Expressing Regression Significance Through R-Squared

Why R-Squared and Regression Significance Are Related

In simple linear regression, R-squared summarizes how much of the observed variation in the response is accounted for by the fitted line. The overall F test addresses a different question: whether the data provide evidence that the population slope is nonzero. The two quantities are connected by an exact formula, but they are not interchangeable. In particular, the same R-squared can lead to different test conclusions when sample sizes differ.

The connection is useful when a regression summary reports R-squared but not the F statistic. It also clarifies why a visually strong fit, measured by R-squared, is not by itself an inferential conclusion. The test uses both the amount of variation explained and the sample size, as well as a model and its assumptions.

Model, Sums of Squares, and R-Squared

Consider \(n\) observations \((x_i,Y_i)\) modeled by a simple linear regression with an intercept:

For the exact, finite-sample F test, assume the \(x_i\) values are fixed and not all equal, and that the errors are independent with \(\varepsilon_i\sim N(0,\sigma^2)\) for a common variance \(\sigma^2>0\). Equivalently, for random predictors, the same test is exact conditional on the observed predictor values when these error assumptions hold. The fitted values are \(\hat{Y}_i\), the response mean is \(\bar{Y}\), and the residuals are \(e_i=Y_i-\hat{Y}_i\).

R-squared is a descriptive measure of this sample’s fit. For a non-perfect fit with \(SST>0\), it lies between 0 and 1. If the response values are all identical, \(SST=0\) and R-squared is undefined. If \(SSE=0\), the fit is perfect and the usual finite F ratio has a zero denominator; the finite-statistic formula below presumes \(SSE>0\).

Deriving the F Statistic from R-Squared

The overall regression test for simple linear regression with an intercept tests \(H_0:\beta_1=0\) against \(H_A:\beta_1\ne 0\). Under the null, the predictor adds no linear slope to the intercept-only model. The overall F statistic compares the regression mean square, which has 1 numerator degree of freedom, with the residual mean square, which has \(n-2\) degrees of freedom.

The identity \(SST=SSR+SSE\) gives \(SSR=R^2SST\) and \(SSE=(1-R^2)SST\). Substitution shows why the total variation cancels and leaves a formula involving only \(n\) and \(R^2\):

The test is upper-tailed: larger values of \(F\) provide more evidence against \(H_0\). At significance level \(\alpha\), reject when the observed statistic exceeds the upper-tail critical value \(F_{1,n-2;1-\alpha}\), or equivalently when its upper-tail p-value is less than \(\alpha\). The degrees of freedom must accompany the statistic; the same numerical F value does not have the same reference distribution for every sample size.

The formula also works in reverse. Solving for R-squared gives \(R^2=F/(F+n-2)\). Thus, for fixed \(n\), larger R-squared means larger F. But a test decision also depends on the null distribution and its degrees of freedom. R-squared alone does not supply a significance level or a p-value.

Worked Applications

What R-Squared Does—and Does Not—Say

A value such as \(R^2=0.36\) says that, in the sample and under this fitted regression decomposition, 36% of the total variation in the observed response is accounted for by the fitted line. It does not say that the predictor caused 36% of the response, that predictions will be accurate for new observations, or that the slope is statistically significant. Those claims require different reasoning and, for significance, a specified inferential procedure.

The F test uses a model-based measure of variation relative to the residual variation, adjusted for degrees of freedom. Under its assumptions, its reference distribution lets us quantify how unusual the observed statistic would be if the slope were zero. If the error distribution or other assumptions are unsuitable, the exact F reference distribution may not apply. A high R-squared does not repair a violated assumption, and a low R-squared does not by itself prove that the slope is zero.

There is also a practical distinction. A small effect can be statistically significant in a large sample, while a potentially meaningful effect may not reach a chosen significance level in a small sample. Statistical significance concerns evidence against the specified null under the model; R-squared is a descriptive summary of in-sample fit. Neither quantity alone measures the practical importance of a slope.

Common Exam Traps

• Using the wrong degrees of freedom. In simple regression with an intercept, the overall F test has degrees of freedom \(1\) and \(n-2\), not \(1\) and \(n-1\).
• Applying the formula to a different model. The displayed identity uses one predictor and an intercept. Do not assume the same numerator degrees of freedom or formula for a model with multiple predictors or different restrictions.
• Calling R-squared a test. R-squared has no significance level or p-value by itself. Use the sample size to obtain F, then use the correct reference distribution.
• Interpreting significance as practical importance or causation. A significant slope is evidence against a zero slope under the model assumptions; it is not, on its own, proof of a large effect or a causal relationship.
• Forgetting the intercept condition. The sum-of-squares decomposition and formula here use regression with an intercept. Do not apply them automatically to a regression forced through the origin.