Anne Case and Sir Angus Deaton are two economists of high standing (both are professors at Princeton and Deaton won the Nobel Prize in Economics in 2015). They also happen to be married to each other. They have recently been famous for statistical analyses of the stopped decline or even increase in the mortality rates for middle-age (and perhaps younger) non-Hispanic white Americans when those rates are still declining for both non-Hispanic blacks and Hispanics in the United States and for whites in European countries and Canada. Their first article on the topic came out in 2015, and a Brookings Institute conference paper (or a conference draft) was released only a few weeks ago, in March of 2017.
The latter paper concludes that the increased mortality of middle-aged non-Hispanic whites applies to both men and women and that it is completely attributable to rising mortality among those non-Hispanic whites whose highest education level is a high school degree or less.
It's that 2017 working paper I want to talk about here, and especially the parts of it which cause me to ask questions. Thus, this post is one of criticisms.(1)
Before I launch into it I want to stress that I admire the contribution Case and Deaton have made by both having the ability to get their message heard in the public conversations and by what they have contributed to the wider epidemiological and statistical literature on the topic of mortality rates and how they change over time.
On that count I have nothing but admiration for their work. Still, presenting a working paper to the world at large is a little like Coco Chanel presenting a half-finished dress, cut, pasted and pinned together, to the woman who ordered it as the finished couture creation. Working papers are not subjected to rigorous peer review, and that means that they rather resemble the pieces of the dress basted or pinned together at the first fitting, not the final dress. In other words, there's work still to be done on the Case-Deaton conference paper and its presentation.
My questions or criticisms fall into three groups. The first is about general methodological and presentation concerns, the second about the racial and ethnic comparisons as they appear in the Case-Deaton working paper, and the third about the way differences between male and female mortality rates are sometimes ignored, sometimes brought forward in inconsistent ways.
General Methodological And Presentation Questions
1. Picking a Particular Age Group for More Analysis
The 2015 Case-Deaton article on the increasing mortality rates of middle-aged whites focused most of its attention on the age group between 45 and 54 and how its mortality rate changed between 1990 and 2013.
Andrew Gelman, a statistics professor at Columbia University, wrote a response to that 2015 article asking what the impact of rising average age inside that 45-54 bin might be when we compare different cohorts of individuals aged 45-54 over time. After all, the aging of the baby boomers could mean that the average person now in the age category 45-54 might be closer to the 54-end than was the case, say, ten years ago. And increasing average age would by itself explain at least a part of the rising mortality rates Case and Deaton found.
The 2017 Case-Deaton conference paper acknowledges this criticism:
Early commentary on our work focused on our lack of age adjustment within the age group 45-54 (Gelman and Auerbach 2016). Indeed the average age of white non-Hispanics (WNH) aged 45-54 increased by half a year between 1990 and 2015 so that part of the mortality increase we documented is attributable to this aging.
The solution to this dilemma in the 2017 paper appears to be the use of narrower age categories. The one which the authors highlight is the 50-54 group, though data for other age groups are present in the tables (2), and the 2015 selection of 45-54 is also widely used. It seems that the 50-54 group was picked because it does not show aging within the group over time. It is difficult to tell, because the average ages are not shown for, say, the 45-49 group.
2. Lagged Selection
The 2017 Case-Deaton conference paper analyzes white non-Hispanic middle-age mortality rates separately by three educational categories (3):
Those whose highest educational achievement is a high school diploma or less, those who have some college attendance, and those whose highest educational credential is a BA degree or higher.
It is the lowest of those categories which is most associated with increasing middle-age mortality rates, Case and Deaton argue. But remember that the analysis compares people in a particular age group over time (roughly, from 1999 to 2015). If the composition of those educational categories change over time, then the results might suffer from what is called lagged selection. This quote explains the context:
It’s easy to see how lagged selection bias could apply to a mortality study that breaks down by educational attainment. In fact, the term was created to address this exact case. In 2012, a paper (usually referred to by its lead author S. Jay Olshansky’s last name) led to similar headlines as the new Case and Deaton one, but a few public-health researchers thought something looked off about the findings, which included rapidly increasing mortality for whites without high school degrees.
In the International Journal of Epidemiology, researchers Jennifer B. Dowd and Amar Hamoudi suggested the Olshansky results could reflect increasing high school graduation rates more than increasing mortality. As a greater proportion of Americans finish 12th grade, lagged selection bias (a term that Dowd and Hamoudi coined) means the demographic of non-completers shrinks over time, and the longitudinal comparison gets less valid. “In terms of mortality risk, those excluded from high school in the early part of the 20th century are not comparable with those excluded from high school a generation later,” they write, “because those left behind by the high school expansions in mid-century likely had childhoods that were more disadvantaged along many dimensions, and so were at higher mortality risk all along.”
Emphasis is mine.
How can a study avoid lagged selection? One way is to replace the absolute education classes by percentages or percentiles. For example, rather than analyze the category of individuals with only a high school diploma or less, the researchers could replace that category with individuals who belong to the bottom 25% on the basis of their educational achievement:
There are ways to guard against lagged selection bias if you’re conscious about it. A group of researchers based out of the Population Studies Center at the University of Michigan has been working with some of the same material, and they took up Dowd and Hamoudi’s challenge and the Olshansky data sets. The Michigan group was able to reproduce the Olshansky results, but they then tried to account for the effect of lagged selection bias. They used percentile of educational attainment rather than simple attainment — based on Census data, a white American in the 25th percentile would have a 10th-grade education if they were born in 1935, but a high school diploma if they were born after 1950.Case and Deaton did not use that solution. Instead, they argue that there is no lagged selection: that the shares of the three educational categories they employ have not essentially changed since the early 1990s among non-Hispanic whites aged between 45 and 54.
It’s a smart way to account for the expansion of education access, and, when they tried it, they found that the Olshansky findings looked overstated. At the 25th percentile of educational attainment, they saw a 1.2-year decrease in life expectancy for white women between 1990 and 2010 (compared to 3.8 years with the Olshansky “no high school graduation” measure), and a small increase for white men (0.4 years vs. -2.2). The media did not exactly trip over itself delving into the new data.
While that may well be true (or not), there is no reference attached to this argument, and it still doesn't explain why so much of the analysis concerns the group between 50 and 54. Neither is it clear if the same argument would apply to the educational classes when they are applied to black non-Hispanics and Hispanics.
3. Aggregation vs. Disaggregation
Most of the Case-Deaton analysis takes place on a fairly aggregate level. That this might not be the best approach in looking for the explanations to a possibly rising mortality rates among non-Hispanic whites is suggested by these graphs by Jonathan Auerbach and Andrew Gelman:
Although the graphs I picked as examples are not for a particular education category and only show white female mortality rates, it's fairly clear that the trend lines show great regional variation. Any attempt to explain the underlying causes of rising mortality should take this geographical (and urban vs.rural) variation into account (4). Case and Deaton discuss several possible explanations for their findings but do not employ regional data in their testing. In this particular example aggregation can mean that we lose sight of the trees while contemplating the forest.
4. The Misery Index
Both Case and Deaton papers and almost all media summaries of their research focus on what might be called "deaths of despair:" suicides and deaths from drug and alcohol abuse. They argue that the increase in the deaths of despair, combined with a slower decrease in mortality attributed to heart disease or cancer has the net effect of increased or stagnant mortality rates for middle-aged non-Hispanic whites.
Both the 2015 and the 2017 papers supplement this with morbidity data, in particular self-reported health evaluations. These show the expected pattern between educational categories, with those in the lowest category reporting increasingly worse health, more mental suffering and more back pain, when compared to those non-Hispanic whites with more education.
The corresponding data on non-Hispanic blacks and Hispanics, disaggregated by the three educational levels, doesn't seem to have been used in the study, or in the 2015 paper. This is unfortunate, because Case and Deaton argue that greater self-reports of chronic pain, say, are likely to be linked to greater likelihood of suicide, one of the deaths of despair, which are more common among non-Hispanic whites than non-Hispanic blacks. An analysis of the morbidity data for other demographic groups could have provided proof for that argument, or not.
Some of the tables in the 2017 Case-Deaton conference paper should have been constructed with more care, lest the readers feel that they are asked to look away from the fact that mortality rates are still higher for non-Hispanic blacks in the United States than for non-Hispanic whites (5).
Take Figure 1.1 from the paper:
The red line in the figure is apples, the other lines are pears. Only the red line limits the mortality rates to a sub-group inside a demographic group: Those with at most a high school education.
That line is then to be compared to the overall mortality rates for all non-Hispanic whites, non-Hispanic blacks and Hispanics. The last three groups include people with some college and people with college degrees.
Figure 1.2 in the paper presents a better comparison by showing the all-cause mortality rates for both non-Hispanic whites and non-Hispanic blacks with only a high school degree or less:
It shows that the mortality rates are approaching each other, but it also shows that the black rates are still higher (6).
Another table which would benefit from reworking is Figure 2.3 in the paper. It shows the median household income per household member for individuals between the ages 45 and 54. One is to read the figures by using different vertical axes for blacks and whites: The white income levels are on the left axis and the black income levels on the right axis:
Similar arguments apply to the following two graphs in the Appendix Figure 5:
Case and Deaton mention white non-Hispanic women's higher mortality rates at the beginning of their 2017 paper when they respond to criticisms made about the 2016 article by Gelman and Auerbach (about the need to adjust the mortality rates for average age within each age category):
Gelman and Auerbach’s age-adjusted mortality rates for WNHs (white non-Hispanic whites) in the 45-54 year age group show that the increase in all-cause mortality is larger for women, a result we have confirmed on the data to 2015 (36 per 100,000 increase for women, 9 per 100,000 increase for men,between 1998 and 2015, age-adjusted using 2010 as the base year,with little variation in the increases across different base years).
They then promise to analyze this sex difference in closer detail. But as far as I can tell that analysis is fairly short and doesn't mention the excess female mortality rates as something to be explained. Rather, much of the analysis focuses on the "deaths of despair" which show greater male mortality rates from those causes:
Case and Deaton promise to address the overall mortality differences in a later paper. I eagerly look forward to it, in order to learn more about the women in that age group.
In this post I have barely skimmed the long part of the paper which speculates about the possible causes for an increased mortality rate among non-Hispanic whites with few education qualifications, though the table (above) on incomes belongs to that section. The theories are interesting and I recommend that readers fascinated by the question read them.
In this post I want to draw out one aspect of the theories that Case and Eaton use to explain their hypothesis that the deaths of despair among middle-aged less educated whites are due to cumulative deprivation: The loss of well-paying jobs, then the loss of marriages and the simultaneous loss of traditional social and economic support, such as churches.
When Case and Deaton discuss marriage, they refer, among other things, to earlier work by David Autor, David Dorn and Gordon Hanson, who argue that working class men become less valuable in the marriage markets when manufacturing disappears, because they no longer can command higher incomes.
This lack of "high quality males" in the marriage markets is then used to explain the decline of marriage among working-class whites and also the increase in single parent (female) households by stipulating that traditionally-minded women abstain from marriage and childbirth altogether under such conditions and that less traditionally-minded women choose to have children without getting married to a "lower quality male."
The Autor-Dorn-Hanson model of heterosexual marriage reminds me of the weirder types of evolutionary psychology arguments. They both share the assumption that the decision to marry or not is made by women alone (7), that the value of a potential husband depends on his ability to bring home the bacon (that his income must be higher than hers) and on him not using drugs and alcohol. That's it.
In that framework the role of men in marriage is largely as sources of income. Even evolutionary psychologists argue that men have different mating strategies, including the strategy of companionship and sharing child-rearing duties, but apparently not all American economists agree.
A model which assumes that it is the women, alone, who make decisions about cohabiting or marriage and about having children, and which assumes that the value of men in the marriage markets is almost solely income-based will find it very important to analyze men's labor market participation in contexts such as the miseries of American working class whites, and that is what Case and Deaton also do.
Thus, we get graphs about non-Hispanic white men's labor market participation rates (Figure 3.1), but not about non-Hispanic white women's labor market participation rates, and we get graphs about the earnings development for those same men (Figure 3.2) but not for the women.
We also get this extremely intriguing quote:
Lower wages made men less marriageable, marriage rates declined, and there was a marked rise in cohabitation, then much less frowned upon than had been the case a generation before.Emphasis is mine.
Figure 3.2 hows that, beyond the cohort of 1940, men and women with less than a BA degree are less likely to have ever been married at any given age. Again, this is not occurring among those with a four-year degree. Unmarried cohabiting partnerships are less stable than marriages. Moreover, among those who do marry, those without a college degree are also much more likely to divorce than are those with a degree.The instability of cohabiting partnerships is indeed their raison d’être, especially for the women, who preserve the option of trading up, see also Autor, Dorn and Hansen (2017)
I waded through the Autor, Dorn and Hansen paper twice, to try to find the reference to cohabiting partnerships as something women choose so that they have the option of trading up, but I couldn't find it. Perhaps the statement I emphasized is by Case and Deaton?
If so, what about the benefits of cohabitation for men who can also preserve the option of trading up to a more recent model? Or what about the disadvantages of cohabitation for women with children who are much less likely to have the option of "trading up," with children in tow, whatever "trading up" might mean.
I write so extensively on the use of a particular framework in the analysis of marriage choices because it shows that Case and Eaton have a particular tilt in how they analyze mortality data on both men and women.
Despite the contents of this post, the Case-Deaton analyses are of value both because they have made a particular health problem more widely noticed and because the ethnic, racial and international comparisons do pose questions which require answers:
Why are the mortality rates of non-Hispanic blacks and Hispanics in the United States still declining while the non-Hispanic white rates have either stayed constant, increased or declined less (depending on whose study we rely on)? And why are European countries and Canada still experiencing declining mortality rates when the economically and culturally fairly similar American non-Hispanic whites are not?
Much more work is needed to answer those questions, and that work requires the cooperation of epidemiologists, statisticians and other experts in the measurement of health outcomes.
(1) A major caveat: I am not an epidemiologist and neither am I familiar with the data sources Case and Deaton used. Thus, I quote others when it comes to those areas. Most of my own criticisms are about the presentations in the working paper and about the implicit underlying modeling in the third group of criticisms. And many of my concerns are about the need to see more evidence for certain assertions before I would be happy with the basic statistical findings.
(2) Mortality rates for the younger age groups should be interpreted and compared across groups with great care for at least two reasons:
First, overall mortality is low among the young after the first year of life. This means that various one-time shock events can cause a large-seeming statistical blip in the tables, simply because we begin from a low basic mortality rate.
Second, statistical analyses based on comparing individuals with different education levels must take into account the fact that for ages below 45 the fraction of both black and white non-Hispanics whose highest educational achievement is a high school degree or less has declined, so that the composition of that group may have changed toward those with greater health problems.
(3) Much of the media decided to popularize the Case-Deaton results by equating the lowest educational group with working class, perhaps to provide a hook to the recent election chatter over the rebellion of the white working class voters who went for Trump in certain trigger states But the two concepts are not identical.
(4) If you are interested in learning more on the geographical variation, Auerbach and Gelman have collected mortality rates from many states separately be race, age, sex and ethnicity.
(5) The rates are the lowest for Hispanics among the three demographic groups that are compared in the paper.
(6) Keep in mind that any lagged selection might or might not affect blacks and whites differently.
(7) Autor, Dorn and Hansen write:
We consider a setting where unmarried women have a preference to become married mothers but face uncertainty about the availability of high-quality men who may serve as marital partners.12
A substantial literature documents that the marriage decision tends to follow the fertility decision: upon becoming pregnant, a woman may choose to marry the child’s father but absent pregnancy would not elect marriage.13
We impose this setting on decision-making by assuming that women choose to remain childless, to have a child and marry, or to become a single mother. Removing the option of marriage without children narrows the generality of the model but is not restrictive empirically since nearly 90% of women ages 18-39 are either mothers, or unmarried without children.14