First time that I ask a question on this platform! Here I go... I have a dataset with two random variables X1 and X2 and an output Y which comes from a discrete Weibull distribution. I've been trying to build a Metropolis-Hasting sampler to sample the following regression parameters for a discrete Weibull regression (reference: https://www.tandfonline.com/doi/abs/10.1080/02664763.2017.1342782?journalCode=cjas20): $$\theta = (\theta_0, \theta1, \theta2)$$ $$\gamma = (\gamma_0, \gamma1, \gamma2)$$ Both theta and gamma have a Laplace prior with two additional parameters lambda and tau which each also have a prior distribution (hyperprior) being the InvGamma(shape = 2, rate = 1) distribution. $$\lambda, \tau$$

My questions:
1. If I want to use a multivariate normal distribution as my proposal distribution, should I include all 8 parameters in this distribution? Generating random draws with a good acceptance rate seems hard...
2. I've set a covariance matrix with all diagonal elements equal to 0.001 to start off. Once the chain starts running, I update this covariance matrix by the covariance of the parameters drawn from the proposal distribution. Would this be a correct approach to update my stepsize while the MCMC chain is running? Or are there better ways?
3. When looking at the acceptance probability: $$ \alpha = min(1, \frac{l(x,y |\pi_{prop}) p(\pi_{prop})g(\pi_{curr}|\pi_{prop}))}{l(x,y |\pi_{curr}) p(\pi_{curr}) g(\pi_{prop}|\pi_{curr})} )$$ (with pi: all parameters) Would it be correct to use the exp(log(...)) trick to have log likelihoods and avoid likelihoods of zero (underflow)? With non-informative priors and for a symmetric proposal distribution, I believe the prior and g() probabilities would cancel out?


Your Answer

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

Browse other questions tagged or ask your own question.