Performance pay, stress, and alcohol and drug use

Employers use performance pay to incentivise their workers to work harder. However, that increases workers' stress and uncertainty, both of which may in theory lead to an increase in workers' use of alcohol and drugs as a coping mechanism. In a new article published in the Journal of Population Economics (ungated earlier version here), Benjamin Artz (University of Wisconsin-Oshkosh), Colin Green (Norwegian University of Science and Technology) and John Heywood (University of Wisconsin-Milwaukee) test this theory, using data from the U.S. National Longitudinal Survey of Youth 1997 (NLSY). The NLSY collects data on whether the respondents received performance pay ("tips, commissions, bonuses, incentive pay, and a small “other” category") in each wave, as well as marijuana or alcohol use within the previous 30 days, and other drug use (which they refer to as 'hard drugs') within the last year. The sample size is over 60,000 observations.

The problem with a simple analysis here is that the types of people who are more likely to accept a job with performance pay (i.e. younger people, and people who are less risk averse) are also those who are more likely to consume alcohol and drugs. So, a positive correlation between performance pay and alcohol and drug use is to be expected because of 'selection bias'. Artz et al. use a number of techniques to get around this bias:

First, we include proxies for risk preferences, ability, and personality incorporating sophisticated error structures... Second, we use the survey’s panel structure to hold constant time-invariant worker fixed effects that could include unmeasured risk preferences or ability. Third, we recognize that changes in unmeasured worker characteristics can lead to both job change (and so performance pay receipt change) and to a change in substance use. We respond by controlling for job match fixed effects. Thus, we examine the change in individual workers’ substance use when their employer changes their performance pay status (even as they remain in the same detailed occupation).

So, they control for risk preferences (albeit imperfectly - I'll come back to that point), and then in their most restrictive specification the fixed effects reduce their comparisons to the same worker, working for the same employer. Essentially, they look at what happens when a worker goes from not having performance pay, to having performance pay (or vice versa), while still working for the same employer. In the first specification (excluding risk preferences and the fixed effects), they find that:

The main estimates of interest reveal large, positive, and statistically significant, relationships between performance pay and all types of substance use. The odds ratio... indicate that holding other determinants constant, performance pay workers have odds that are 29% higher for marijuana use, 35% higher for hard drugs use, and 45% higher for alcohol consumption.

The inclusion of risk preferences (which are measured only in the 2010 wave) do not change the estimated relationships greatly. Then, in their most restrictive specification (with worker-employer-job fixed effects), they find that:

The results indicate a 29% increase in the odds ratio for marijuana use, a 26% increase in the odds ratio of hard drug use, and a 34% increase in the odds ratio alcohol use... In sum, the results... indicate that the relationship between PRP receipt and marijuana, hard drugs, and alcohol use persists despite worker sorting on time fixed unobserved worker characteristics, or worker sorting across employers.

A potential problem with these analyses is that risk aversion decreases as people get older, and alcohol and drug use also decrease, and so even though they control for risk aversion measured at a single point in time, or control for job-match fixed effects, they still potentially don't fully eliminate the selection bias. However, it's difficult to see how this research design could be much improved upon. While jobs can be randomised to receiving performance pay, researchers couldn't force workers to accept the performance pay condition, so a randomised experiment would not work.

Performance pay does increase incentives for work effort in some jobs, but that doesn't mean that it comes without cost. Some of the cost is of course borne by the employer, but workers also must endure a more uncertain and stressful work environment. The results of this study are consistent with that story, and that some workers respond by self-medicating with alcohol and drugs.

[HT: Marginal Revolution]

Starbucks and the tragedy of the bathrooms

Back in 2018, Starbucks hit the news for its bathroom policies. As Vox reported at the time:

Starbucks will treat anyone who walks into one of its cafes as a customer, whether or not they buy anything, the company said on Saturday. The announcement is the latest step the coffee company is taking as part of its ongoing response to the public outcry over the arrest of two black men at a Starbucks in Philadelphia. The men were waiting for a business associate to arrive and had asked to use the bathroom in April when Starbucks employees called the police, eventually leading to them being arrested and escorted out...

The arrest of Rashon Nelson and Donte Robinson on April 12 in Starbucks kicked off a major firestorm. The company initially issued a less-than-satisfying apology, and CEO Kevin Johnson later issued a lengthy statement on the incident in which he apologized to the men arrested, laid out plans to investigate the incident, and affirm Starbucks’ stance against discrimination and racial profiling. “You can and should expect more from us,” he wrote. “We will learn from this and be better.” 

What effect did the change in bathroom policy have? In a 2020 paper, Umit Gurun (University of Texas at Dallas), Jordan Nickerson (MIT), and David Solomon (Boston College) investigated that question. They first collated anonymised cellphone location data from SafeGraph (which I really wish was available for New Zealand, but it turns out is only available for the US, UK, and Canada), and compared the change in monthly visits to Starbucks before and after the policy change, with the change in monthly visits to other nearby coffee shops, and the change in monthly visits to restaurants (essentially, this is what we refer to as a difference-in-differences analysis). Using data covering the period from January 2017 to October 2018, they find that:

...Starbucks stores experienced a 7.0% decrease in visits after the enactment of the policy, compared with similar coffee shops and restaurants... After the policy change, Starbucks saw a small time-series increase in visits, whereas absent the policy a much larger increase would have been expected. Put differently, the general boom in visits to all coffee shops at the time helped disguise the fact that the new policy appears to have significantly reduced visits to Starbucks.

So far, so unfortunate for Starbucks. However, Gurun et al. aren't done. They look at how the effect of the policy differed depending on the distance to the nearest homeless shelter, and find that:

Strikingly, the decrease after the policy enactment is significantly larger the closer the location is to a homeless shelter. Stores less than two km away experienced declines of 8.5% relative to nearby coffee shops, while stores more than 10 km away experienced declines of only 4.8%. Again, this decline in attendance is not from worsening economic conditions in these areas – rather it captures the change in Starbucks relative to nearby coffee shops experiencing the same local economic conditions.

Their results hold when they switch to a synthetic control method as a robustness check. Interestingly, they have some evidence that different types of customers are affected differently as well:

Starbucks also experienced a significant change in the demographics of who visited the store. Relative to other coffee shops and restaurants, Starbucks saw a larger decline in visitors from relatively wealthier home locations. The estimated income of Starbucks customers declined by 0.4%, relative to changes in other coffee shops and restaurants... Despite the racial angle on the initial controversy, we find no difference in the racial demographics of the home locations of Starbucks visitors after the policy. In other words, the new policy appears to have deterred both black and white customers in roughly equal amounts.

To summarise those results so far, Starbucks' change in bathroom policy decreased their customers relative to other coffee shops, and the effect was greater for Starbucks stores closer to homeless shelters, and presumably had a larger effect on its wealthier customers.

Was there anything good that came out of this policy change? Gurun et al. look at the effect on crime for a subset of cities where appropriate crime data are available (Austin, Denver, and Pittsburgh), and find that there was:

...a decrease in public urination citations near Starbucks locations relative to other areas after the policy change. By contrast, a wide range of other minor public order crimes show no significant changes or consistent signs of effects.

The other crimes that they looked at included disturbing the peace, simple assaults/fighting, marijuana possession, shoplifting, theft of service, threats/harassment, and vandalism. The overall question is whether this was a good policy change for Starbucks - Starbucks face all of the cost, but aside from some good press (or, more accurately, a reduction in bad press), the benefits are public. Gurun et al. link this to the private provision of public goods, concluding that:

Our results suggest that companies may be better off focusing on donating money to worthwhile causes, and effectively using a division of labor, whereby Starbucks specializes in making and selling coffee, and engages in CSR by supporting organizations who specialize in social policies. Our results show that trying to incorporate the two within a single company may result in outsized negative externalities for the underlying business that makes CSR possible in the first place.

I don't agree. Despite the adjective, public bathrooms are likely not a public good. Public goods are good that is non-rival (where one person using the good doesn’t reduce the amount of the good that is available for everyone else) and non-excludable (where the good is available to everyone if they are available to anyone). Opening your bathrooms up to non-customers changes the bathrooms from excludable to non-excludable. However, I'm not convinced that they are non-rival. Anyone who has had to queue for a public bathroom would have to agree - one person using the bathroom reduces the amount of bathroom capacity available for everyone else (and makes you wait). So, at least during peak times, public bathrooms are a rival good. Goods that are rival and non-excludable are referred to as common resources.

Common resources are vulnerable to a problem that we refer to as the Tragedy of the Commons, a problem that was first described by William Forster Lloyd in 1833, but was brought to modern attention by Garrett Hardin's 1968 article of that title published in the journal Science. In the Tragedy of the Commons, private incentives and social incentives differ. The social incentive is to keep the common resource well-maintained, so that it is available in sufficient quantities for everyone. The private incentive is to use as much of the common resource as possible, because each user faces the full cost of restraining their activity, but receives only a small share of the benefits of their restraint.

Turning back to the example of public bathrooms, users have a low incentive to keep them clean and tidy. Which is why, and I'm sure you can relate to this experience, public bathrooms may in general be some of the grossest places on the planet. It's little wonder that fewer people would want to visit Starbucks, if the quality of their bathrooms has degraded. The paper by Gurun et al. seems to focus attention on the types of people using the bathrooms (hence, the homeless shelters angle). However, bathroom users don't have to be homeless to fail to keep the bathroom clean.

So, Starbucks likely faces costs on two sides from their policy change - reduced foot traffic (relative to other coffee shops) as shown empirically by this study, and theoretically higher costs of bathroom clean-up as well. And Starbucks' only benefit appears to be avoiding negative attention. Is it paying off for Starbucks? It must be, because the policy remains in place today.

[HT: Marginal Revolution, last year]

The procyclical relationship between mortality and economy

Since the beginning of the pandemic, I've gotten more than a little tired of the whole economy vs. public health debate. Especially when lockdowns were new, the media seemed to have this lazy public-health-experts-favour-lockdowns vs. economists-argue-against-lockdowns theme going on. But what if both sides were wrong? Lockdowns may save lives from coronavirus, but reduce preventive healthcare and diagnosis of medical problems like cancer, which might result in more deaths overall. Not having lockdowns may reduce the economic impact from business closures, but what about the economic impact of increased death and illness from coronavirus? I think it will take us a while to disentangle everything, and hopefully we'll know better for next time.

In the meantime, here's something else to think about. In this discussion paper from late last year, Kadir Atalay, Rebecca Edwards, Stefanie Schurer, David Ubilava (all University of Sydney) looked at the relationship between the business cycle and mortality, using annual state-level data from Australia over the period from 1979 to 2017 (see also the non-technical summary on The Conversation). Interestingly, past research has shown that there is a negative relationship between unemployment (a measure of the business cycle) and mortality - that is, mortality is higher when the economy is doing better and unemployment is lower. That might seem surprising at first, but Atalay are able to confirm this relationship with their Australian data, but only for a subset of age groups:

Overall, we find no effect of unemployment on all-cause mortality... A 1 p.p. increase in the unemployment rate is associated with a zero impact on mortality (-0.02 percent and statistically insignificant). The effect is likewise statistically insignificant when estimated separately for men and women...

We observe that mortality is procyclical for the youngest age group (0-24 years) in the pooled sample (-1.8 percent, p-value<0.05) and for men (-1.9 percent, p<0.10) and women (-1.5, p-value<0.10) separately. Hence, there are fewer deaths in times of economic growth for the young. For those 25-64 years and 65 years of age and older, we do not observe any significant relationship.

Drilling down further into the five-year age groups, they find that it's essentially a result of men aged 20-39 years, as shown in their Figure 4, Panel A:

Notice that the only statistically significant bars are among the age groups 20-39 years. The corresponding figure for women shows no statistically significant effects. So, what is going on here? Atalay et al. next look at cause-specific mortality, and find that:

...higher unemployment is associated with fewer vehicle accident (road) deaths. An increase in the unemployment rate by 1 p.p. is significantly associated with a 6 percent decrease in transport accidents (p-value<0.05)... The number of lives saved are five time [sic] larger for men (73 fewer deaths) than for women (15 fewer deaths).

In other words, mortality is positively associated with the business cycle, that effect is concentrated among young men, and is driven by changes in the number of motor vehicle accidents. The mechanism seems clear - when the economy is doing better, people drive more, and that increases the risks of accidents, injuries and deaths. That is both interesting and plausible. However, Atalay et al. then probably over-reach in their conclusion:

Our findings allow us to propose an estimate for the likely impact of the recession associated with the pandemic and the Great Lockdown on mortality. If unemployment rates rise from the February 2020 rate of 5.1 percent to 10 percent as predicted by the Reserve Bank of Australia, we would expect almost 425 fewer deaths due to vehicle transport accidents. This reduction in the number of deaths is equivalent to approximately 30 percent of all transport accidents in 2017.

Although lockdowns did create a snap recession in Australia, and people drove less because they were confined to home, there is no reason to believe that the relationship observed over the period 1979 to 2017 would continue to hold. The pandemic period really is an out-of-sample event. However, we should take this as suggestive evidence in favour of lockdowns reducing mortality, and through a mechanism that isn't directly related to coronavirus.

It's the demography, stupid

I don't often write about macroeconomics. A lot of the research on macroeconomics is highly mathematical, and a lot of it is, frankly, voodoo. Macroeconomics has rightly taken a beating since the Global Financial Crisis for unrealistic models (see here and here, for example). One of the things that most surprises me about macroeconomic models is the absence of demography. So many models ignore the population, as if it's only something that matters in terms of the size of the labour force, and as a denominator to turn GDP into GDP per capita. But the population age structure matters. Labour force participation, and saving and borrowing behaviour, depend on the age structure of the population. So population ageing matters (for example, see this post).

So, I was interested to recently read this 2018 article entitled "The demographic deficit", by Thomas Cooley (New York University) and Espen Henriksen (BI Norwegian Business School), published in the Journal of Monetary Economics (ungated presentation version here). Cooley and Henriksen calibrate overlapping generations models for the U.S. and Japanese economies for 1990 and 2007, and include increasing life expectancy and the age distribution of the populations as key inputs into the model. As they explain:

Growth accounting shows that that growth differentials both across countries and over time are not only driven by TFP and capital accumulation, but labor supply on the extensive margin, labor supply on the intensive margin, and (obviously) population growth. One straightforward way in which demographics impact changes in aggregate economic activity is through their impact on aggregate factor supply. Data show that households steadily decrease labor supply both on the intensive and extensive margin in the latter part of their working lives. This is in contrast to the usual assumption in overlapping-generations models, that households supply labor inelastically until retirement age. Changes in life expectancy and cohort distributions will therefore affect both labor market participation and average hours worked. Faced with in- creases in life expectancy individuals need to provide for more years in retirement during their working life. In addition, aging populations means more people will be in their highest savings years. This may lead to changes in aggregate capital supply. Lastly, demographic change affect the composition of the work force and its productivity. Changes in the average efficiency of the individuals working will manifest itself in changes in TFP.

Cooley and Henriksen show with their calibrated models that:

...about 1/6 of the level of growth for both United States and Japan, net of population growth, can be accounted for by changes in life expectancy and in the cohort distributions.

There is clearly a non-trivial share of economic growth attributed to demographic change. And that means that demographic change might help explain some of the growth slowdown since the Global Financial Crisis. Older populations have fewer people working, workers working fewer hours, and less saving. All of those contribute to slower economic growth. The contribution of demographic change to slowing economic growth is complementary to the other recent explanations for the growth slowdown, including secular stagnation (from Larry Summers), and technological slowdown (from Robert Gordon).

Cooley and Henriksen's article is complemented by a comment in the same issue of the journal, by Etienne Gagnon, Benjamin Johannsen, and David López-Salido (all Federal Reserve Board). If you don't understand Cooley and Henriksen's article, then Gagnon et al. provide a summary that is actually much simpler to understand. However, what struck me was this from the conclusion of the comment:

As populations in the advanced economies continue to age, understanding the consumption and labor supply decisions of older workers is becoming an increasingly urgent task for deriving projections of aggregate variables. Henriksen and Cooley’s paper is a most-welcome step in this direction.

That is both good, and bad. It's good that macroeconomics is taking steps towards meaningfully including demographic change in its models, but bad because we're not already there.

Read more:

• Will population ageing lower economic growth?