UPDATE: Dr. Monnat has left a comment pointing out that I made a major error in reading her methods (I assumed she used non-standardized rates but in the methods she specifies that she did). So I have removed one criticism of her paper and modified another about regression. This doesn’t change the thrust of my argument (though if Dr. Monnat is patient enough to engage with more of my criticisms, maybe it will!)

Since late 2016 a theory has been circulating that Donald Trump’s election victory can be related to the opioid epidemic in rust belt America. Under this theory, parts of mid-West America with high levels of unemployment and economic dislocation that are experiencing high levels of opioid addiction switched votes from Democrat to Republican and elected Trump. This is part of a broader idea that America is suffering an epidemic of “deaths of despair” – deaths due to opioids, suicide and alcohol abuse – that are part of a newfound social problem primarily afflicting working class white people, and the recent rapid growth in the rate of these “deaths of despair” drove a rebellion against the Democrats, and towards Trump.

This theory is bullshit, for a lot of reasons, and in this post I want to talk about why. To be clear, it’s not just a bit wrong: it’s wrong in all of its particulars. The data doesn’t support the idea of a growing death rate amongst white working class people; the data does not support a link between “deaths of despair” and Trump voting; there is no such thing as a “death of despair”; and there is no viable explanation for why an epidemic of “deaths of despair” should drive votes for Trump. The theory is attractive to a certain kind of theorist because it enables them to pretend that the Trump phenomenon doesn’t represent a deep problem of racism in American society, but it doesn’t work. Let’s look at why.

The myth of rising white mortality

First let’s consider the central framework of this story, which is the idea that mortality rates have been rising rapidly among middle-aged whites in America over the past 20 years, popularized by two economists (Case and Deaton) in a paper in PNAS. This paper is deeply flawed because it does not adjust for age, which has been increasing rapidly among white Americans but not non-white Americans (due to differential birth and migration patterns in earlier eras). Case and Deaton studied mortality in 45-54 year old Americans, differentiating by race, but failed to adjust for age. This is important for surprising reasons, which perhaps only epidemiologists understand, and we’re only figuring this out recently and slowly: ageing is happening so fast in high-income countries that even when we look at relatively narrow age categories we need to take into account the possibility that the older parts of the age category have a lot more people than the younger parts, and this means that even the small differences in mortality between say 53 year olds and 45 year olds can make a difference to mortality rates in the age category as a whole. If this seems shocking, consider the case of Japan, where ageing is so advanced that even five year age categories (the finest band of age that most statistical organizations will present publicly) are vulnerable to differences in the population. In Japan, the difference in the size of the 84 year old population to the 80 year old population is so great that we may need to adjust for age even when looking at narrow age categories like 80-84 years. This problem is a new challenge for epidemiologists – we used to assume that if you reduce an analysis to a 10 or 15 year age category you don’t need to standardize, because the population within such a band is relatively stable, but this is no longer true.

In the case of the Case and Deaton study the effect of ageing in non-hispanic white populations is so great that failure to adjust for it completely biases their results. Andrew Gelman describes the problem  on his blog and presents age-adjusted data and data for individual years of age, showing fairly convincingly that the entire driver of the “problem” identified by Case and Deaton is age, not ill health. After adjustment it does appear that some categories of white women are seeing an increasing mortality rate, but this is a) likely due to the recent growth of smoking in this population and b) not a likely explanation for Trump’s success, since he was more popular with men than women.

White people are dying more in America because they’re getting older, not because they have a problem. I happen to think that getting older is a problem, but it’s not a problem that Trump or anyone else can fix.

The myth of “deaths of despair” and Trump voting

Case and Deaton followed up their paper on white mortality with further research on “deaths of despair” – deaths due to opioid abuse, suicide and alcohol use that are supposedly due to “despair”. This paper is a better, more exhaustive analysis of the problem but it is vulnerable to a lot of basic epidemiological errors, and the overall theory is ignorant of basic principles in drug and alcohol theory and suicide research. This new research does not properly adjust for age in narrow age groups, and it does not take into account socioeconomic influences on mortality due to these conditions. But on this topic Case and Deaton are not the main offenders – they did not posit a link between “deaths of despair” and Trump voting, which was added by a researcher called Shannon Monnat at Pennsylvania State University in late 2016. In her paper, Monnat argues for a direct link between rates of “deaths of despair” and voting for Trump at the county level, suggesting that voting for Trump was somehow a response to the specific pressures affecting white Americans. There are huge flaws in this paper, which I list here, approximately in their order of importance.

  • It includes suicide: Obviously a county with high suicide mortality is in a horrible situation, which should be dealt with, but there is a big problem with using suicide as a predictor of Trump voting. This problem is guns. Uniquely among rich countries, the US has a very high prevalence of gun ownership and guns account for a much larger proportion of suicides in America than elsewhere – more than half, according to reputable studies. And unfortunately for rural Americans, the single biggest determinant of whether you commit suicide by gun is owning a gun – and gun ownership rates are much higher in counties that vote Republican. In America suicide is a proxy for gun ownership, not “despair”, and because gun-related suicide depends heavily on rates of gun ownership, inclusion of this mortality rate in the study heavily biases the total mortality rate being used towards a measure of gun ownership rather than despair.
  • It uses voting changes rather than voting odds: Like most studies of voting rates, Monnat compared the percentage voting for Trump with the percentage voting for Romney in 2012. This is a big flaw, because percentages do not vary evenly across their range. In Monnat’s study a county that increased its Republican voting proportion from 1% to 2% is treated exactly the same as a county that went from 50% to 51%. In one of these counties the vote doubled and Trump didn’t get elected; in the other it increased by 2% but Trump got elected. It’s important to account for this non linearity in analysis, but Monnat did not. Which leads to another problem …
  • It did not measure Trump’s success directly: In a first past the post electoral system, who wins is more important than by how much. Monnat used an ordinary least squares model of proportions voting Trump rather than a binomial model of Trump winning or losing, which means that meaningless small gains in “blue” states[1] had the same importance as small gains in “red” states that flipped them “blue”. This might not be important except that we know Trump lost the popular vote (which differences in proportions measure) but won the electoral college (which more closely resembles binary measures of win/lose). Not analyzing binary outcomes in a binomial model suggests you don’t understand the relationship between statistics and the political system you live in, i.e. your analysis is wrong.
  • It did not incorporate turnout: A 52% win for Trump can reflect two things – a change in attitude by 2% of the voters, or a non-proportionate increase in the number of people who chose to turn out and vote. If you analyze proportions (or differences in proportions) you don’t account for this problem. In addition, you don’t adjust for the overall size of the electorate. If you analyze proportions, an electorate where 52 people voted Trump and 48 people voted Clinton is given the same weight as an electorate where 5200 people voted Clinton and 4800 people voted Trump. If you use a proper binomial model, however, the latter electorate gets more weight and is implicitly treated as more meaningful in the assessment of results. A reminder of what is fast becoming a faustusnotes rule: the cool kids do not use ordinary least squares regression to analyze probabilities, we always use logistic regression.
  • It did not present the regression results: Although Monnat reports regression results in a footnote, the main results in the text are all unadjusted, even though in at least some states the impact of economic factors appears to eliminate the relationship with mortality rates. Given that people who own guns are much much more likely to vote Republican, and the main predictor variable here incorporated suicide, adjustment for gun ownership might have eliminated the effect of “deaths of despair” entirely. But it wasn’t done as far as I can tell, and wasn’t shown.
  • It did not adjust for trends: Monnat openly states in the beginning of the paper that “deaths of despair” have been rising over time but when she conducts the analysis she uses the average rate for the period 2006-2014. This means that she does not consider the possibility that mortality has been dropping in some counties and rising in others. A mortality rate of 100 per 100,000 could reflect a decline over the period 2006-2014 from 150 to 50 (a huge decrease) or an increase from 25 to 175. We don’t know, but it seems likely that if “deaths of despair” is an issue, it will have had more influence on electoral decisions in 2016 in counties where the rate has risen over that time than where it has declined. There are lots of policy reasons why the death rate might have increased or decreased, but whether these reflect issues relevant to Republican or Democrat policy is impossible to know without seeing the distribution of trends – which Monnat did not analyze[2].

So in summary the study that found this “relationship” between “deaths of despair” and voting Trump was deeply flawed. There is no such relationship in the data[3].

There is no such thing as a “death of despair”

This study has got a fair bit of attention on the internet, as have the prior Case and Deaton studies. For example here we see a Medium report on the “Oxy electorate” that repeats all these sour talking points, and in this blog post some dude who fancies himself a spokesperson for ordinary America talks up the same issue. The latter blog post has some comments by people taking oxycontin for pain relief, who make some important points that the “deaths of despair” crew have overlooked. To quote one commenter[4]:

I too am a long time chronic pain sufferer and until I was put on opiate medications my quality of life was ZERO. I’ve heard horror stories of people actually being suicidal because they can no longer deal with the constant pain. It took me two years before I realized I could no longer work as an account manager with a major telecom company. I was making decent money but leaving work everyday in pain. I finally started going to a pain management doctor who diagnosed me with degenerative disc disease. I had to go on medical leave and now am on SSDI. My doctor prescribed me opiates in the fall of 2006 and I’ve been on them ever since. I have to say, I totally AGREE with you. I don’t know how I would be able to manage without these medications. At least I’m able to clean my house now and now without being in horrible pain. I don’t know what I would do if suddenly I was told I could no longer be prescribed opiates.
Who is someone that will champion those of us who legitametly need these medications? Do we write to our senators?? I sure hope Trump takes into consideration our cases before kicking us all to the curb!

This person (and others) make the valid point that they are taking pain medication for a reason, and that they were in despair before they got hooked on opioids, not after. Unfortunately for these commenters, we now have fairly good evidence that opioids are not the best treatment for chronic pain and that they are very, very dangerous, but regardless of whether this treatment is exactly the best one for these patients they make the valid point that it is the treatment they got and it works for them. To use an Americanism, you can take the opioids from their cold dead hands. In stark contrast to other countries, a very large proportion of America’s opioid deaths are due to prescription drugs, not heroin, reflecting an epidemic of overdose due to legally accessible painkillers. It’s my suspicion that these painkillers were prescribed to people like the above commenter because they could not afford the treatment for the underlying cause of their pain, because America’s healthcare system sucks, and these people then became addicted to a very dangerous substance – but in the absence of proper health insurance these people cannot get the specialist opioid management they deserve. America’s opioid epidemic is a consequence of poor health system access, not “despair”, and if Americans had the same health system as, say, Frenchies or Britons they would not be taking these drugs for more than 6 months, because the underlying cause of their condition would have been treated – and for that small minority of pain patients with chronic pain, in any other rich country they would have regular affordable access to a specialist who could calibrate their dose and manage their risks.

The opioid death problem in America is a problem of access to healthcare, which should have been fixed by Obamacare. Which brings us to the last issue …

There is no theory linking opioid addiction to voting Trump

What exactly is the theory by which people hooked on oxycontin are more likely to vote Trump? On its face there are only two realistic explanations for this theory: 1) the areas where oxycontin is a huge problem are facing social devastation with no solution in sight, so vote for change (even Trump!) in hopes of a solution; or 2) people who use drugs are arseholes and losers. Putting aside the obvious ecological fallacy in Monnat’s study (it could be that everyone in the area who votes for Trump is a non-opiate user, and they voted Trump in hopes of getting the druggies killed Duterte-style, but the data doesn’t tell us who voted Trump, just what proportion of each area did), there are big problems with these two explanations even at the individual level. Let’s deal with each in turn.

If areas facing social devastation due to oxycontin are more likely to vote Trump, why didn’t they also vote Romney? Some of these areas were stronger Obama voters in 2012, according to Monnat’s data, but opioid use has been skyrocketing in these areas since 2006 (remember Monnat used averages from 2006-2014). The mortality data covers two election cycles where they voted Obama even though opioid deaths were rising, and suddenly they voted Trump? Why now? Why Trump and not Romney, or McCain? It’s as if there is something else about Trump …

Of course it’s possible that oxycontin users are racist arseholes – I have certainly seen this in my time working in clinics providing healthcare to injecting drug users – but even if we accept such a bleak view of drug users (and it’s not true!) the problem with this theory is that even as opioid use increases, it remains a tiny proportion of the total population of these areas. The opioid users directly cannot swing the election – it has to be their neighbours, friends and family. Now it’s possible that a high prevalence of opioid use and suicide drives people seeing this phenomenon to vote Trump but this is a strange outcome – in general people vote for Democrats/Labour in times of social catastrophe, which is why they voted Obama to start with – because he promised to fix the financial crisis and health care. There has to be some other explanation for why non-opioid using people switched vote in droves to Trump but not Romney. I wonder what it could be?

American liberals’ desperate desire to believe their country is not deeply racist

The problem is, of course, that Trump had a single distinguishing feature that no one before him in the GOP had – he was uniquely, floridly racist. Since the election this has become abundantly clear, but for Donnat writing in late 2016 I guess it still seemed vaguely plausibly deniable. But the reality is that his single distinction from all other GOP candidates was his florid racism. Lots of people in America want to believe that the country they live in – the country that just 150 years ago went to war over slavery, and just 50 years ago had explicit laws to drive black people out of the economic life of the nation – is not racist. I have even recently seen news reports that America is “losing its leadership in the movement for racial equality.” No, dudes, you never showed any leadership on that front. America is a deeply racist nation. It’s racist in a way that other countries can’t even begin to understand. The reason Trump won is that he energized a racist base, and the reason his approval remains greater than 30% despite the shitshow he is presiding over is that a large number of Americans are out-and-out fascists, for whom trolling “liberals” and crushing non-whites is a good thing. That’s why rural, gun-owning Americans voted for Trump, and if the data were analyzed properly that fact would be very clear. Lots of people in America want to believe second- or third-order causes like the rustbelt or opioids, but the reality is staring them in the face: it’s racism. Don’t blame people with chronic pain, blame people with chronic racism. And fix it, before the entire world has to pay for the vainglorious passions of a narrow swathe of white America.


fn1: I refuse to take the American use of “blue” and “red” seriously – they get scare quotes until they decide that Republicans are blue and Democrats are red. Sorry, but you guys need to sort your shit out. Get proper political colours and get rid of American Football, then you’ll be taken seriously on the world stage. Also learn to spell color with a “u”.

fn2: I’m joshing you here. Everyone knows that Republicans don’t give a flying fuck if an electorate is dying of opioid overdoses at a skyrocketing rate, and everyone knows that the idea that Republicans would offer people dying of “deaths of despair” any policy solutions to their problem except “be born rich” is a hilarious joke. The only possible policy intervention that could have helped counties seeing an increasing opioid death rate was Obamacare’s Medicaid expansion, and we know republicans rejected that in states they controlled because they’re evil.

fn3: Well, there might be, but no one has shown it with a robust method.

fn4: I’m such a cynic about everything American that I really hope this commenter isn’t a drug company plant…

Advertisements