In March of 2024 the state legislature considered a bill that would have required counties to verify the immigration status of applicants for emergency rental assistance. I was working as a data analyst for a county housing agency at the time, and my director asked me a question I could not answer: if this passes, how many households do we lose?1
I spent two weeks trying. The agency had 11,000 assistance records. None of them contained immigration status, correctly and by design. What they did contain was the intake language, the household composition, and the census tract. I built an estimate by joining our records to the American Community Survey five-year tract-level tables on nativity and language spoken at home, then adjusted for the composition difference between our applicants and the tract population. My answer was that we would likely lose between 600 and 1,400 households, with the wide band reflecting how badly tract-level inference travels down to the household.2
My director presented the low end of that range in testimony. The bill did not pass, for reasons that had nothing to do with my number.3
I have thought about those two weeks more than any other work I have done.4 Three things bother me about them.
The first is that I do not know if the estimate was any good. I had no way to validate it. I had never been taught how to reason about the error in a small-area estimate, and I was reconstructing something from first principles that I now know is a well-developed literature with a name.5
The second is that the range mattered enormously and the range is exactly what gets dropped. A policy audience hears 600. I produced 600 to 1,400 and I did not know how to make the uncertainty survive the trip to a committee room.
The third is the one I keep returning to. The reason our records do not contain immigration status is a privacy decision made by people who were protecting our applicants, and it is the right decision. It is also the reason I could not answer the question, and being unable to answer the question was very nearly the thing that let the bill through. I do not think that tension resolves cleanly, and I would like to spend two years with people who have thought about it longer than I have.6
I am applying to the MPP with the quantitative methods concentration for a specific sequence of courses.7 Program evaluation and causal inference, because almost every question my agency asks is a causal question dressed as a descriptive one and I have been answering them descriptively. The small-area estimation and survey methods coursework, for the direct reason above. And the data privacy and public records seminar, which is the only course I have found at any program that treats the third problem as a problem rather than as a constraint to work around.8
I know what I want to do afterward and it is not ambitious in scope.9 I want to return to state or county housing policy in the Midwest, in an analytic role with enough seniority to decide which questions get asked rather than only answering them. The rental assistance programs built during the pandemic are being wound down or made permanent right now, state by state, mostly on the basis of evidence that does not exist yet.10 I have spent three years close enough to that decision to see what is missing from it.
I did the work with a five-year ACS table and two weeks. I would like to know what I could have done with the right training.11