I am reading Comparing Open and Sealed Bid Auctions: Evidence From Timber Auctions (working paper version) and struggling with getting a log likelihood function about a Gamma-Weibull distribution. In particular, the simple version of the distribution used in paper (p.20 Eq (7)) is:


where $\lambda(X,N,n) = \exp(X\beta)$, $\rho(n)=\exp(\gamma n)$ and $u$ has a Gamma distribution with unit mean and variance $\theta$. While I can understand the log-likelihood without $u$, the log-likelihood for auction $t$ would be the following with $u$ (simplified version of p.32):

$$\ln L_t = n\ln\theta + \ln\Gamma\left(\frac{1}{\theta}+n\right)-\ln\Gamma\left(\frac{1}{\theta}\right) + \sum_{i=1}^n\ln\left(\rho_{it}\lambda_{it}\left(\frac{b_{it}}{\lambda_{it}}\right)^{\rho_{it}-1}\right) + \left(\frac{1}{\theta} + n\right)\ln\left(1+\theta\sum^n_{i=1}\left(\frac{b_{it}}{\lambda_{it}}\right)^{\rho_{it}}\right)$$

How do I derive this equation?


1 Answer 1


I was not able to find the result. Here is my attempt.

The random term $u$ is not observed, and so it is not possible to rely on the conditional density $$g(b|X,u,N,n)$$ to derive the likelihood function whose parameters have to be estimated using observed explanatory variables only. The following unconditional likelihood function is of interest: $$L(b|X,N,n)= \int_0^\infty g(b|X,u,N,n)h(u|X,N,n) du,$$ with $h$ being the density of $u$. We obtain the Weibull density for $b$ from the cdf by $g=G'$ which yields $$ g(b|X,u,N,n)= \frac{u\rho(n)}{\lambda(X,N,n)}\left(\frac{b}{\lambda(X,N,n)}\right)^{\rho(n)-1} \exp\left(-u\left(\frac{b}{\lambda(X,N,n)}\right)^{\rho(n)}\right). $$ For the random term $u$, the authors consider a "Gamma distribution with unit mean and variance $\theta$, and which is independent of $X$, $N$, and $n$", that is $$ h(u|X,N,n)=h(u)= \left(\frac{1}{\theta}\right)^{1/\theta} u^{1/\theta-1} \frac{\exp(- u/\theta)}{\Gamma(1/\theta)}. $$ Once inserted into the likelihood function we have (omitting the conditioning variables): \begin{align*} L(b|X,N,n) = & \frac{\rho}{\lambda}\left(\frac{b}{\lambda}\right)^{\rho-1} \left(\frac{1}{\theta}\right)^{1/\theta} \frac{1}{\Gamma(1/\theta)} \cdot \int_0^\infty \exp\left[-u\left(\left(\frac{b}{\lambda}\right)^{\rho}+\frac{1}{\theta}\right)\right] u^{1/\theta-1} du \\ = & \frac{\rho}{\lambda}\left(\frac{b}{\lambda}\right)^{\rho-1} \left(\frac{1}{\theta}\right)^{1/\theta} \frac{1}{\Gamma(1/\theta)} \cdot \frac{\Gamma(1/\theta)}{ \left(\left(\frac{b}{\lambda}\right)^{\rho}+\frac{1}{\theta}\right)^{1/\theta}} \\ = & \frac{\rho}{\lambda}\left(\frac{b}{\lambda}\right)^{\rho-1} \left(\frac{1}{\theta}\right)^{1/\theta} \cdot \left(\left(\frac{b}{\lambda}\right)^{\rho}+\frac{1}{\theta}\right)^{-1/\theta}. \end{align*} The second equality is achieved after use of the following result: \begin{align*} & \frac{\Gamma(\beta+1)}{\alpha^{\beta+1}} = \int_0^\infty \exp(-\alpha u)u^\beta du. \\ \end{align*} For one observation, the log-likelihood is: \begin{align*} \ln L(b|X,N,n) &= \ln \left( \rho/\lambda (b/ \lambda)^{\rho-1} \right) - 1/\theta\ln \theta - 1/\theta\ln( (b/\lambda)^{\rho} + 1/\theta) \\ &= \ln \left( \rho/\lambda (b/ \lambda)^{\rho-1} \right) - 1/\theta\ln \theta - 1/\theta \left( \ln( 1+ \theta(b/\lambda)^{\rho} ) - \ln\theta \right) \\ &= \ln \left( \rho/\lambda (b/ \lambda)^{\rho-1} \right) - 1/\theta \ln( 1+ \theta(b/\lambda)^{\rho} ). \end{align*}

Summation over $n$ yields the sample log-likelihood. There are obvious differences between this expression and the one expected to be found... Where is the error ?


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.