This week in research #8

I've been busy again this week, which is why this post is coming a day later than usual. Here's what caught my eye in research over the past week:

• Ivandić et al. (open access) look at the dynamics of domestic abuse calls around the time of football matches in Manchester, and find a decrease in incidents during the two-hour duration of the game (football and domestic violence are substitutes), but an increase after the game, which peaks ten hours later
• Barrington-Leigh (open access) shows that research participants with lower education levels are more likely than those with more education to ignore points on a Likert scale (such as for measuring subjective wellbeing or happiness), and reduce their choices to top, middle, and bottom, and then proposes a solution to this problem (relates to other problems with subjective wellbeing - see here, and in the links at the end of that post)
• Bursztyn et al. (with ungated earlier version here) show that decades-long exposure to immigrant groups (focusing on Arab Muslims in the US) decreases explicit and implicit prejudice, among other positive outcomes
• Cabral and Dillender (with ungated earlier version here) find that, compared to male patients, female patients randomly assigned to a female doctor rather than a male doctor are 5.2 percent more likely to be evaluated as disabled and receive 8.6 percent more subsequent cash benefits on average

And another paper from my own research:

• Again with a large number of co-authors led by John Oetzel, our new article (open access) in the journal BMC Geriatrics reports on a strengths-based programme to improve health among kaumātua (Māori elders), and finds that the programme was both effective in improving health-related quality of life, and cost effective at doing so 

Cheating in the lab and in the field

A commonly used test for dishonesty or cheating in lab experiments is to ask research participants to privately (unobserved by the researcher) roll a number of dice and report the number of sixes they roll, or to privately toss a number of coins and report the number of heads. If research participants are paid for each six they roll, or each head they toss, then they have an incentive to cheat, by reporting more sixes or more heads than they actually obtained. Now, this doesn't give an individual measure of dishonesty, since a research participants might really have tossed six heads in a row (on average, one in every 64 research participants should have that outcome), but it does provide a measure of the extent of cheating across the whole sample (for example, if half of research participants report tossing five or more heads out of six, you can be pretty sure there is a lot of cheating going on).

The real question, though, is does this measure of dishonesty replicate outside of the lab environment? That is essentially the question addressed in this 2018 article by Alain Cohn (University of Michigan) and Michel André Maréchal (University of Zurich), published in the Economic Journal (ungated earlier version here). Using a sample of 162 public middle and high school students in Switzerland, Cohn and Maréchal started with the lab measure of dishonesty:

Subjects first opened an envelope containing 10 coins, each worth 0.5 Swiss francs (about US $0.55). Then, they were instructed to toss each coin in private and report their outcomes on paper. For every coin toss for which subjects reported the outcome ‘heads’ they were allowed to keep the coin; they had to put the coin back into the envelope otherwise. Participants thus faced a financial incentive to cheat by misreporting the outcomes of their coin flips without any risk of getting caught... The stakes were considerable as the maximum possible payoff in this task corresponds roughly to half the amount students of similar age receive in pocket money every week.

Cohn and Maréchal then look at the relationship between the number of heads reported and measures of school misconduct (reported by their teachers): (1) disruptiveness in class; (2) non-completion of homework; and (3) absenteeism. Since the three measures of misconduct were highly correlated, they created a single index from them (but their results also appear to hold for each measure individually). They found that:

On average, the students took 62.8% of the coins in the envelopes (95% confidence interval: 60.0%, 65.7%)...

Given that on average, students should have taken 50 percent of the coins, there was a substantial amount of cheating. Who cheated most? Cohn and Maréchal report that:

...female students behaved more honestly than male students as they took significantly less coins (p < 0.000, t-test)... Moreover, we found that high school students took significantly less coins than those from middle school after controlling for age (p = 0.011, t-test), which could be explained by less deviant students selecting into higher education. Earnings in the coin tossing task and the two measures of cognitive ability are negatively correlated. However, the correlations do not reach statistical significance, neither for crystallised nor for fluid intelligence (p = 0.599 and p = 0.744, t-tests).

What about school misconduct (cheating outside the lab)? Cohn and Maréchal report that:

...behaviour in the coin tossing task is significantly related to school misbehaviour when controlling for age, gender, nationality, education level and parental education. A higher number of coins taken is associated with increased behavioural problems in school (p = 0.015)... The coefficient estimate implies that the difference in school misbehaviour between students who took 10 coins (presumably cheaters) and those who took five coins (presumably honest individuals) is more than 0.7 points (or 0.53 standard deviations) on average. For comparison, it would require students to differ by 2.7 standard deviations in cognitive ability (i.e. crystallised intelligence) to produce the same difference in school misbehaviour.

So, the important takeaway from this paper is that there is support for the external validity of the lab measure of cheating or dishonesty. At least, to the extent that misbehaviour by school students is the same as dishonesty (which, of course, it isn't). So, while this should give experimental economists and others a little bit more comfort in using the lab measures of dishonesty, more studies of the external validity of these measures, in other contexts, are sorely needed.

An extraordinary (but weak) claim about China's poverty reduction performance

Over the last forty years, global poverty has reduced dramatically. To see how dramatically, run this animation from Our World in Data (source here):

That animation is based on a poverty line of US$30 per day, which is about the average poverty line in a developed country. A poverty line is a level of income that separates the poor (who have income below the poverty line) from the non-poor (who have income above the poverty line). The poverty rate is then the proportion of the population who are poor. For example, the World Bank uses a poverty line that is currently US$2.15 per day (which is US$1 per day in 1996, adjusted for inflation) to determine rates of extreme poverty. By that measure, over the last 40 years, 800 million people have been lifted out of poverty, and China contributed close to three-quarters of that number.

The level of the extreme poverty line has been the subject of significant debate over the years (with many articles and several books devoted to the topic). Nevertheless, that global poverty has declined, and that China has contributed significantly to this success, is generally accepted. So, I was very surprised to read this article in The Conversation by Dylan Sullivan (Macquarie University), Jason Hickel (Autonomous University of Barcelona), and Michail Moatsos (Maastricht University), which presents a counter-view. They argue that:

In contrast to the World Bank, we find that from 1981 to 1990 – at the end of the socialist period – China’s rate of extreme poverty was one of the lowest in the developing world. It averaged only 5.6%, compared to 51% in India, 36.5% in Indonesia and 29.5% in Brazil.

We find extreme poverty increased dramatically during the market reforms of the 1990s. It reached a peak of 68% as price deregulation pushed up the cost of basic food and housing, cutting the buying power of low-income people.

Extreme poverty then slid during the 2000s, but has yet to fall to the levels calculated by the World Bank.

This is an extraordinary claim, so I read the underlying research article, which was just published in the journal New Political Economy (open access). The key difference between Sullivan et al.'s results and those of the World Bank lie in how the poverty line is calculated. As they explain:

In recent years, scholars have developed an alternative approach to measuring extreme poverty, which compares incomes against the cost of basic needs in different contexts (Moatsos 2016, Allen 2017, 2020). In 2021, the OECD published estimates of the share of the population below this ‘basic needs poverty line’ (BNPL), for all countries with available household survey based data from 1981 to 2008.

It is the use of this Basic Needs Poverty Line (BNPL) that explains the extraordinary results, and that is because when it comes to basic needs:

Socialist policies of public provisioning and price controls may keep the cost of meeting basic needs quite low compared to capitalist contexts characterised by high levels of commodification and privatisation. This means that any given level of broad-gauge PPP income would have a greater welfare purchasing power – in terms of basic needs – under socialism than under capitalism; people would have better access to the key goods, such as food and housing, that are necessary for escaping extreme poverty.

So, as China opened their economy up in the 1980s, privatising housing and other markets, the cost for a household to satisfy their basic needs increased, meaning that more households would be defined as extremely poor based on the BNPL measure. So, they end up with this comparison between China and several other 'middle income' countries (Figure 4 from the article):

Sullivan et al. then go on to compare those countries' performance across a range of measures that should be correlated with living standards, including literacy, school enrolment, life expectancy, health access (physicians and hospital beds per 1000 people), and calorie availability. They compare China with those other countries in 1981 and in 1990, a period where extreme poverty was reducing based on the World Bank poverty line, but where poverty based on the BNPL was relatively flat (as shown above). With these comparisons, they show that:

In sum, the empirical data on social indicators raises significant questions about the validity of the World Bank’s estimates of the poverty rate in China during the 1980s. In 25 out of 29 comparisons, China achieved the 1st or 2nd best score of the countries reviewed here, as we would expect from China’s relative performance on the BNPL.

There is a problem with the comparisons that they make though. They look at how China ranked in 1981 and 1990 compared with these other countries, and use China's declining ranking compared to these other countries as evidence in support of the BNPL data. However, they don't make the very obvious comparison of China in 1990 with China in 1981 on those same measures. If extreme poverty was decreasing, as the World Bank data suggests, then we would expect an improvement in these other measures of living standards. And indeed, that is what we see. The literacy rate in China increased from 66% to 78%, life expectancy increased from 67 to 69, and calorie availability increased from 103% to 119% of basic needs. None of that is inconsistent with a reduction in extreme poverty. Of course, that is also consistent with the BNPL data, which shows an (albeit slight) decrease in extreme poverty between 1981 and 1990 as well. An even better comparison would have been to look at the 1990s, where poverty based on the World Bank data and poverty based on the BNPL dramatically diverged. However, Sullivan et al. don't make that comparison (presumably because it doesn't support their argument).

The setting of a global extreme poverty line is quite fraught, and there will always be disagreements about how it should be determined, and at what level it should be set. However, as the American astronomer Carl Sagan noted, extraordinary claims require extraordinary evidence. The evidence that Sullivan et al. have brought to support their extraordinary claim is not extraordinary. Before we overturn the consensus that China dramatically reduced extreme poverty during the past 40 years, we need something more.

The post-COVID-19 housing boom as a cause of the US Great Resignation

The term 'Great Resignation' came into use last year to refer to a sustained decline in the labour force participation rate in the US. As I noted last year, there is little evidence that there has been a Great Resignation in New Zealand, despite media commentary suggesting that there has been. However, this working paper from last year, by Jack Favilukis and Gen Li (both University of British Columbia), is making me wonder why New Zealand didn't experience a Great Resignation.

Favilukis and Li look at whether the COVID-19 housing boom explains the US Great Resignation, using data from the American Community Survey. They start by showing that the Great Resignation was concentrated among older workers. Specifically:

First, up to 2019, the labor force participation rate was rising for all age groups, but especially for older Americans. Second, in 2020, labor force participation fell for all groups, but most dramatically for the youngest and oldest Americans. Third, with the exception of the oldest Americans, all groups returned to the labor force in 2021 and were near or even above 2019 rates by 2022. On the other hand, the oldest Americans further reduced labor force participation in 2021 and continued to stay out of the labor force in 2022. By 2022, nearly the entire reduction in the labor force participation rate was due to the 65+ group. This is especially striking given the pre-2020 trend.

Favilukis and Li then look at whether employment, or weekly hours worked, are related to the housing market return over the past 4.5 years. They estimate this relationship separately for each year, age group, and homeowners/renters, allowing them to look at whether these various groups were affected in different ways by the post-COVID-19 housing boom. The results don't appear to be particularly sensitive to the choice of 4.5 years of housing market returns as the key variable, with other lengths of time showing similar results.

They find that, for 2021:

Renters tend to increase their labor force participation in response to house price increases; a 40 year old renter increases her probability of being in the labor force by approximately 0.07×0.10 =0.7% for every 10% increase in house prices (e.g. from 80% to 80.7% participation). This is statistically significant... Older renters are less reactive to house price changes, with slopes slightly positive but rarely statistically significant.

Younger owners also increase their labor force participation in response to house price increases, although their response is much lower than that of younger renters... Middle aged owners are relatively unresponsive to house price changes – their labor force participation falls slightly in response to higher prices but this is not statistically significant for one year age buckets.

However, individuals above 60 have a strong negative response; a 65 year old owner decreases her probability of being in the labor force by approximately 0.11 × 0.10 =1.1% for every 10% increase in house prices (e.g. from 15% to 13.9% participation). This is strongly statistically significant...

So, it appears that the 'Great Resignation' was concentrated not only among older workers, but among older workers who are homeowners, rather than renters. And, the propensity to be out of work within that group is strongly related to housing returns. Moreover, in metropolitan areas where housing returns were higher, the Great Resignation among older homeowners was bigger.

Which brings me back to my comment from the start of this post. Post-COVID-19 housing returns have been relatively strong (according to the QV House Price Index), and yet we haven't experienced the same substantial decrease in the labour force participation rate as seen in the US. Here's the labour force participation rate among people aged 65 years and over in New Zealand over the period since 2016 (source here):

There's little evidence of a Great Resignation among older people in New Zealand. The trend remains generally upwards over time. Unfortunately, the freely available data don't allow us to look at the difference between homeowners and renters, but if there were similar effects to those reported in the Favilukis and Li paper, the Great Resignation among older homeowners would be apparent in the overall statistics without looking at those homeowners and renters separately.

Favilukis and Li conclude that:

High house prices allowed many older Americans to retire early; if not for the high house prices, their labor force participation in 2021 would have been similar to 2019.

However, clearly there is more to the story than simply a housing boom leading to Great Resignation, otherwise we would have likely seen the same effect in New Zealand. New Zealand house prices are high. Why didn't that induce older New Zealand homeowners to retire early? Hopefully, someone is investigating that question.

[HT: Marginal Revolution, last year]

Read more:

• The myth of the Great Resignation (in New Zealand)