the response variable under the usual regression interpretation.

The remaining 19% represents variation not accounted for by the model. It should not automatically be interpreted as being caused by specific “other factors,” because it can include random variation, measurement error, omitted variables, and model misspecification.

R vs. R²: Key Differences

Simple Way to Remember

Think of R as answering:

How strongly are these variables linearly related, and in which direction?

Think of R² as answering:

How much of the variation in the outcome is accounted for by this regression model?

Example: Simple Linear Regression

Suppose a researcher wants to examine the relationship between hours studied and exam score.

Consider the following data:

After fitting a simple linear regression model, suppose the results are approximately:

[
r=0.99
]

and:

[
R^2=0.98
]

Interpreting r

An (r) value close to +1 indicates a very strong positive linear association between hours studied and exam score.

Interpreting R²

An (R^2) of 0.98 means that approximately 98% of the variation in exam scores is accounted for by the fitted linear relationship with hours studied in this sample.

This does not prove that studying causes higher scores. Causal conclusions require an appropriate study design and assumptions.

Why R² Is Not the Same as Accuracy

A common misconception is that an R² of 0.90 means that the model is “90% accurate.”

That is incorrect.

R² measures the proportion of variation accounted for by the model. It does not directly tell you the size of prediction errors.

For regression problems, it is often useful to examine additional metrics such as:

• MAE
• RMSE
• MSE
• MAPE, when appropriate
• Out-of-sample R²

For example, two models can have similar R² values but very different practical prediction errors.

R in Multiple Regression

When a regression model contains multiple predictors, the interpretation of R requires additional care.

For example:

[
Y=\beta_0+\beta_1X_1+\beta_2X_2+\beta_3X_3+\epsilon
]

In this setting, regression software may report Multiple R, which represents the correlation between the observed response values and the values predicted by the fitted model.

The relationship is:

[
R=\sqrt{R^2}
]

when the conventional nonnegative multiple correlation coefficient is being reported.

Therefore, the statement:

[
R^2=R\times R
]

is valid in this context, but R² should not be confused with the square of the correlation between the outcome and any one individual predictor.

Example: Multiple Linear Regression

Suppose exam scores are predicted using:

• Hours studied
• Previous grade
• Attendance

A fitted model might produce:

[
R^2=0.92
]

This means the model accounts for approximately 92% of the sample variation in exam scores.

The corresponding multiple correlation coefficient is:

[
R=\sqrt{0.92}
]

[
R\approx0.959
]

Unlike a simple Pearson correlation between two variables, this multiple R summarizes the relationship between the observed response and the model’s fitted values.

Adjusted R²

Adding predictors to a regression model can increase or leave unchanged the ordinary R², even when the additional variables provide little useful information.

This is one reason Adjusted R² is often reported in multiple regression.

Adjusted R² accounts for the number of predictors and the sample size.

A common formula is:

[
R^2_{\text{adj}}

1-(1-R^2)
\frac{n-1}{n-p-1}
]

where:

• (n) = number of observations
• (p) = number of predictors

Example

Suppose two models produce:

Although Model B has a higher R², its much lower Adjusted R² suggests that the additional predictors may not provide sufficient improvement relative to model complexity.

Adjusted R² should still not be treated as a universal model-selection criterion.

Does a Higher R² Always Mean a Better Model?

No.

A higher R² on training data does not necessarily mean better predictive performance.

A complex model can achieve a high R² by fitting noise in the training data.

This is particularly important in machine learning.

For predictive modeling, evaluate performance on data that were not used to fit the model, using approaches such as:

• Train-test splits
• Cross-validation
• External validation
• Out-of-sample R²
• MAE
• RMSE

Can R² Be Negative?

For ordinary least-squares regression evaluated on the same data used to fit the model, with an intercept, R² is generally between 0 and 1.

However, R² can be negative when evaluated on new or validation data.

A negative out-of-sample R² means the model predicts worse than a simple baseline that always predicts the mean of the training response, under the conventional definition.

This is an important distinction when discussing model performance.

High R² Does Not Mean Causation

A high R² indicates a strong statistical fit, not a causal relationship.

For example, ice cream sales and drowning incidents may both increase during warmer months.

A regression model could find a strong association between them, but that does not mean ice cream consumption causes drowning.

A third variable, such as temperature or season, can influence both.

Therefore:

[
\text{Association} \neq \text{Causation}
]

Causal conclusions require an appropriate research design or causal-inference framework.

Low R² Is Not Necessarily a Bad Result

A low R² does not automatically mean that a model is useless.

Some outcomes are inherently difficult to predict because they are influenced by many variables, measurement error, random variation, or complex human behavior.

In areas such as:

• Economics
• Psychology
• Social sciences
• Consumer behavior
• Healthcare

moderate or low R² values can occur even when predictors have meaningful relationships with the outcome.

The usefulness of a model should therefore be evaluated in context.

R and R² in R Programming

You can calculate Pearson correlation in R using:

hours <- c(1,2,3,4,5,6,7,8,9,10,11,12)

scores <- c(55,60,65,70,75,80,85,89,92,94,96,98)

cor(hours, scores)

To fit a linear regression model:

model <- lm(scores ~ hours)

summary(model)

Extract R²:

summary(model)$r.squared

Extract Adjusted R²:

summary(model)$adj.r.squared

The regression summary also provides coefficients, standard errors, t-statistics, p-values, residual information, and other diagnostics.

What Should You Look at Alongside R²?

R² should rarely be interpreted alone.

For a complete regression assessment, consider:

1. Adjusted R²

Useful when comparing models with different numbers of predictors.

2. RMSE

Measures the typical magnitude of prediction errors in the same units as the response variable.

3. MAE

Provides the mean absolute prediction error and is generally easier to interpret than squared-error metrics.

4. Residual Diagnostics

Check whether residuals show:

• Nonlinearity
• Heteroscedasticity
• Outliers
• Dependence
• Systematic patterns

5. Cross-Validation

Useful for estimating how well the model generalizes to unseen observations.

6. Confidence Intervals

Confidence intervals for coefficients provide information about the uncertainty associated with estimated effects.

7. Domain Knowledge

A statistically strong model may still be unsuitable if its assumptions or predictions do not make practical or scientific sense.

Common Misconceptions

Is R² Always the Square of R?

For simple linear regression with one predictor and an intercept:

[
R^2=r^2
]

In multiple regression, the reported multiple R satisfies:

[
R=\sqrt{R^2}
]

But R² is not simply the square of the Pearson correlation between the response and one predictor.

Is R² a Percentage Accuracy?

No.

An R² of 0.80 does not mean the model is 80% accurate.

It means that, under the standard in-sample interpretation, approximately 80% of the variation in the response is accounted for by the fitted regression model.

Is a Higher R² Always Better?

Not necessarily.

A higher training R² can result from adding unnecessary predictors or overfitting.

Is There a “Good” R² Value?

There is no universal threshold.

A useful R² depends on:

• Research field
• Data quality
• Measurement error
• Outcome variability
• Prediction objective
• Cost of prediction errors
• Model purpose

A model with R² = 0.40 can be useful in one application, while R² = 0.90 may be inadequate in another.

R vs R²: Quick Summary

Conclusion

R and R² are related but fundamentally different statistics.

The correlation coefficient describes the strength and direction of a linear association, while R² describes the proportion of variation in the response accounted for by a regression model.

In simple linear regression:

[
R^2=r^2
]

But regression analysis should not stop at R².

A reliable evaluation should also consider Adjusted R², residual diagnostics, MAE, RMSE, confidence intervals, cross-validation, and out-of-sample performance.

Most importantly, neither a high correlation nor a high R² establishes causation.

Understanding these distinctions helps analysts and machine learning practitioners interpret regression results more accurately and avoid one of the most common mistakes in statistical modeling: treating a single model metric as the complete story.