Why not both?

]]>https://www.washingtonpost.com/news/rampage/wp/2017/09/28/free-speech-and-good-vs-bad-polls/

Andrew Gelman, a professor of statistics and political science at Columbia University who is arguably the statistics field’s biggest public intellectual.

And I thought you were the biggest methodological terrorist.

]]>However, if you do something like (1 + x1 + x2 | g), then there is a 3×3 covariance matrix to estimate. Stan developers have long encouraged decomposing a covariance matrix into a correlation matrix and standard deviations and more recently be encouraging people to decompose the correlation matrix into its Cholesky factors. But people are still putting independent (often half-Cauchy, which is dubious) priors on the standard deviations, whereas **rstanarm** has always put a Dirichlet priors on the proportions of the unknown trace of the covariance matrix. By setting the concentration hyperparameter of the Dirichlet distribution to some number greater than 1, you can encourage the variances to be similar to each other. The unknown trace is set equal to the size of the matrix multiplied by the square of a scale parameter, which has a Gamma prior.

]]>