# Why p-value of coefficients of each variable are insignificant but the overall F is significant is a indicator of multicollinearity?

Today I read a document about multicollinearity here.

In the first page, the author runs a simple regression that

$$y = \beta_1x1+ \beta_2x2$$

the p-value( or t-statistic) of coefficients of $$x1$$ and $$x2$$ are insignificant but the overall F-test is significant. And he said that it is an indicator of multicollinearity.

I do not understand what is the reason for this conclusion, would you please help me to understand it intuitively?

Much appreciated.

• If $y = x_1+ x_2$ there is no need for a regression. Oct 27, 2021 at 8:25
• @Betrand sorry, I forget to add the coefficient, it is a regression Oct 27, 2021 at 8:40

## 1 Answer

The reasoning behind it is simple:

1. Multicolinearity distorts t-test and consequently p-values because it inflates standard errors of the $$\beta$$ coefficients. Since in multivariate regression the variance of estimated coefficients depends on correlation between independent variables.

2. F-test is not affected by multicolinearity.

3. the t-test tests whether coefficients are individually significant (with null $$\beta_i=0$$ and alternative hypothesis $$\beta_i\neq0$$

4. The F-test tests the null of joint insignificance $$\beta_1 = \beta_2 = … = \beta_k=0$$ against alternative that at least one of the coefficients is not equal to zero.

Given 1-4 it is very strange to observe situation where using F-test we cannot reject the hypothesis that at least one coefficient is non-zero, while individual t-test we cannot reject the hypothesis that both coefficients are individually equal to zero.

In that case results become a bit suspicious which warrants further investigation. One likely explanation is that there is a multicolinearity present because as mentioned above multicolinearity distorts t-tests but not F-test.

However, this itself is not a test for multicolinearity just common symptom. If you want to make sure that multicolinearity is present you should make further investigation by for example calculating variance inflation factors. VIF>5 (some sources say VIF>10 these values are just rule of thumbs) tell you that multicolinearity is indeed present.

• Thanks @1muflon1, from the calculation of p-value based on std.err (bmj.com/content/343/bmj.d2304), I know that the higher the std.err, the lower p-value. But I don't understand why multicollinearity cause the std.err inflation Oct 26, 2021 at 20:38
• and can I ask why small N is also an indicator of multicollinearity in this case if it is convenient to you Oct 26, 2021 at 20:47
• @Louise 1. Small N is not an indicator of multicolinearity. 2. It is because in formula for se in a multivariate regression there is also term $(1/(1-r_{x1x_2}^2$
– 1muflon1
Oct 27, 2021 at 18:19
• Thanks @1muflon1 Oct 27, 2021 at 19:47