线性模型中系数子集为零的适宜检验方法咨询

Great question! When you're testing whether a subset of regression coefficients are jointly zero (instead of all predictor coefficients like the global F-test you mentioned), the standard and appropriate method is the restricted F-test (also commonly referred to as the partial F-test). Here's a breakdown of how to apply it to your specific model:

Step 1: Fit the two key models #

• Unrestricted Model: This is your full original model where no coefficients are constrained:
  $$ y = \beta_0 + \beta_1x_1 + \beta_2x_2 + \beta_3x_3 + \beta_4x_4 $$
  Calculate its residual sum of squares (denoted $RSS_u$) and degrees of freedom $df_u = n - k$, where $n$ is your sample size and $k$ is the number of parameters in the full model (here, $k=5$, so $df_u = n-5$).

• Restricted Model: This is the model that enforces your null hypothesis $H_0': \beta_1 = \beta_4 = 0$, so you drop $x_1$ and $x_4$ from the model:
  $$ y = \beta_0 + \beta_2x_2 + \beta_3x_3 $$
  Calculate its residual sum of squares ($RSS_r$) and degrees of freedom $df_r = n - k'$, where $k'$ is the number of parameters in the restricted model (here, $k'=3$, so $df_r = n-3$).

Step 2: Compute the F-statistic #

The F-statistic measures how much worse the restricted model fits compared to the full model, adjusted for the number of constraints. The formula is:
$$ F = \frac{(RSS_r - RSS_u)/(df_r - df_u)}{RSS_u/df_u} $$
For your hypothesis, $df_r - df_u = (n-3)-(n-5) = 2$ (the number of coefficients you're testing as zero), so this simplifies to:
$$ F = \frac{(RSS_r - RSS_u)/2}{RSS_u/(n-5)} $$

Step 3: Evaluate the test result #

• Compare the calculated F-statistic to the critical value from an F-distribution with numerator degrees of freedom $2$ and denominator degrees of freedom $n-5$.
• Alternatively, calculate the p-value associated with your F-statistic. If the p-value is less than your chosen significance level (e.g., 0.05), you reject $H_0'$, meaning at least one of $\beta_1$ or $\beta_4$ is non-zero—so $x_1$ and $x_4$ jointly have a significant effect on $y$.

Key Note #

This test assumes the standard OLS regression assumptions hold: normally distributed errors, constant variance (homoscedasticity), and independent observations. If these assumptions are violated, you may need to use robust or alternative testing methods, but for standard linear regression, the restricted F-test is the gold standard for this scenario.

内容的提问来源于stack exchange，提问作者Frank Vel