Chain convergence and successful inference highly dependent on initial values


Hi everyone,
I am currently testing my model against simulated data. Due to “identifiability” issues and divergence caused due to correlated parameters I have fixed 3 of 7 parameters. The model combines two datasets.

Now I am running into the problem that it is highly dependent on the initial values if a chain converges towards the “true” value and doesn’t get stuck in a low probability area of the parameter space. I am using a init function which generates random initial values based on the model constraints.

However, the inference works if the initial values are fixed exactly or close to values used to simulate the data.

What could cause this initial value dependency? Any ideas?