Model weight of mixture components

[martinmodrak]

So if I understand you correctly, your problem could actually be stated as 4 completely separate inference problems:

1. Estimate the proportion of data points that go into each component - since you have 3 components, this is IMHO well handled directly fitting an intercept-only multinomial model on a single-row dataset with just the total counts of data points in each component (see Example with family "multinomial" for a simple example)

2. For each component separately fit the parameters of the lognormal.

If this is right, i.e. if you see no way how those separate problems could share information, you can just fit 4 separate models (your current solution fits 3 completely independent means, but one shared sigma - this might or might not be desirable).

@GianniRewer

I do not think you can assign a point to a component when fitting mixture models, am I correct?

You are correct, mixture models are designed for the case where you don’t know which component the data point belongs to.