Joint Significance

Can predictors be statistically insignificant individually but jointly significant, and what tests do you use? Give two reasons why this can happen.

Answer

Yes they can be.

Tests:

  • Individual significance: use \(t\)-tests for each coefficient \(H_0:\beta_j=0\).

  • Joint significance: use an \(F\)-test for \(H_0:\beta_1=\beta_2=\cdots=\beta_p=0\).

Why can this happen?

  • Multicollinearity: Predictors are highly correlated, so each coefficient’s standard error is large \(\Rightarrow\) \(t\)-stats are small hence large p-values (statistical insignificance). But together they explain variation in \(Y\), so the joint \(F\)-test can still be large.

  • Shared signal spread across predictors: The true effect is distributed across several predictors; each one alone looks weak, but combined they matter.

Back to collection