Hi, I get two warnings from cmdstan: 38 of 16000 (0.24%) divergent transitions, E-BFMI = 0.27 is smaller than 0.3. Is it ok to publish with these results? If not, how can I fix this. Is there a way to reparametrize the model, as the warnings suggest? Thanks! Code gaussian_mixture.stan // G…

Maybe a non-centered parameterization of the mixture components works better? // Guassian mixture model data { int<lower=1> N; // number of data points real y[N]; // observed values real uncertainties[N]; // uncertainties } parameters { real<lower=0,upper=1> p; // mixin…

Not references sorry. It’s the same principle as the uncentered parametrization and one of the tricks people recommend here for autoregressive models. The differences are sometimes more independent than the means and reparametrizarion in terms of differences can help Stan. It did not work in this ca…

[image] Max_Mantei: Can’t make anything of the results though… Do they make sense? Yes, thank you very much! With your reparametrisation, the model runs without any warnings now. This issue is solved! :) I had to make a little change to your code. I changed from this: for (n in 1:N) { t…

Hmmm, but I’m guessing that your uncertainties is the standard deviation from the mean and that y is a draw from a normal distribution with a known standard deviation but an unknown mean, isn’t that right? If so, what you did the first time seems correct to me because it takes into account the uncer…

[image] emiruz: Hmmm, but I’m guessing that your uncertainties is the standard deviation from the mean and that y is a draw from a normal distribution with a known standard deviation but an unknown mean, isn’t that right? Yes. [image] emiruz: If so, what you did the first time seems corr…

If I’m not mistaken, by eye it looks like the original has a higher variance at every data point. I suppose it’s down to does it fit your use case and whether the extra variance is important. It’s worth trying to run your original model for more iterations and increasing adapt_delta to .99, which I …

[image] emiruz: If I’m not mistaken, by eye it looks like the original has a higher variance at every data point. That’s good point. In order to check that, I’ve made a test by simulating observations that have 5x larger measurement uncertainties, and ran the models. The original and @Max_Man…

[image] Evgenii: I assumed it was a typo. The way I found the issue is that the code gave unrealistically small spreads of 0.01 for mu1, mu2 and sigma parameters. With the “[n]” fix, the results are the same as with my original code (except that we have no warnings now). Good catch! Yes, this…

Thanks Max, I hadn’t realised that your reformulation was equivalent (but it is clearly) so I learned something useful too!

[image] Max_Mantei: Good catch! Yes, this is what I meant to write. Glad to hear that the results makes sense after your correction! :) No problem, do we also need uncertainties[n] instead of uncertainties, in the loop? This is what I used. [model]

That’s correct, yes. Man I feel a bit stupid. Sorry, I answered way too quickly. With everything indexed correctly, you probably have to up to number of iterations and the adapt_delta. Luckily, you don’t have a huge data set, so that should not be a big problem. I got zero divergent transitions with…

How to reparameterize a Gaussian mixture model?

Modeling

Evgenii March 2, 2020, 4:14am 33

Update: @avehtari suggested an improvement to the model that makes sampling even better.

Topic		Replies	Views
Request for help understanding model behavior Modeling	17	1686	March 24, 2018
Trouble with Gaussian Mixture Model Modeling	10	4369	December 7, 2018
Divergent transitions & BFMI low in a state-space model Modeling	6	1093	August 17, 2017
Problems remain after non-centered parametrization of correlated parameters Modeling	34	4423	July 3, 2019
Mixture Models Modeling fitting-issues	9	2361	September 16, 2020

How to reparameterize a Gaussian mixture model?

Related topics