Filter questions

What gets asked first, and who is routed past the rest — and why that routing decision shapes every number that follows.

The gate before the count

Every large-scale disability survey begins not with a direct question about disability but with a screen. A respondent who answers "no" to a preceding question about health or functioning is routed around the detailed module entirely. The people filtered out at that stage are not in the denominator of the finer questions. They are, in effect, invisible to the instrument.

The design choice is practical: asking dozens of functional questions of every respondent in a general household survey is expensive and exhausting. A filter keeps questionnaire length manageable. But it has a statistical consequence that is easy to overlook — the filter question's wording and placement determine who reaches the disability question at all.

In the Family Resources Survey, the sequence that feeds official disability statistics begins with questions about long-standing illness or impairment. Respondents who report none are not asked whether that condition substantially limits daily activity. The survey definition is therefore applied only to those who have already identified themselves as having some health condition — which means the limitation question can only catch people the filter has already admitted.

Change the filter and the population changes with it. Research comparing different filter phrasings — "do you have a health problem or disability?" versus "do you have a condition expected to last twelve months or more?" — produces meaningfully different groups even before the substantive disability question is posed. The wording decides the number at every stage, including this earliest one.

The filter also interacts with self-identification. Some people with conditions that substantially affect daily activity do not think of themselves as having a long-standing illness, and so answer "no" to the filter. They are routed out. Others with relatively mild conditions answer "yes" and proceed. The filter does not measure impairment; it measures how a respondent categorises their own health at the moment of answering. That is not a flaw so much as a feature of self-report methodology — but it is one that should be named when citing the resulting count.

Census disability questions face the same architecture at scale. Once a decade, near-complete coverage is achieved — but the filter logic embedded in the question order still shapes who the disability question reaches. Administrative data avoids filter questions structurally, because it counts records rather than surveying individuals, which introduces a different set of boundaries.

A filter question is infrastructure. Statisticians reading a bulletin who want to understand what a disability figure actually measures need to go back before the disability question — to the question that decided who was asked it.