Write the hypothesis down before you measure it

The first attempt was a guess, and it was wrong

Vehicles per household went into a feature table because it sounded like a disability proxy: two of the six census disability questions concern mobility, so vehicle access seemed a reasonable stand in. It moved disability by 0.0096, well under the threshold. It moved race, marital status and citizenship instead, and accuracy went up, so every metric said ship it.

The failure is kept in the register as a control. A register containing only the hypotheses that worked would be a results table.

Two numbers per hypothesis

Alone asks whether there is anything in the column at all. Adds asks the question that actually matters: how much more of the attribute would enter the model if this column were added to the feature table. A column can score well alone and add nothing, because the features already present carry the same information.

The baseline rebuilds disability at 0.7321 from the deployed model's own features before any candidate is added, and the threshold is 0.0222. That threshold is not chosen for this page: it is the same one the exposure check fires on, four times the measured spread across refits with a floor. Using a looser one here would fill the register with findings the check downstream would then ignore.

The strongest evidence had no statistics in it

The candidate that fires hardest after the union is Medicaid, and the reason is not a correlation anyone reported. It is the wording of the question that collects it:

ACS questionnaire, health insurance coverage, item d:

  Medicaid, Medical Assistance, or any government-assistance plan
  for those with low incomes or a disability

A respondent living with a disability is being asked, in the same sentence, whether they are the kind of person the programme is for. The association is built into the instrument. No study is needed to justify it, only the questionnaire, and that is a far more durable source than an effect size from one paper.

The mirror image is Medicare, which had a good source and is silent anyway. The eligibility rule is real, but the features already present carry most of what it knows, so adding it would move very little. A published association is a hypothesis about your warehouse, never a fact about it.

Sources, verbatim in the register: IPUMS, health insurance variables in the ACS, SSA, Medicare and disability, Census, how ACS collects disability.

The register is the input, not the write up

research/proxies.yml is read by tools/screen.py. Nothing on this page was typed by hand: the numbers come from the run and the reasons come from the file.

- id: medicaid_question_names_disability
  claim: >
    Medicaid coverage identifies disability because the survey question that
    collects it names disability in its own text, so the association is built
    into the instrument rather than observed after the fact.
  column: medicaid_coverage_flag
  pums: HINS4
  strength: definitional
  source:
    title: Health Insurance Variables in the American Community Survey
    publisher: IPUMS USA
    url: https://usa.ipums.org/usa/acs_healthins.shtml
    quote: >
      Medicaid, Medical Assistance, or any government-assistance plan for
      those with low incomes or a disability

Two entries could not be tested because the columns are not in this warehouse, and they stay in the file. An untested hypothesis is not a refuted one, and deleting it would quietly turn a gap in the data into a finding.