Reports have been filtering out recently of a study that found a relationship between US unemployment rates and deaths due to opioid use. The Washington Post reported on these results, suggesting that there is a connection between unemployment and death due to “diseases of despair” (their words), and citing the unfortunate Case and Deaton study that found increasing mortality rates among non-hispanic whites in the USA. The implication is that some kind of post-2008 economic depression-related despair has driven the white working class to drugs, with an attendant high death toll. This is particularly poignant in light of the recent election, since some of the states (like West Virginia) that voted heavily for Trump are also heavily affected by opioid abuse. The implication here is that the economic despair supposedly driving Trump voting is also driving high mortality in these communities, which have also supposedly been hollowed out by globalization, immigration and Democratic neglect (only Democrats can neglect poor white people; Republicans ride in to save them with trickle down economics while Democrats abandon them for groovy inner-city Black Lives Matter activists and funky Chicago law professors). But is any of this true?

The news reports are based on the findings of a study by Hollingsworth, Ruhm and Simon, Macroeconomic conditions and opioid abuse, published in my bete-noir, the National Bureau of Economic Research (NBER) working papers series. This is where economists publish their brain farts before they are shot down in peer review, and this paper is a typical economist brain fart. This study suffers from the usual problems of NBER papers: it has a ludicrous model, uses the wrong modeling approach, does some dubious data manipulation, and probably isn’t representative. Worse still, the study is based on a failed and useless model of drug addiction that eschews a balanced understanding of drug addiction in favour of a lazy just-so story about the causes of drug addiction that has no basis in reality. I will briefly discuss the modeling problems that make this study useless, and then discuss in more detail the problem of its underlying theoretical structure.

Modeling problems with the study

The study is a classic example of how economists just cannot handle data well. First, the authors have presented a ludicrous model which has an enormous number of explanatory variables – one for every county in their data set, one for every year, and an additional term for the combination of states and years – which means that the model has a huge number of terms to be estimated. Worse still, they do not include age or sex in the model, so they don’t adjust at all for differences in age structure between different counties and states or ethnic groups. Non-heroin opioid addiction in the USA seems to be clustered in rural whites, and probably reflects addiction pursuant to pain relief for real health problems. If so the problem is likely more prevalent in older groups (which have higher levels of chronic pain) who may well be more vulnerable to early death – so adjustment for age is important in these studies. The authors find mortality rates in whites increasing much faster than blacks or hispanics but this could well be because these groups are younger and thus earlier into their drug addiction, or simply less likely to die. This complexity is further compounded by the authors decision to impute drug types to drug-related deaths where the drug is not specified – they simply statistically estimate what drug caused the death, which makes all their results highly vulnerable to the quality of the model by which they impute 30% of all drug-related deaths. So the authors have estimated a model with a huge number of terms and have not properly adjusted for the age structure of the population. This is extremely important, since the CDC has shown that opioid-related mortality is much higher in older people, and if areas with many old people also have high unemployment there will be a spurious relationship between unemployment and mortality if age is not adjusted for.

Incidentally, this paper gives completely different crude opioid mortality rates to the CDC, probably because it uses a subset of states with unusually high mortality rates. So there is a huge generalizability problem right there.

The other big problem with the model is that, of course, being economists, the authors do not use the correct modeling approach. Opioid mortality is a rare even with very small numbers of deaths when disaggregated by race at the county level – even the authors admit that many of their data points have zero deaths – but the authors have chosen to divide the counts of mortality by the population of the area, to get crude rates, and then to model these using ordinary least squares linear regression. As I have repeatedly said here, OLS regression is completely the wrong method to use on data that is constrained. In this case the data is constrained to be greater than or equal to zero, and is likely very close to zero in most cases. OLS regression assumes a completely different probability structure to the correct method, Poisson regression, and applying OLS regression to rates means that you are assuming all zero rates have the same probability. In contrast, a Poisson regression adjusted for population size models a zero count with a different probability depending on the population size, so a zero event in a large population has a different meaning to a zero event in a small population. It also models a non-linear relationship between the underlying death rate and the unemployment rate, which is crucial to understanding how the underlying death rate is related to unemployment. By not using a Poisson regression for rare events the authors have mushed together a bunch of very different mortality patterns as if they were all the same, and completely changed the nature of the relationship between unemployment and mortality.

Big no no!

So the modeling is completely flawed, but this isn’t the worst part of this study. The worst part of this study is that the underlying theory is completely flawed.

Opioid use is not a disease of despair

The fundamental problem with this model is the assumption that macroeconomic conditions drive opioid use. Figure 1 shows the observed and modeled number of monthly deaths due to heroin overdose in New South Wales, Australia between 1995 and 2003, taken from Degenhardt and Day, Impact of the Heroin Shortage: Additional Research (I prepared this figure for this technical report).

Figure 1: Monthly observed and modeled heroin overdose deaths in New South Wales, 1995-2003

This figure shows a clear rapid peak occurring in 1999, followed by a gradual decline and then a sudden downward step in January 2001. This downward step is even more evident in heroin possession offences (Figure 2, also prepared by me, from Gilmour et al, Using intervention time series analysis to assess the effects of imperfectly identifiable natural events: A general method and example, BMC Medical Research Methodology 2006; 6:16).

Figure 2: Observed and modeled trend in heroin possession offences in New South Wales, 1995-2003

Is it really conceivable that trends in unemployment were so intense over the 8 years of this data series that they caused heroin possession offences to more than double, and heroin mortality to double, within 2 years, and to then decline by 50% before halving in one month? What are the macroeconomic effects driving this phenomenon? In fact youth unemployment in NSW declined consistently over the 1990s, and was at a historic low when heroin mortality peaked. What changed over the 1990s was the availability of heroin, which was flooding the market in the mid-1990s; and what changed in 2001 was that new models of drug interdiction and cooperation between police agencies led to unprecedented success in fighting drug traffickers, so that in the early ’00s they pulled out of Australia in favour of easier targets. The result was a sudden precipitous decline in heroin availability, a massive increase in cost, a temporary increase in street-based sex work and cocaine use, and a rapid flight of young people from the market. This occurred against a backdrop of readily available harm reduction services and widespread, free methadone treatment, to which many drug users fled when the price skyrocketed.

The reality is that drug addiction patterns are driven primarily by availability of the drug and availability of treatments for drug addiction. Far from being a “disease of despair” as the Washington Post described it, with patterns of use determined by social dislocation and poverty, heroin addiction is a disease of opportunity, driven primarily by the presence of the drug, its ease of use, and the economic potential to purchase it. There is no relationship between drug use and unemployment or poverty, and we have known this since Robin Lee did her groundbreaking work on returning heroin addicts after the Vietnam war. I suspect the truth of the American opioid epidemic is much more boring, and much more difficult to explain, than unemployment: It is a problem of availability. I don’t know what causes that problem but my guess is that sometime in the 2000s legislative changes made opioids much more easily available. In 2003 the Medicare Prescription Act was passed, and my guess is that it made it much easier for middle-aged poor people to get access to pain relief – pain relief they desperately needed for a wide array of real problems. With access to affordable opiates but with no corresponding access to specialist pain management professionals a cohort of middle-aged workers became addicted to opioids, and in the subsequent 10 years they started dying. It’s a boring health policy explanation for a terrible problem, and it can only be fixed by improvements in quality of care, access to specialists, and careful attention to modern strategies for pain relief.

Unfortunately this story doesn’t fit with a narrative – popular on left and right – of drug addiction as a disease of despair. In this narrative the left sees drug addiction as a product of an alienating and destructive society, best solved by improvements in welfare and labour rights, while the right sees drug addiction as a consequence of unemployment and poverty, which are best solved by getting everyone into work (since good welfare programs are anathema to the right). For economists both of these stories show the primacy of economics as a driver of social problems, and make a good just so story. But the reality of opioid addiction is that it is a complex health policy problem best solved by careful attention to the way that opioids are dispensed and pain is managed. True, this policy prescription requires potentially quite radical changes in the way that doctors approach chronic illness, poverty and occupational health – but it’s completely boring outside of health policy. Stories of a “generation left behind”, forced to vote for Trump because of the carnage sweeping through their blighted communities, are much more interesting than “oh yeah, we made dangerous drugs cheaper and didn’t train doctors how to manage them.”

This article and the interest it drove are another example of two pernicious problems in modern debate: economists can’t be trusted with health data, and journalists are too quick to believe economists. When this is tied with a problem that is easily amenable to sensationalism and patronizing assumptions, of course you get a narrative that is completely divorced from the truth. In this case we don’t know what the truth of the numbers is, since the economists in question made a model so bad it has no bearing on the truth; and we were led into believing that this model could ever explain the very real problems facing these communities by credulous economists and journalists all too willing to believe lazy stereotypes about drug users and drug use.

Let’s score that as another failure for two of the worst professions, and hope we can make some real changes to prescription laws and pain management so that the people affected by this problem can find better, safer ways of managing their chronic pain. And please, please please, can economists please stop touching health data until they learn a method other than OLS regression?

Save

Save