Should this citizenship question be added to this census?

Tyler Cowen writes, in a blog post and Bloomberg opinion piece, that

If you can’t ask about citizenship on your census, as indeed Canada and Australia do, it is a sign that your broader approach to immigration is broken.

That sentence contains a number of errors.

But first, I agree entirely with his conclusion: America’s “broader approach to immigration” is broken. There are more than 10 million undocumented people living in the United States, the vast majority of whom have lived here more than a decade. These people have established lives here and deserve a path to citizenship. But in the meantime, their existence feeds a weirdly adversarial attitude to citizenship and immigration in the US, from both the right and left: ICE raids and sanctuary cities and so on. People on the left shy away from sharp notions of citizenship because of the inhumane effect it would have on undocumented people. Racists embrace citizenship for precisely the same reason. It’s total political failure leading to a toxic climate for serious discussion.

Okay, on to some misconceptions about the citizenship question.

You can’t ask about citizenship on the census #

You can. The Supreme Court said so. It is entirely constitutional (2, in the headnote), and an administration probably doesn’t even need a very strong reason to do so (4, in the headnote). The only reason this attempt was blocked is that the administration lied and lied and lied about why they were adding it (6, in the headnote).

I think this was more-or-less the right decision. You can ask a citizenship question on the US census. You just can’t blatantly lie to federal courts about why you’re asking it. That’s a pretty low bar to clear.

But Canada and Australia ask about citizenship #

First, despite what you’ve read elsewhere, Canada does not—at least not in the relevant sense. Canada’s census is split into two parts: a “short form,” with only 10 questions for each person, and a “long form,” with 49 questions for each person. The short form goes to everyone in the country, whereas the long form is only sent to a sample of 25% of households. The citizenship question only appears on the long form questionnaire.

So what, you may say, it is still on the census. Well, by that standard, the citizenship question is already on the US census too. Until 2000, the US also had a short form / long form census as Canada does. But then, the once-a-decade long form was replaced by a “rolling” survey, the American Community Survey (ACS). It’s a little bit complicated: the ACS is ongoing, reaching out to around 3.5 million households every year. But it’s basically still the long form census (the Census Bureau considers the ACS to be part of the census program, and France does its entire census like this). Over a decade it is sent to a quarter of all households, the same general reach that the Canadian long form census has. And the ACS already asks the exact citizenship question about which we are arguing. So the US and Canada actually treat the issue very similarly.

Now Australia does, indeed ask every person, not just a sample, about citizenship. But that’s because Australia, in effect, has only a long-form questionnaire. It asks each person more than 50 questions. It’s a very different questionnaire for the US short form, which excluding the citizenship question asks only 5 real questions—age, hispanic origin, race, relationship within household, sex—plus one, tenure, for the household in general.

Adding a new question to a form that only asks 5 questions is a big deal. That’s why the Census Bureau was resistant to the original, late request.

Fine. Why can’t the US use a longer census for everyone like Australia does? #

It used to. Have a look at the 50 questions on the 1940 census. The census form ballooned over roughly the century from 1850, as politicians, academics, businesspeople and others demanded to know more and more about the population.

Two things interrupted this trend, from around the late 1940s. Firstly, the Census Bureau began to realize that they were undercounting people. And not just any people—the census tended to systematically miss minorities and other disempowered people: black, hispanic, recent immigrants, etc. As the civil rights era dawned, this was not acceptable, and so the Bureau increasingly sought ways of reducing that undercount: for example, translating the form and other materials, and working with community organizations on outreach.

The second thing that killed the old longer census was advancements in statistical sampling methodology. This meant that the Bureau could safely ask some questions of only a sample of households, but still know how accurate the sample would be. For less important information (“Do you have a clothes washing machine?”), this was a good compromise.

It turned out that making the census form much shorter would increase response rates, because it took less time to complete and because it seemed less intrusive. So the long form / short form distinction was born.

So how does Australia do it? Doesn’t the Australian Bureau of Statistics care about undercounting people? #

It does, but the stakes are much lower in Australia (and Canada for that matter, to which all of this also applies) for two reasons. All three countries—US, Canada, Australia—redistribute electoral representation, after a census, in two parts: a formulaic part, which more-or-less automatically determines how many representatives each state or province, receive; and a subjective part, where the actually within-state boundaries are drawn.

First on the subjective part: electoral boundaries in Australia and Canada are drawn by independent commissions, not politicians. Gerrymandering is not a big deal in these countries. Because the process is seen as impartial, ta lot of heat is taken out of the whole debate. The fine details become less important if you trust the process that produced them.

Second, and perhaps more importantly here, both Australia and Canada adjust their population results to correct for undercount. This is important because no census of anything larger than a village can possibly expect to count everyone: up to 10 percent of people being missed is not that unusual. But there are robust, widely-accepted statistical methods to measure that undercount, and the same methods can be used to adjust the raw counts from the census, making them more accurate.

Both Australia and Canada use those adjusted figures for the equivalent of Congressional apportionment, which means that, while it is still important to try to count everyone in the first place, the political consequences of failing to do so are reduced. (Here’s Canada’s 2012 allocation, and the population data it used: note the phrase “counts adjusted for census net undercoverage.”)

Undercount adjustment sounds like a great idea. Why doesn’t the US Census Bureau just do that? #

It has tried. The Bureau collects the data necessary to do it, from something called a post-enumeration survey, that re-surveys a sample of census respondents. But the Supreme Court has found, in a series of decisions from the 1980s forwards, that sampling methods may not be used in the calculation of population figures that are used for apportionment.

There was a fairly sensible proposal on the table prior to the 2000 census, which would have cut back the extremely costly “non-response follow-up”—that is, census employees driving around knocking on doors and quizzing neighbors of people who fail to respond to the census the first time—while boosting the post-enumeration survey. It would, in all likelihood, have reduced costs, improved accuracy, and taken the political sting out of the undercount.

The Supreme Court killed it. It concluded that federal Census Act prohibited sampling for the purposes of the apportionment counts. The court didn’t rule on whether the constitution, in principle, might allow it, although hints given then and the current make-up of court makes it likely that the words “actual enumeration” would be read narrowly, as excluding sampling.

What’s the principled argument against post-census adjustment? #

Broadly, I think, that adjustment brings in—or seems to bring in—an element of subjectivity, so it could lead to endless arguments about the exact method used.

Imagine that: endless arguments about the census.

What about the argument that the citizenship question has often been asked in the US / has never been asked in the US? #

Here, I think both sides have it wrong. Conservatives who claim that census question has often been asked are ignoring the subtle variation in how it has been asked, and more importantly, the Bureau’s long, careful history of pruning the census back to its bare essentials.

But people who claim that “the citizenship question has never been asked of everyone” are splitting hairs somewhat disingenuously. Take the 1950 census—which remember, was conducted as personal interviews, rather than self-enumeration. The enumerator asked first, “What state (or foreign country) was he born in?” and only if the answer was a foreign country, did they ask the next question “Is he naturalized?”

So, sure, technically that question was not asked of everyone. But since the United States has birthright citizenship, you can fairly safely assume that somebody born here is a citizen. So they did collect citizenship information from everyone, they just didn’t ask directly.

Amongst survey-design people, this kind of thing is called “skip logic”: you include instructions to only ask questions of people to whom it is relevant (i.e. you don’t ask men how many children they have given birth to). It’s easier to use complex skip logic in an interview situation or with an online survey where it can be automated. Doing it on a paper, self-enumeration form can be tricky because people don’t read instructions. (How many times have I inadvertently answered that I am not pregnant on a medical form?)

There are probably several reasons they asked it in two stages like this in 1950. First, they wanted to collect the place of birth anyway, so they could save enumerator time by using it to skip the naturalization question. Secondly, citizenship might not have been as well understood a concept in 1950, so it was probably clearer to ask it like this.

But the effect was the same as the proposed question.

What about the argument that other countries ask this question? #

This is a weak argument. It’s true that the UN publishes a 300-page book on Principles and Recommendations for Population and Housing Censuses. It’s also true that that book lists “country of citizenship” as a “core topic”. But it lists some 26 core topics, and remember the US census has only 5 questions. The US census (as opposed to the ACS) does not closely follow the UN recommendations.

In the 2010 round of censuses, 20 percent of countries (p. 14) did not ask about citizenship (not including the United States: the UN Statistics Division also appears to consider the ACS part of the census).

While the census is a fairly international undertaking, local conditions and sensitivities matter a lot. For example, Australian and the UK ask about religion and ethnicity/ancestry, but this is taboo in France (because of Vichy history and a long tradition of secularism).

The UN guidelines say this:

Prime importance should be given to the fact that population censuses should be designed to meet national needs. … The prime consideration is that the census should provide information on those topics that are of greatest value to the country, with questions framed so as to elicit data of maximum utility.

In other words: