A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
The F-test evaluates whether a set of coefficients is jointly statistically significant, testing the null hypothesis that all slope coefficients equal zero simultaneously. The overall F-statistic compares the explained variance in the restricted model (intercept only) to the full model; individual t-tests cannot perform this joint test without inflating Type I error. F-tests also apply to linear restrictions — for instance, testing whether two coefficients are equal. The F-statistic follows an F-distribution with (q, n−k−1) degrees of freedom, where q is the number of restrictions being tested.
Suppose you estimate a regression with ten explanatory variables and find that seven of them have t-statistics below 2 — individually, they appear statistically insignificant. Should you conclude the model has no explanatory power and drop all seven? Not necessarily. From your hypothesis testing background, you know that each t-test has a false positive rate (Type I error). If you run ten separate tests, each at the 5% level, the probability that at least one falsely rejects the null is much higher than 5%. The F-test solves the complementary problem: it tests whether variables are collectively significant in a single, unified null hypothesis.
The F-test for overall significance asks: is the full model (with all predictors) meaningfully better than the restricted model that includes only an intercept? In other words, do any of the slope coefficients differ from zero? The F-statistic compares the variance explained by the full model to what remains unexplained, adjusted for degrees of freedom. Formally, it equals (RSS_restricted − RSS_unrestricted)/q ÷ RSS_unrestricted/(n−k−1), where q is the number of restrictions tested (the number of slope coefficients you're testing jointly) and n−k−1 are the degrees of freedom of the full model. A large F-statistic means that adding the predictors reduced unexplained variance substantially — more than would be expected by chance.
The most powerful application of this logic appears when regressors are correlated (multicollinear). Imagine you're predicting exam performance using both hours studied and hours of tutoring. These two variables are highly correlated — students who study more tend to get more tutoring. In OLS, multicollinearity inflates the standard errors of individual coefficients, making each look separately insignificant. Yet together, they clearly explain a lot. The F-test picks this up: it asks whether the joint contribution of both variables is significant, not whether each stands alone. This is why you can have a high F-statistic (significant overall model) alongside individually insignificant t-statistics — and why dropping all the individually insignificant variables would be a mistake.
The F-test also generalizes beyond overall model significance to testing linear restrictions. The same framework applies when you want to test whether two coefficients are equal (for instance, whether the return to an extra year of education is the same for men and women), or whether a subset of variables can be jointly dropped. In each case, you compare the restricted model (imposing the constraint) to the unrestricted model (without the constraint). The F-statistic follows an F-distribution with (q, n−k−1) degrees of freedom under the null, where q is the number of restrictions. If you're only testing a single restriction (q=1), the F-statistic equals the square of the t-statistic — a useful sanity check connecting F-tests back to the t-tests you already know.