Extremely large person errors for some rows using non_census_synthesis

[werdnabae]

I am getting extremely large errors using the sample data (hh_marginals.csv, household_sample.csv, person_marginals.csv, person_sample.csv) and I'm generating the synthetic population using the non_census_synthesis notebook. The generated households match the marginals very well, but the persons are not matched well at all.

In this picture, I calculate the percent difference between the synthesized and actual marginals. As you can see many of the differences are very large.

I've also tried generating synthesis using my own queried data, and I'm having the same problem with the person distributions not matching well.