,
14 tweets,
3 min read
Read on Twitter
This is temporarily good, but not decisive. We will now have a chance to force Secretary Ross to publicly justify his decision, but it could still come back before the 2020 census. HOWEVER... for purely statistical reasons, they should have ruled more decisively. Thread:
1/ Consider a basic principle of survey design: the sequence you ask questions matters. The way you frame questions matters. Adding one question sometimes changes the way people answer another.
2/ Market researchers and pollsters know that and work very hard to ensure that the structure of questionnaire doesn't impact the results. (And pollsters seeking to get skewed outcomes know how to use this to their advantage).
3/ The Census bureau also knows that and as a matter of good practice they generally don't make wholesale revisions to the census in a single year.
4/ In the last census there were some spot tests of internet data collection. Those tests were done for small subsets of census takers to learn how changing format affected results. Having tested in a small sample, that's now being rolled out at a larger scale 10 yrs later.
5/ That is the statistically sound and responsible way to change the census. Do it on a small scale that doesn't impact overall results in one period, see how that sample compares with others, then (if appropriate) roll out next time to factor in lessons learned.
6/ Now consider the alternate. When the UK rolled out a new question on their census asking people for their religion, they got a lot of silly results (Jedi knight, etc.) ons.gov.uk/peoplepopulati…
7/ That's kind of funny and makes for good internet fodder. But as a statistical design question, it should scare UK officials that people thought it was OK to lie on this question. Might that increase their propensity to lie on other questions?
8/ Now consider all that in light of the current discussion on citizenship. Setting aside the (real and significant) problems with asking citizenship questions on the Census, we should not roll out questions census-wide that haven't been tested. On any subject.
9/ Will that question depress willingness to answer the Census (as many suspect is it's intent)? Will that question encourage people to lie on this question and others? We can speculate, but we don't know.
10/ AND... the only way we can know is to ask the question for small, randomly collected samples and then see how those respondents compare to the larger control group. Which by definition we won't know if we do census-wide.
11/ So if you want the census to provide an accurate representation of the country, and if you want to ensure people are truthful, and if you want to ensure that results from 2020 can be compared apples:apples to 2010... don't add the citizenship question.
12/ If, on the other hand, you support the continuing fetishization of ignorance and erosion of objective truth that has been at the heart of the current administration... you shouldn't be allowed to come anywhere near the census bureau.
13/ So while today's decision is a start, it should have been decisive. Now keep pushing. /fin
