3
$\begingroup$

How do I compute the standard error if I only know the p-value, the beta (i.e., the regression coefficient from a linear regression), the sample size, and the number of regression parameters?

$\endgroup$
  • $\begingroup$ What does "beta" represent here? I'm guessing regression coefficients? $\endgroup$ – COOLSerdash Mar 27 '18 at 16:06
  • $\begingroup$ Yes it does! Regression coefficients from a linear regression. $\endgroup$ – Abdel Mar 27 '18 at 16:06
  • $\begingroup$ Then please edit your question and add this information. Short answer: No, you need the residual degrees of freedom (or sample size and number of regression parameters). Just the coefficient and the p-value is not enough. $\endgroup$ – COOLSerdash Mar 27 '18 at 16:08
  • $\begingroup$ Thank you! I will edit the question, as I actually do have the sample size, just not in the file I was looking at :) Thank you! $\endgroup$ – Abdel Mar 27 '18 at 16:09
  • $\begingroup$ As I said: The sample size alone is still not enough: You need to know how many variables were included in the regression models. $\endgroup$ – COOLSerdash Mar 27 '18 at 16:13
6
$\begingroup$

In a linear regression, the $p$-value is calculated from a $t$-value, which is the coefficient divided by its standard error ($t=\hat{\beta}/\mathrm{SE}_{\hat{\beta}}$). The degrees of freedom used in the $t$-distribution for calculating the $p$-value are the residual degrees of freedom ($\mathrm{SE}_{\hat{\beta}}=\hat{\beta}/|t|$). The residual degrees of freedom, on the other hand are the total degrees of freedom of the variance $N-1$ minus the model degrees of freedom $p-1$, where $p$ is the number of parameters including the intercept. So the residual degrees of freedom are $(N-1)-(p-1) = N-p$.

From this, you can use the quantile distribution of the $t$-distribution to calculate the standard error. Example: Assume that $\hat{\beta}=5.47, p = 0.004, N = 100, p = 4$. The residual degrees of freedom are $100-4 = 96$. We assume that the $p$-value is two-sided.

Using R, the calculations are:

t_val <- qt(0.004/2, df = 96) # Calculating the t-value using quantile function
5.47/abs(t_val) # Calculating standard error
1.854659

So the standard error was 1.85.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.