Consider the Weibull sampling model for $X_1,\ldots,X_n$ iid, where $$p(x|\lambda,k)=k\lambda^kx^{k-1}e^{-\lambda^kx^k}$$ for $x>0$. Assume $k$ is known and $\lambda$ is unknown. First, if I adopt a $\text{Gamma}(a,b)$ on $\lambda$, this is not a conjuage prior for the Weibull sampling model, since the leading posterior has the term $\lambda^k$ on the exponential, so the posterior is not a Gamma distribution. I'm thinking of if I reparametrize $\theta=\lambda^k$, whether I can find a conjugate model for $\theta$. Then the sampling model will be $p(x|\theta,k)=k\theta x^{k-1}e^{-\theta x^k}$. But I don't know how to proceed. Something that I have in mind currently is that the reparametrized sampling model can be written as an exponential family $h(x)c(\theta)e^{\theta t(x)}$, with $h(x)=kx^{k-1}$, $c(\theta)=\theta$ and $t(x)=-x^k$. I'd like to know how I can derive the conjugate prior for $\theta$. Thank you.


1 Answer 1


It is more usual to take a conjugate prior for $\theta= \lambda^{-k}$. So the Weibull likelihood becomes $$\frac{k}\theta x^{k-1}e^{-x^k/\theta}$$ and the prior for $\theta$ is an inverse gamma distribution with density $$\frac{\beta^\alpha}{\Gamma(\alpha)}\frac{1}{\theta^{\alpha+1}} e^{-\beta/\theta}$$ which with observations $x_1,x_2, \ldots, x_n$ gives a posterior density proportional (after dropping multiplicative terms not involving $\theta$) to

$$\frac{1}{\theta^{\alpha+1}} e^{-\beta/\theta} \times \frac1{\theta^n}e^{-\sum x_i^k /\theta} = \frac{1}{\theta^{\alpha+n+1}} e^{-(\beta+\sum x_i^k)/\theta} $$

i.e. with $\alpha$ becoming $\alpha+n$ and with $\beta$ becoming $\beta+\sum\limits_1^n x_i^k$. You can translate this back to $\lambda$ if you must.

  • $\begingroup$ Hi @Henry, I'd like to know how to derive it. Can you include more details? Does your $\alpha$ and $\beta$ correspond to $a$ and $b$ in my question? $\endgroup$ Commented Oct 30, 2022 at 18:40
  • $\begingroup$ @PseudodifferentialOperators You do not have an $a$ or $b$ in your question. I had some typos in my answer and have expanded it; note your Weibell distribution has a reciprocal $\lambda$ to Wikipedia's - I have stuck with yours $\endgroup$
    – Henry
    Commented Oct 31, 2022 at 2:32
  • $\begingroup$ I don't take the reciprocal of $\lambda$ for each of computation $\endgroup$ Commented Oct 31, 2022 at 3:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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