Hello, I am implementing a GLM with a monotonic link growth function, because for my data a generalised sigmoid function, includes all the possible dynamics interesting to me. ( @Bob_Carpenter this is a dummy example that isolates the issue, on real data I am trying applying this link function in …

Martin discusses this problem in his blog post on identifiability and divergences. [image] Blog post: Identifying non-identifiability Publicity Hi all, I wrote a blog post, surveying some broken models in hope it helps people spot, uncover and understand issues with thei…

Thanks a lot, [image] stijn: intuitively it makes sense for me to have a prior that says if w is not very large, it is unlikely that b is very large I will try. ( @martinmodrak any news about your experience ? )

With this method I managed to resolve the banana shape, however I still have linear correlation, and it seems to result in divergent transitions. Anything more I could look at? [image] data { int<lower = 0> G; // all genes int<lower = 0> T; // tube …

I don’t really have anything to add beyond what is in the StanCon submission . The problem with slope = 0 is already partilly mitigated with reparametrizing via mean and sd of the input to the sigmoid (X[t] * beta in your case) - see the submission for the math. However, there might be situations whe…

[image] martinmodrak: reparametrizing via mean and sd of the input to the sigmoid ( X[t] * beta in your case) - see the submission for the math I have to admit that this reparametrization is not clear to me. Could you make an example for my model, or expand on this? [image] martinmodrak: …

[image] stemangiola: P.S. At the moment I am trying a 2 step solution, where I detect non-identifiable genes (i.e., slope ~ 0), and I fix the intercept to 0 for those genes, and rerun the whole model with more meaningful “tricks free” hierarchical priors That’s then basically identical to m…

I realized I was probably a bit wrong: the non-identifiability with different signs of beta probably does not apply to your case and arises only because I handle time series and solve the differential equation over time, linear models should be identifiable in this respect.

Thanks for this. This issue discouraged me, in favour of my “two step process”. I will attempt your trick

[image] martinmodrak: vector[G] mean_regulatory_input; I don’t see this parameter anywhere [image] martinmodrak: intercept = observed_mean_input - beta * observed_mean_input; Is this supposed to be like this?

[image] stemangiola: vector[G] mean_regulatory_input; I don’t see this parameter anywhere [image] martinmodrak: intercept = observed_mean_input - beta * observed_mean_input; Is this supposed to be like this? Sorry, there is a mistake, it should be: intercept = mean_regulatory_input -…

GLM with growth function - problem of indetermiinability when slope = 0

Modeling

stemangiola May 19, 2018, 8:31am 3

Thanks a lot,

I will try.

(@martinmodrak any news about your experience ? )

Fitting a non-linear growth model with stan

Topic		Replies	Views
Blog post: Identifying non-identifiability Publicity	11	5316	June 6, 2018
Many regressions: unique (stable) function for exponential and s-shape behaviour Modeling	21	3173	June 12, 2018
Difficulties with logistic population growth model Modeling fitting-issues	5	3009	October 28, 2019
Fitting generalized logistic function Modeling fitting-issues	8	3065	October 2, 2017
Fitting an unidentified linear model Modeling	16	1490	February 25, 2022

GLM with growth function - problem of indetermiinability when slope = 0

Related topics