Like other power-law-type distributions, it gets more diffuse as the exponent is lowered. Below quadratic (exponent 2), it has infinite variance, so that’s pretty darn diffuse.

I just followed the BUGS manual. There’s also Wolfram: Pareto Distribution and Wikipedia: Pareto Distribution.

]]>What I can’t do with Brendan et al.’s RTE data is get a good estimate of the beta priors if I use moment-matching inside of EM; likelihood diverges with zero-variance beta priors. The “right” thing to do is a better estimate of the beta priors based on another layer of priors; that’s what I did in the paper and in R/BUGS (reparameterized, I used Beta(1,1) on the beta mean (alpha/(alpha+beta)), and Pareto(1.5) on the beta scale (alpha+beta). It’s easy when BUGS does all the work, and I may just need to implement a proper beta estimator with priors.

I’m in the middle of writing up a longer blog post with more details.

]]>