Wednesday 12 December 2018

Census woes

I don't know what normal people talk about at drinks. For the past few months, whenever I'd catch up with other economists over drinks, it's been rumours about just how bad things are at Census. 

The mess hit the front page of the Dom Post today.
Documents released under the Official Information Act show the department is planning for an August 2019 deadline, and continues to stare down a high risk it may not produce viable results.

A low turnout during Census 2018 has caused considerable pain for Stats NZ, which last week announced it would not make the third deadline it set - July 2019 - and promised to announce a release date in April.

Stats NZ failed to count an estimated one in 10 New Zealanders in Census 2018, previously providing an interim response rate of at least 90 per cent of individuals providing full or partial information.
Some tales have it that the incomplete census responses are on top of the 10% non-response rate rather than part of the 10%.
It was announced in June that other sources of Government data would be used to fill gaps in the results.

Both the actual number of responses received and the number of additional data for 2018 were redacted from documents, and Stats NZ declined to release them.

The documents show "high risks" Stats NZ continues to face include: the failure of methods to patch census data; census data failing to be fit for purpose, leading to "less than ideal" decisions being made; and a failure to provide information for the re-drawing of electorate boundaries.
I really hope that they're flagging interpolated data in the back-end systems so that analysts know whether they're dealing with generated regressors. So, for example, suppose you're trying to estimate the relationship between a bunch of admin back-end variables and some Census result. If the "Census" result was really generated by those back-end variables in the first place for some of the respondents, you're going to botch your analysis unless you know that that's gone on. You could just be rediscovering the algorithm used to generate the data rather than any real relationship.
Also described are a "severe incident", where some data fields on the census forms failed to cross over to another system and contributed to a month-long setback.
... BERL chief economist Dr Ganesh Nana said Stats NZ needed to "draw a line" under Census 2018.

"We all might have to put our hands together and say, 'Okay, the data won't be as rigorous and robust as previous census'. We might just have to admit that."

A standard census release would have BERL using the data intensively over a six-month period from about now, he said.

"That's clearly been put out of kilter this time around. The biggest concern for us is the timing of the data … the data is already starting to become stale."
And drinks chatter has speculation over whether the whole thing will yet require a do-over.

It isn't the first year that they've combined online and paper - I filled in my Census online in Christchurch last time. Quite why everything's gone wrong this time - I'd love to know.

I'd made brief allusion to the problems at Census in my column at NewsRoom yesterday ($). That column wasn't about Census but about the fun they're having with the Indicators Aotearoa project. Will post on that one soon.

Update: Ganesh Nana talks with Kathryn Ryan about it on Nine to Noon.

Update 2: Hoisted from the comments (on the mobile version, where I've not been able to get Disqus to work):
One big difference between this Census and the last one was that last time the online form was opt-in, but this time the paper form was opt-in. In 2013 everyone received a paper form, but people who registered could ignore that and do it online instead. In 2018 you had to register to receive a paper form. If you didn't get one in time then it was do it online or not at all. That's a big difference in practice and could be part of the reason for the problems. 

2 comments:

  1. That is a concern. I am sure the problems in Hillmorton hospital where I work are in part related to a larger pop increase than we are funded for...

    ReplyDelete
  2. One big difference between this Census and the last one was that last time the online form was opt-in, but this time the paper form was opt-in.

    In 2013 everyone received a paper form, but people who registered could ignore that and do it online instead. In 2018 you had to register to receive a paper form. If you didn't get one in time then it was do it online or not at all. That's a big difference in practice and could be part of the reason for the problems.

    ReplyDelete