I have a simple Tweedie model in GLM, or $Y_i=\beta_0+\beta_1x_1$. As usual, $Y$ is the output and $x_i$ is its dependent variable. Suppose that $p>1$, my link function would be $\frac{\mu^{1-p}}{1-p}$.

I thought this would be a simple exercise of $\frac{\mu^{1-p}}{1-p}=\beta_0+\beta_1x_1$. However, I can't solve for $\mu$ since I can't take natural log of a negative value. I'm wondering if anyone ran into this situation before.

  • 1
    $\begingroup$ At which point are you taking the logarithm of a negative value? $\endgroup$ – Sextus Empiricus Jul 8 at 20:29
  • $\begingroup$ In my model, all $\beta$s are positives. My $p$ is 1.5 since $p>1$. After I moved $1-p$ to the right side, I realized I cannot take $ln$ of both side. $\endgroup$ – tkhu Jul 8 at 21:00
  • $\begingroup$ Why are all $\beta$s positive? If the left side is negative due to that $1-p$ term then just make the right side negative as well. If you don't like that, then get rid of the $1-p$ term. $\endgroup$ – Sextus Empiricus Jul 9 at 1:19
  • $\begingroup$ My $\beta$s were positive because that's what the rxGLM summary results gave me. Turns out the package also got rid of $1-p$ on the left too, so I think $1-p$ is already baked into the estimated values. $\endgroup$ – tkhu Jul 9 at 2:56

I was able to answer my own question. I was using rxGLM in R with Tweedie distribution. The code indicates that "Family-link: Tweedie-mu^-0.5 ". Thus, the link function is actually $\mu^{-0.5}$.

It is then simply $\mu^{-0.5}=β_0+β_1x_1$. Flipping side gives $\mu^{0.5}=\frac{1}{β_0+β_1x_1}$. Finally the answer is $\mu=\frac{1}{(β_0+β_1x_1)^2}$. I checked this by hand and it works!

| cite | improve this answer | |

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.