A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
The F-statistic F = (ESS/k) / (RSS/(n-k-1)) tests H₀: all slopes equal zero; it follows an F(k, n-k-1) distribution under the null. High F values indicate the model explains significant variation, though this does not imply causal effects.
The F-statistic for overall model significance answers a deceptively simple question: does this regression model explain anything at all? You have built a normal linear regression model with k regressors, and you want to know whether those regressors collectively have any explanatory power. The null hypothesis is maximally skeptical: H₀ says that every slope coefficient equals zero simultaneously — meaning all those regressors are jointly useless. The F-statistic is a formal measure of how much evidence the data provide against this skeptical null.
To understand the formula intuitively, think about how variation is partitioned. Total variation in your outcome (TSS) splits into two pieces: variation explained by your model (ESS, explained sum of squares) and variation left unexplained (RSS, residual sum of squares). If the model is worthless, ESS should be near zero and RSS should be nearly equal to TSS. The F-statistic is essentially a ratio of average explained variation to average unexplained variation: F = (ESS/k) / (RSS/(n-k-1)). The denominators k and (n-k-1) are degrees of freedom — they adjust for the fact that adding regressors mechanically improves fit even when those regressors are garbage. A model with many predictors and a modest R² might have a low F, while a lean model with fewer, more relevant predictors can have a high F.
Under H₀ (all slopes are truly zero), this ratio follows an F(k, n-k-1) distribution. A large observed F-value means your data are far into the right tail of that distribution — unlikely to arise if the null were true. You compare your computed F to critical values from the F-distribution, or look at the p-value, to decide whether to reject H₀. This connects directly to your prior work on the F-test for joint significance: the overall model F-test is just a special case where you're jointly testing that every slope equals zero at once.
Two important caveats complete the picture. First, a statistically significant F does not tell you which individual coefficients matter — some regressors may be doing all the work while others add nothing. That question requires individual t-tests. Second, and more importantly, a high F-statistic says nothing about whether the regression estimates causal effects. A model that uses zip codes and household income to predict house prices will have an enormous F-statistic, but none of that association implies that giving someone a richer zip code would raise their house price. The F-test is about statistical explanatory power, not identification of causal mechanisms.