Disaggregation y


I’m somewhat new in bayes analysis and wondering if there is a reasonable approach for following problem.

Let’s say:

  • I have an outcome for an upper level (e.g. states of us).
  • But I i’d like to have this outcome for a lower level (e.g. counties inside every state).
  • I also have some characteristics (features) for the counties.
    X1[county] … Xn[county]

I saw kind of an expected maximization approach.

  • Generate a starting condition Y[state]/len(counties) for every state and now process EM steps a model Y[county] = X1[county] … Xn[county] and rescale sum(Y[county]) == Y[state].

Is this reasonable and I’m wondering if there is a possibility to try something with pymc3?

Thanks for a starting point.

Yes, you can do this kind of EM approach in PyMC3, here is an example:

Note that this is currently not very easy to do due to the API design.

1 Like