tag:blogger.com,1999:blog-8547374028803716084.post6286301319191281848..comments2024-03-27T01:44:55.692-07:00Comments on Modern Statistical Workflow: A few simple reparameterizationsjavagehttp://www.blogger.com/profile/16427725069671113980noreply@blogger.comBlogger4125tag:blogger.com,1999:blog-8547374028803716084.post-54521189143999950412017-07-21T10:17:34.633-07:002017-07-21T10:17:34.633-07:00I don't think it's that strong. One of the...I don't think it's that strong. One of the things I do to get an idea of how the degrees of freedom parameter affects the distribution is to export the lkn_corr_rng() function to R (using expose_stan_functions()) and then draw some random numbers. lkj(4) isn't very tight. But yes, nothing to stop you from using a looser hyperprior. javagehttps://www.blogger.com/profile/16427725069671113980noreply@blogger.comtag:blogger.com,1999:blog-8547374028803716084.post-51919387655753940152017-07-20T17:28:27.065-07:002017-07-20T17:28:27.065-07:00Isnt lkj_corr(4) too strong? i.e. you might be for...Isnt lkj_corr(4) too strong? i.e. you might be forcing it to be a diagonal covariance unwittingly. Maybe add a uniform prior over that parameter between 1 and 4?Anonymoushttps://www.blogger.com/profile/18376557599002699434noreply@blogger.comtag:blogger.com,1999:blog-8547374028803716084.post-82392832590454978042017-07-18T08:39:13.245-07:002017-07-18T08:39:13.245-07:00Thanks Ben -- updated. Thanks Ben -- updated. javagehttps://www.blogger.com/profile/16427725069671113980noreply@blogger.comtag:blogger.com,1999:blog-8547374028803716084.post-86995902673026437102017-07-17T16:28:55.539-07:002017-07-17T16:28:55.539-07:00Not: Theta[n] = mu + diag_matrix(tau)*L_Omega * z[...Not: Theta[n] = mu + diag_matrix(tau)*L_Omega * z[n];<br />Rather: Theta[n] = mu + diag_pre_multiply(tau, L_Omega) * z[n];<br />But actually diag_pre_multiply(tau, L_Omega) should be executed just once before the loop.Anonymoushttps://www.blogger.com/profile/13622001562201260948noreply@blogger.com