Can adjusted R-squared increase when a predictor is added?

Yes, if the reduction in residual error is sufficient to offset the degrees-of-freedom adjustment.

Can adjusted R-squared be negative?

Yes. Its variance-ratio formula allows negative values; those values are not negative squared correlations.

Does adjusted correlation remove confounding?

No. This sample-size and complexity adjustment does not identify causal effects or substitute for choosing an appropriate causal adjustment strategy.

AI Glossary / Statistics and probability / Intermediate

Adjusted correlation ( r adj )

Definition

Adjusted correlation, in the source glossary's regression sense, is a correlation measure corrected for the number of predictors relative to the sample size. It addresses the tendency of a fitted multiple correlation to look stronger when more predictors are fitted. The related adjusted R-squared is the quantity most commonly reported.

What it means

A regression can fit its training data better simply because it has more flexibility. Ordinary in-sample R-squared cannot decrease when predictors are added to a nested ordinary least-squares model fitted to the same observations with an intercept. That mechanical improvement does not establish better prediction on new data.

The adjustment incorporates residual degrees of freedom. It is different from a partial correlation, which measures an association after controlling for specified variables. Because "adjusted correlation" can have other meanings in other settings, identify the correction and formula whenever reporting it.

How it works

For an ordinary least-squares regression with an intercept, let n be the number of observations and k the number of fitted predictor coefficients excluding the intercept. A standard formula is adjusted R-squared = 1 - (1 - R-squared)(n - 1)/(n - k - 1). The residual degrees of freedom must be positive, and the usual total variation must be nonzero.

If an adjusted multiple correlation is required and adjusted R-squared is nonnegative, its nonnegative square root can be reported, with that convention stated. Multiple correlation does not carry the positive or negative direction of an individual regression slope.

Adjusted R-squared can be negative. Despite its name and notation, it is not literally the square of a real-valued correlation in that situation. Do not silently take an invalid square root or replace the result with an apparently ordinary correlation. Also, count fitted coefficients correctly: a categorical predictor can require several coefficients, and polynomial terms add parameters.

Three practical examples

1. An analyst compares two sales regressions fitted to the same records, checking whether a small improvement in raw R-squared survives the penalty for additional coefficients.

2. A reporting team distinguishes adjusted multiple correlation from a correlation adjusted for age, preventing two different statistical operations from sharing an ambiguous label.

3. A model reviewer checks sample size and parameter counts before interpreting a very high training correlation from a model containing many predictors.

Think of it this way

Think of awarding points for how well a model fits while charging for each extra adjustable setting. Adding a setting may improve the fit, but the improvement has to offset its cost. The charge discourages superficial improvements; it does not prove that the resulting model will work well in a different situation.

Worked example

Consider an original regression example with n = 50 observations, k = 4 predictor coefficients and R-squared = 0.64. Adjusted R-squared is 1 - 0.36 times 49/45 = 1 - 0.392 = 0.608. The raw multiple correlation is the square root of 0.64, or 0.80; the square root of the adjusted value is approximately 0.780.

Now add a fifth predictor coefficient using the same 50 observations. Suppose R-squared rises to 0.645. Adjusted R-squared becomes 1 - 0.355 times 49/44, approximately 0.6047. The raw fit improved, but the adjusted measure fell because the improvement was too small relative to the added parameter.

For a separate weak-fit example, n = 20, k = 3 and R-squared = 0.05 produce 1 - 0.95 times 19/16 = -0.128125. Report that negative adjusted R-squared directly; it does not imply a negative directional correlation between predictions and outcomes.

Illustrative case study

Fictional case study: Forecastlane evaluates a regression with several extra customer attributes. Its training R-squared rises slightly, and an early presentation describes the expansion as a clear improvement. A reviewer asks for the sample size, fitted coefficient count and adjusted R-squared.

The adjustment falls, prompting the analysts to inspect whether the added complexity is justified. They also compare predictions on held-out observations and assess whether the extra fields will be reliably available in operation. Their final report separates training fit, complexity adjustment and predictive validation.

This fictional case illustrates a useful diagnostic sequence. It does not establish that the model with the highest adjusted R-squared is always the best operational choice.

Common misconceptions and mistakes

Adjusted correlation is not automatically a partial correlation or a correction for measurement error. Those are different operations requiring different information. Adjusted R-squared is also not a probability that the model is correct, a causal effect, or an exact guarantee of future explanatory performance.

Comparisons can become misleading when models use different observations because missing values change the sample. Searching through many candidate models can still overfit the selection process even when adjusted R-squared is used. The ordinary formula should not be transferred uncritically to penalized models, generalized linear models or other settings with different effective degrees of freedom.

Business use

The adjustment can improve the discipline of regression reporting by making complexity visible. Show raw and adjusted fit measures alongside the sample size and model specification, especially when comparing nested models on the same outcome and observations.

Use validation aligned with the intended deployment, such as later periods for forecasting, and examine prediction errors in relevant units. Check residual patterns and data quality rather than reducing the assessment to one statistic. A modestly lower adjusted measure may accompany a model preferred for interpretability or operational reliability, while a higher measure may still describe an impractical or unstable model.

Check your understanding

Two nested models use the same observations. Raw R-squared increases but adjusted R-squared decreases after adding a coefficient. What does that indicate?

Answer: The added coefficient improved training fit, but not enough to compensate for the standard complexity adjustment. Additional evidence is needed to decide whether the expanded model is useful.

Related terms

Coefficient of determination
Partial correlation
Overfitting
multiple correlation
adjusted R-squared
residual degrees of freedom
cross-validation and model selection

Learn next

calculate raw and adjusted R-squared for nested regressions on an identical sample. Then compare their held-out prediction errors and explain why the rankings can differ. State whether the parameter count includes the intercept.

R-squared

Sources

Primary sources: Howell, Statistical Methods for Psychology glossary, Adjusted correlation: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

Penn State, STAT 462, The Multiple Linear Regression Model: https://online.stat.psu.edu/stat462/node/131/

Penn State, STAT 501, Model Building: https://online.stat.psu.edu/stat501/Lesson10

Previous: Adjoint state method Next: Adjusted means