I am new to Bayesian modeling and trying to “replicate” and compare the results of a hierarchical linear model in Pymc3 to the results fitting this same model in R (lmer). I believe I am properly setting up a case where there are fixed effects for the intercept, exper and hgc.9, there are random eff…

Im not sure, I essentially copied the approach from 13.22 of rethinking example

I did find one mistake in the code. But even with this correction, the estimates (especially for the correlations) are off (the first rho should be -0.3 ish not positive 0.4). I can get “correct” estimates using brms or rstanarm so I must be doing something wrong in terms of the pymc3 code? id_idx=…

Could you print the brm stan generated code?

Sure, it looks like this (I am not a stan programmer so I cant say that I follow it completely): Here is the R code: #brms wages <- read.table("https://stats.idre.ucla.edu/stat/r/examples/alda/data/wages_pp.txt", header=T, sep=",") #center wages$exper_c<-(wages$exper-mean(wages$exper)) wages$hgc.…

There are quite a few place the Stan model and the pymc3 model differ, but what makes the most differences for model convergence is usually the standardization of the predictor matrix: transformed data { int Kc = K - 1; matrix[N, K - 1] Xc; // centered version of X vector[K - 1] means_X; …

I center the two predictors in the code I pasted above: wages['hgc.9']=wages['hgc.9']-wages['hgc.9'].mean() wages['exper']=wages['exper']-wages['exper'].mean()

It seems really odd to me that any of the priors make much difference given the amount of data. Seeing completely flipped signs for the correlation component… From the standpoint of the code for pymc3 does this look correct?

Parameterization does make a huge difference, and programmatically generated model like the one from brm often apply many heuristics in preprocessing etc to make it work well.

Thanks. It seems like unless there is a mistake in the code, in a pretty simple case like this I should be able to get close to what comes from lmer, brm etc. Frustrating for us newbies :slight_smile:

Random Intercepts and Slopes (Correlated) Convergence Issues

Questions

lwahedi March 7, 2019, 3:57pm 2

There was a thread on something similar here. It may be helpful in troubleshooting.

Topic		Replies	Views
Cholesky decomposition and correlation among random effects Questions	20	5882	March 11, 2019
Bambi model correlated intercept slopes	2	117	April 10, 2025
Correlated slopes in multivariate model Questions	9	3550	July 14, 2018
Hierarchical logistic regression giving non-sensible results? Do I formulate it correctly? Questions	15	940	August 4, 2020
Multilevel model with covariance of 3 variables Questions	9	966	October 10, 2023

Random Intercepts and Slopes (Correlated) Convergence Issues

Related topics