Household wealth and housing wealth: first quintile oddities edition

In my column over at Newsroom this week, I noted one strange feature of New Zealand's household wealth statistics:

Unfortunately, data on wealth is far worse than data on income – the Government gathers a lot less data on wealth. For example, Statistics New Zealand reports that people in the least wealthy 20 percent have $1.75 in property debt for every $1 in property assets – but no bank in the country would extend a loan on that basis. It is more likely that the survey data misses some houses owned through family trusts where, for example, a 25-year-old takes over the mortgage and effective ownership of their parents’ second home held in the family trust. The mortgage payments and mortgage debt are noted in Household Economic Survey data, but the ownership may be missed. This means their parents’ net wealth will be overstated, their own net wealth will be understated, and measured wealth inequality among younger cohorts would be somewhat understated but overall measured wealth inequality would be somewhat overstated. 

There's an ungated version of the column here.

The effect wouldn't be large because there aren't many households in that situation. But it's odd. So I asked Statistics New Zealand what's up with that, wondering whether the trusts explanation might be what's going on. I'd also asked whether Stats were able to ask follow-up questions of their HES respondents to see what's going on.

Stats' Statistical Analyst Michelle Griffin helps me out:

We’ve had a look at the distribution of owner-occupied property assets and owner-occupied mortgages for those in net worth quintile one. There are about 30 more households (unweighted) with mortgages than homes, and because of this difference, you need to go further along the mortgage distribution to get to the median, resulting in a larger mortgage median and hence a larger debt to asset ratio. So it’s looking like the high ratio is due to a mix of under-reporting and distribution differences.

We’ve had a look at only those with both a home and a mortgage to see what they look like (this is excluding those with a home in a business or trust), and this brings the ratio down to $1.44. This drops even further when you take the median of each household’s individual debt to asset ratio (to $1.26).

We’ve had a quick look at the households with only a mortgage to try and see what might be happening. They appear to be a mix of households with the home in a trust or business which is assigned to someone else (which could be a situation like you’ve described below), and households which say the home is in a trust or business but then don’t mention this trust/business later on (could be partly due to misreporting or respondent burden leading to households not answering fully). I think the fact that there are so few households in net worth quintile one with owner-occupied property is making this quintile more sensitive to these issues. Unfortunately, we can’t follow up with the respondents later on to confirm these situations – although it would certainly be interesting!

So restricting things to those households in the first quintile fixes some of the problem, but we're still at a debt-to-equity ratio on housing debt well out of step with sane banking practice, let alone existing LVR rules. And while there aren't many households with mortgages without homes, it's hard to tell whether debt to asset ratios for a broader set of respondents elsewhere mightn't be wrong too. 

Michelle suggests that keeners might apply for microdata access to really drill down into what's going on. It would be a great Honours thesis for somebody wanting to get microdata experience who has a qualified supervisor.

Many thanks to Michelle and to Statistics NZ for the helpful and detailed response.

Update: An informed reader writes:

A couple of thoughts about the data on wealth:

• There is a wider problem with data for people at the top and bottom of the distribution. Surverys include lots of trade offs between clarity, comprehensiveness and accuracy. For instance survey data on the benefit population substntially under-reports income from second and third tier payments (that is all loans, one off payments, disability allowance and accommodation supplement). Since this is a relatively small part of the population, SNZ makes the reasonable decision not to overcomplicate their survey to deal with the problem, but the consequence is continuous understatement of income for many people. The same applies at teh top where it is hard for surveys to capture the many different assets people with many assets have.


• More technically there is problem using a small number of  discrete categories like quintiles to describe a continuous variable. The general point is that if someone makes a mistake they have have to go somewhere in the discrete categories, so effectively there is a cut off at the top and bottom that will skew the results. It's easy to think of this through an intuition. If a person's income is roughly in the middle, say $50 000, and they miss a zero then they drop to the bottom of the distribution ($5000) while if they add a zero by accident ($500 000) they go to the top of the distribution. However a person whose income really was $5000 who makes a mistake either stays in the same place or goes up (ie income is $500 or $50 000); while a person with income at $500000 either goes down, or stay at the top. Even if we assume the liklihood of a person making an error and its direction are independent of income, the measured income will be pushed to the extreme quintiles. "On average" it may be right, but it will overstate the dispersion. My guess is this will be even more pronounced with wealth because it is so much harder to measure and thus more likely to generate error-prone answers.


• Finally, they are using snapshots for variables that change over time. I know this is a problem with flows like income and employment, but I sometimes wonder what it means for wealth. In particular, there is an issue with measuring transitions. The following example is made up, but gives a clue what might be happening. Say some % of people selling a house to buy another spent a couple of weeks where they had two mortgages while the paper work was settled. For the individuals the cost might be a few hundred dollars - which is not a lot when you get real estate agents involved! - but two weeks is approx 4% of the year so you would expect your snapshot survey to include some small % of people with double the mortgage debt relative to their asset. It is not many people, but it does not need to be because they will be a small percentage of the population but they will concentrated in decile/quintile. 

To me the deeper problem is that surveys tend to be designed to be "on average right", whereas the data is often used to make distributional statements. And Bryan Perry’s incomes report notes the problems with the extreme tails of the distribution.