Death rates have been increasing for middle-aged white women, decreasing for men

Here’s the deal (data from CDC Wonder, age-standardized to a uniform distribution in the age range):

focus_group_3

Hoo boy. Looky here, something interesting: From 1999 to 2013, the death rate for middle-aged white women steadily increased. The death rate for middle-aged white men increased through 2005, then decreased.

Since 2005, the death rate has been rising for middle-aged white women and declining for middle-aged white men. Not by a lot—we’re talking a change of 4% over a decade—but this is what we see.

It’s funny. We’re so used to the narrative that things are getting worse for men, it’s so hard to be a guy in the modern era, etc. But in his particular case it’s the middle-aged women who are doing worse (relatively speaking; of course the absolute death rates remain much higher for men than for women, that’s just how things always are).

Background: Why age adjustment is needed

As Anne Case and Angus Deaton noted in a much-talked-about recent paper, the mortality rate among middle-aged white Americans has been roughly constant in recent decades, even while it’s dropped dramatically among other other groups and other countries.

Here’s the graph of the raw data of mortality among 45-54-year-old non-Hispanic whites in the U.S.:

focus_group_2

But that curve, which shows a steady increase since 1999, is wrong—or, should I say, misleading. As we discussed recently in this space (see here, here, and here), it can be tricky to interpret raw death rates binned across ages, especially in the U.S. What with the baby boom generation moving through, the average age in the 45-54 group crept up from 49.3 in 1999 to 49.7 in 2013.

An increase of 0.4 years might not sound like much, but mortality rate increases a lot by age—more than doubling between the ages of 45 and 54—so even a small shift in average age can cause a big shift in the observed trends.

Here’s what we get after adjusting for age:

focus_group

The flat pattern after 2005 is the sum of the increasing trend for women and the down slope for men.

What’s the point?

The published curves were biased because they did not correct for the changing age distribution within the 45-54 bin. When we make the adjustment we find something different: no longer a steady increase. And when we look at men and women separately, we find something more.

This update has not yet percolated through the news media.

For example, here’s Paul Krugman in the New York Times:

There has been a lot of comment, and rightly so, over a new paper by the economists Angus Deaton (who just won a Nobel) and Anne Case, showing that mortality among middle-aged white Americans has been rising since 1999.

Ross Douthat in that same newspaper yesterday:

Starting around the turn of the millennium, the United States experienced the most alarming change in mortality rates since the AIDS epidemic. . . . concentrated among less-educated, late-middle-aged whites.

Julia Belluz writes in vox.com about “the shocking rise in mortality rates among middle-aged white Americans.”

And Angus Deaton quoted in the Times the other day:

If we want to be more precise about the age range involved, we could say that for all single years of age from 47 to 52, mortality rates are increasing.

All these reports should be corrected to make it clear that the increase stopped in 2005. Since 2005, mortality rates have increased among women in this group but not men.

The age-aggregation bias did come up in this online NYT article, but the focus there was on the comparison between 1999 and 2013, so it did not come up that the net increase stopped after 2005, and that men and women’s mortality rates have been going in opposite directions since then.

Where does age adjustment make a difference?

First, I followed Deaton’s advice and downloaded death data from the CDC Wonder site. Second, I looked not just at the range 45-54 but also at the age decades before and after. Third, I looked at non-Hispanic whites, also at Hispanic whites, also at African Americans.

Then I computed the raw and age-adjusted death rates for each decade of age for each group, to get a sense of where age adjustment matters.

I plotted death rates since 1999, and here’s what I found:

effect_of_age_adj

It turns out that the only place where a lack of age adjustment really changes the story is . . . non-Hispanic whites aged 45-54. Too bad about that! But good that we checked.

Of course I may well have some “gremlins” in my analyses too. Anyone who wants can and should feel free to go to the data and find out what I garbled or missed.

Bring on the data

Finally, I broke down the numbers by sex and single year of age. Here’s what happened from 1999-2015 among all three ethnic groups:

death_rates_by_age_and_eth_1

death_rates_by_age_and_eth_2

death_rates_by_age_and_eth_3

And here’s a summary:

decades

That pattern among 45-54-year-olds? It was happening in the younger decade too.

One more time

Let me emphasize that this is all in no way a “debunking” of the Case and Deaton paper. Their main result is the comparison to other countries, and that holds up just fine. The place where everyone is confused is about the trends among middle-aged non-Hispanic white Americans.

The story being told is that there was something special going on, with an increase in mortality in the 45-54 age group. Actually what we see is an increasing mortality among women aged 52 and younger—nothing special about the 45-54 group, and nothing much consistently going on among men. Perhaps someone can inform Douthat and Krugman and they can modify their explanations accordingly. I’m sure they’ll be up to the task.

103 thoughts on “Death rates have been increasing for middle-aged white women, decreasing for men

    • Note that the state and county level data also do bolster the smoking hypothesis. In the more culturally liberal areas, white women both started and stopped smoking in earlier dates than in culturally conservative areas. (Men, and black and Hispanic women, starting smoking at earlier dates than white women, and the effect would not be reflected in th current data.) The state and county level data tend to show that the worst results for this data on white women are in the culturally conservative areas where the proportion of white women smokers experienced an increase in the relevant age groups.

      • I believe the CDC breaks out lung cancer as a specific cause of death, so direct smoking–>death can be charted.

        However, I would hardly be surprised though if smoking has indirect, broadly detrimental effects with uncertain lag periods. For example, I nagged my mother into stopping smoking the year after the Surgeon General’s 1964 report against cigarettes. But I wouldn’t be terribly surprised if her death several decades later was related to her having once been a smoker.

      • To summarize the single year of age graphs for white women:

        – for each age from 35 to 52, the death rate is higher in 2013 than in 1999.

        – then, ages 53 to 56 are a transition zone with little change in the death rate

        – finally, ages 57 to 64 show lower death rates in 2013 than in 1999

        I don’t know what implications should be taken from this pattern, but perhaps somebody can up with some hypotheses.

        • Thanks for these beautiful graphs.
          The most plausible hypotheses might have more to do with birth cohorts than with secular time. Perhaps summarizing the data by birth cohort would instigate more enlightening hypothesizing.

        • Given the appearance ONLY in white women of prime working years, I would guess it has something to do with delayed childbirth and a shifting of the relative risks associated with maternal mortality. This has largely been a white phenomenon until recently and would be offset in minorities by improved access to healthcare, but a similar dynamic is just beginning in Hispanic women if you look closely.

        • Given the appearance ONLY in white women of prime working years, I would guess it has something to do with delayed childbirth and a shifting of the relative risks associated with maternal mortality. This has largely been a white phenomenon until recently and would be offset in minorities by improved access to healthcare, but a similar dynamic is just beginning in Hispanic women if you look closely.

      • I covered that back then on my blog, but I was always worried how much of a statistical illusion was involved. Because overall levels of education are increasing, the remaining uneducated whites are more of a hard core of people with problems.

        But now we know that the increasing death rates are so bad that they affect white women overall in the 45-54 (and 35-44 range, and possibly younger). Much of the problem is concentrated among the less educated of course, but the size of the problem is so bad that it affects entire white female age cohorts across all classes.

  1. Your Hispanic group differs from Deaton’s because it only includes white Hispanics whereas Deaton includes all Hispanics regardless of race, which is how the data is typically presented. Fwiw, my back of the envelope found that age adjustment matters for Hispanic men.

    • Is this true?

      Or does the government call Hispanics white in death statistics as a remnant from the long-lost Postwar era when Latino pressure groups insisted on the government classifying Hispanics as white? (Hispanics disappeared from the Census categories in 1950 and 1960, to reappear in the new affirmative action era of 1970.)

      There’s a fair amount of confusion in government documents over Hispanics. For most purposes, the government is extremely diligent about breaking Hispanics out, but, notoriously, not for the purposes of calculating criminal offenders rates. The Bureau of Justice Statistics largely lumps Hispanics in as whites in calculating homicide offender rates, which has the salutary effect of making the black v. white homicide rate look a little less large. Perhaps death statistics are another lagging category?

    • Howie:

      There are different ways of categorizing. Usually we talk about white (meaning non-Hispanic white), Hispanic (meaning Hispanic white), black (including Hispanic), and others. But I agree the population can be sliced in different ways, especially given the rise of the mixed-race category.

      • The Census Bureau long separated “race” and “ethnicity,” with race being implicitly biological while ethnicity was implicitly a group identity typically passed down among biological families but not necessarily so (e.g., a person of Pacific Islander racial background who was adopted by a Hispanic family could self-identify as racially Asian and ethnically Hispanic).

        It was explicitly stated on Census forms that Hispanics/Latinos could be of any race. (Many Spanish-surnamed people are insistent upon identifying as white racially). But for unexplained reasons, “Hispanic” and “non-Hispanic” were the only ethnicities you were allowed to pick from on your decennial Census form.

        My vague impression is that the last Census form stopped trying to rationalize ethnicity and now just asked if you are Hispanic or not.

        The federal government is moving toward treating Hispanic as a de facto race along with white, black, and Asian, but isn’t all that explicit about it, since it doesn’t want white Hispanics to feel like they need to choose between identifying as white or Hispanic since most of the lobby pressures are to maximize the number of Hispanics in government reports (except of course for crime offender reports). You can be proud to be a 100% blue-blooded direct descendant of Queen Isabella and King Ferdinand and the U.S. government wants you to identify as Hispanic.

        On the other hand, deaths and births are typically collected by states and only aggregated by the federal government, and some states have lagged behind the direction the federal government is going.

  2. Men’s death rates are much higher than women’s across these years, right? That is, the baseline rate (normed to 1) is much higher, and particularly due to the causes that Deaton and Case single out. Alcohol related death rates for the whole population are around twice as high for men as women. Suicide rates are around four times higher for men than women. So it seems to me that you are removing a larger portion of the distribution of underlying susceptibility to these causes of death for men, during the years 2000-2005 during which rates are increasing dramatically (recall that these are overall death rates, but the increase is driven by the jump in suicide and drug fatality). The decrease subsequently is not exactly a mechanical result, but does not necessarily tell you much about social conditions, since the individuals most likely to die that way are no longer present in the population.

    • There was also a lot of publicity about drugs and suicide being responsible for the increased death rates. Does this re-analysis cast doubt on that explanation? Has the rise in deaths from these causes been greater in women than men? Even if not, maybe increased prevalence of these causes of death still explains the gap between America and other countries in both genders?

      • From the numbers I posted below, white non-Hispanics commit suicide at much higher rates than other groups, and no group is as likely to commit suicide as white non-Hispanic men (more likely to commit suicide during ages 45-54 than American Indian men are to die from alcohol-related deaths, for example.) The gender ratio of suicide deaths is lower for white non-Hispanics than for other groups, however: only 3:1 as opposed to 4:1 for black and Hispanic gender ratios. While Hispanic and non-Hispanic white men have comparable rates of death from alcoholism, Hispanic women have much lower rates of death from alcohol than non-Hispanic white women.

  3. One thing I remember is that the 35-54 band of women smoke relatively more than both older and younger women. See this: http://www.ncbi.nlm.nih.gov/books/NBK44311/ especially the birth cohort analyses and historical trends. Among women 55+, smoking was thought of as unladylike, with only women’s lib progressives smoking (remember Virginia Slims sponsoring the women’s tennis tour, and going further make, “torches of liberty?” There’s this sweet spot around 35-54 where smoking became as acceptable for women as men, but before smoking tobacco in general became deprecated. I suspect that that’s some of what we’re seeing, though maybe not all. I’d have to check all the dates to make sure that this isn’t off by a decade now as the population keeps aging, but it’s a hypothesis.

    Note that black and Hispanic women are currently less likely to smoke than white women, but before the 1960s black women were more likely to smoke than white women.

  4. It would be interesting to see how the rates are calculated. Since you plotted rates by year I assume the denominator of the mortality rate is some sort of population estimate. Since sub population estimates are subject to a high degree of uncertainty in general it would be interesting to calculate rates based one deaths and population numbers from the last two censuses to see if the trend is maintained.

  5. Interesting and useful analysis. There is one statement you made in passing that perhaps deserves a bit more scrutiny, however: “the absolute death rates remain much higher for men than for women, that’s just how things always are”. Would we say death rates remain higher for blacks than for whites, that’s just how things always are? Or would “that’s just how things always are” be considered a bit too accepting of a differential when the ones on the short end of the stick are a demographic group other than males?

    • The female advantage in life expectancy has fallen a couple of years or so over my lifetime. One reason is the growth of female smoking toward equality with men. Other reasons might be that masculine jobs in industry have tended to become safer, so fewer men get killed in workplace accidents.

      • I would imagine that drunk driving used to kill more men than it does now. This was particularly true for younger Hispanic men. Seatbelts, airbags, and other improvements have made car crashes more survivable, and drunk driving is increasingly cracked down upon legally.

  6. As best as I can calculate it from the CDC Wonder tool, the relative risk ratio is 3:1 for gender within race for alcohol related deaths within this age group, and about 4:1 for gender within race for suicide:

    Alcohol Related Deaths
    Race/Gender Death Rate per 100,000 over 45-54
    American Indian or Alaska Native
    Female 9.1
    Male 26.2
    Black or African American
    Female 5.2
    Male 15.9
    White Hispanic or Latino
    Female 1.6
    Male 9.8
    White non-Hispanic
    Female 3.2
    Male 10.1

    Suicide

    Suicide Rate per 100,000
    White Hispanic
    Female 3
    Male 12.3
    White non-Hispanic
    Female 10.3
    Male 31.9
    American Indian or Alaska Native
    Female 3.7
    Male 20
    Black or African American
    Female 2.4
    Male 9.9

  7. This is a great post. When I started reading, I thought: “Prof. Gelman advises not to focus on one striking result, but search for patterns of variation, but he isn’t doing it himself. Maybe a blog is not the place…” But it turned out just the opposite. My only small suggestion is to check for the cohort vs. age effect. Bump in early 2000s for 35-45 white non-H women may translate to continued increase for 45-55 white non-H women in the late 2000s. Another thing is to widen the time range. What was the pattern on few decades scale? Even if there was a steady decline in all rates we can look at how different were the trends for men and women in any given year/decade. No time to do it myself, sorry!

  8. Prof Gelman-Thank you so much for the excellent clarification on the findings of Deaton and Case.
    As an ER doctor, I was not surprised to read their findings or conclusions. Even though the effect size isn’t as large as they report, I suspect prescription opiates are an important part of this story. Your discovery of the increased mortality in women in this age range is significant…I suspect the culprit is the same:
    http://www.cdc.gov/vitalsigns/PrescriptionPainkillerOverdoses/index.html

    • Excellent point, that was why I came to this very site in fact, to see if anyone else had made the connection you have. Opiate addiction is one huge factor that has changed drastically in this very population, in this very date range. This is also where nearly all the increase in heroin use has come from recently, so it seems very relevant to me, women addicted to pain pills are also more likely to have kids and a lot more likely to feel extreme guilt about their addiction but Purdue Pharma made Billions addicting people to Oxy. They have no liability for That, but they do have a lot of lawyers, so I expect to see the Opiate angle go exactly Nowhere.

      SRRI usage has also risen steadily among that group, and while SRRI’s help some people, they also help others kill themselves, so may be relevant but there are no real numbers to refer to thanks to blanket protections for big pharma guesses, at least we have some idea about the pain pill addiction deaths.

  9. Suggestion: in addition to the first plot, you could also plot the death rate for both men and women relative to mens’ death rate in 1999, which would let you see the relative magnitude of women’s death rates. Aren’t women’s death rates lower overall as you said? How is it that their increasing rate and the men’s decreasing just happens to basically cancel out to give the leveled out overall rate? It’d be nice to see that relative magnitude difference between sexes in a plot.

  10. Another group that should be examined in greater detail is White, non-Hispanic of both sexes age 25-34. The CDC report “Health, United States 2014” shows a significant increase in death rate for 2013 vs. 1998. The Table 23 spreadsheet is available to any amateur (like me), but I will leave it to the professionals to determine if that age group became older, younger or stayed about the same over the 15 years.

    Non-Hispanic White females, deaths per 100,000, 2013 vs. 1998: 70.2; 57.9
    Non-Hispanic White male, deaths per 100,000, 2013 vs. 1998: 149.9; 124.4

    Health, United States, 2014. Table 23. Death rates for all causes, by sex, race, Hispanic origin, and age: United States, selected years 1950-2013

    ftp://ftp.cdc.gov/pub/Health_Statistics/NCHS/Publications/Health_US/hus14tables/table023.xls

  11. So isn’t this exactly one of those situations where you’d want to say something like “death rates across ages but within socioeconomic groups must be smooth” and so you borrow information across parameters (meaning hazard at each age within each socioeconomic group) to a) smooth out the age-profiles of death rates; and b) buy statistical precision?

    I just feel like this is EXACTLY the situation where the hierarchical models you pitch would fit: it is a great exercise in modeling things where we both know there must be some smoothness and we don’t have to know too much about the shape of the age-curve we are smoothing; and where the smoothing over age can give us more precision to quantify the differences in age-profiles across groups and cut down the likelihood we interpret noise as signal.

    I feel like this is how I would’ve done it (just, like, show the raw data) and then I would’ve thought “Andrew would smooth/shrink over all of this and come up with something that is actually better because it can enforce the necessary smoothness in the age profile with minimal structure on the substantive parts of the underlying model (just, like, groups are different and age profiles are smooth).

    Also you could include non-parametric cohort controls and still estimate the age-profile differences by SES group. But then I think the smoothing over age is even more important.

    • I think there’s something to be said for a Lagrangian analysis instead of an Eulerian.

      To clarify for the non-mechanicians: Instead of looking at people who are various ages in given years (year on the x axis). Look at people born in year k and their death rate in year x + k for x from say 30 to 70 (ie. at age on x axis from 30 to 70)

      Draw one curve for each year of birth (k) and think about this as a process of evolution of these curves… How has the passage of time affected people’s “risk of death at age x”

      • I think we agree that the thought of experiment of following 1 cohort over time allows you to get all the information you need to estimate one outcome-age profile, and then when you have variation (by sub-group) within cohort and variation across cohorts, you can get just about anything you want (including within-group changes over time or across-group differences).

        … but my thought was that one should smooth the age profile using information across cohorts (within sub-groups, which you get from following a particular cohort-by-subgroup over time) and that gets you a sub-group specific estimate of the outcome-age profile. Then you can (parametrically) model the shifts in that group’s age-profile over time or (non-parametrically) the differences with other sub-groups (or, parametrically, how those differences change over time).

        I think this is a natural way to think about the age-cohort-survey problem when you have repeated cross sections and useful variation within-cohort.

        • Yes, the next step would be to model the time evolution of those curves by some means. My first thought is as some kind of partial integro-differential equation, using time-lagged integral of exposure to things like cigarette smoke, as well as more immediate effects such as prevailing economic conditions (suicide), changing patterns of transportation, and transportation safety (accidents), changing medical technologies (cancer, treatable infections, etc), with a gaussian process error term to describe the noise.

          But, I’m just crazy like that.

    • Jrc:

      Yes, I’d definitely want to fit a hierarchical model. This is an interesting example because the variation is mostly not from small samples (as, for example, scaled by the variation in the Poisson distribution) but rather it represents different things happening to different demographic groups in different years. But that can be handled by appropriate error terms.

      For an example of the sort of age-period-cohort analysis you might like, see this paper with Yair. I agree that it would be interesting to fit such model to mortality data. I’m sure there are various groups of demographers and actuaries who are already doing this. For example there’s the work of Leontine Alkema.

      • The figures are beautiful.

        But to be all Gelman about it: The shading for (I presume) precision in Figures 7-10 doesn’t quite work right, does it? It just looks like a blur across the X-axis. But something just like that should work…. maybe going from darker to lighter blue/red instead of transparency? I love the idea, and in other contexts I love this particular method, but here with so many repeated crossings of the X-axis it just doesn’t seem to look just right.

        These were done in R?

      • Thanks, Andrew! Demographers are indeed active in this area, with some commonly-used APC methods subject to much debate and controversy. I agree of course that a nice Bayesian analysis, with appropriate borrowing of information and smoothing across cohort-period-age dimensions, could make a great contribution to this discussion. PhD topic anyone :)

  12. If I read it right, the first plot shows convergence in the death rates between men and women. The large difference between men and women in the United States was largely driven by differences in smoking rates, as Sam Preston has shown. So this period of convergence may be related to declining smoking rates for the cohorts of men born in the 50s.

    • Philip:

      We’re nothing close to convergence. The death rates for men and women are far apart at just about every age. For example, non-Hispanic white women age 48 in 2013: death rate is 266 per 100,000. Non-Hispanic white men same age same year: 420 per 100,000.

        • I don’t know if this is helpful at all, but…

          Let’s say there’s an underlying distribution of propensity towards suicide or overdose, and the distributions of men and women overlap, with men higher than women. If the distributions stay still, but the threshold above which people actually act upon their propensity decreases (because opioids are easily available, because people have guns in their house, because they are living alone and can drink dangerously more easily), then I believe you would see rises in rates of all these causes of mortality for both men and women (which Deaton/Case definitely do), but the ratio of male to female deaths would decrease as you move away from the tails of each distribution as well. As above, white Americans have higher suicide rates than other groups, but they also have a lower ratio of male to female as well. Not sure it’s a useful modeling assumption, but it made me think that if you continued to see rises in the overall rates of these causes of death for both men and women, you would continue to see reduction in the ratio of these causes of death, and visa-versa: if you saw a decrease in suicide rates, you would also see an increase in the male to female ratio of suicide.

    • The difference in mortality rates has been around for awhile though. This is from the paper first to use a significance test (wherein the null hypothesis of equal male/female birthrates is rejected and the research hypothesis of “Polygamy is contrary to the Law of Nature and Justice” is accepted):

      “we must observe that the external Accidents to which Males are subject (who must seek. their Food with danger) make a great havock of them, and that this loss exceeds far that of the other Sex occasioned by Diseases incident to it, as Experience convinces us. To repair that Loss, provident Nature, by the Disposal of its wise Creator, brings forth more Males than Females; and that in almost a constant proportion.”

      An Argument for Divine Providence, taken from the Constant Regularity observed in the Births of both Sexes. By Dr. John Arbuthnot, Physician in Ordinary to her Majesty, and Fellow of the College of Physicians and the Royal Society. 1710. http://www.york.ac.uk/depts/maths/histstat/arbuthnot.pdf

  13. For women, I wonder if it’s a change in the age of menopause. Women have been starting their period earlier and I suspect finishing earlier due to becoming heavier earlier. With menopause comes an increased risk in cardio-vascualar diseases.

    Having been involved in genealogy, I still think it’s a problem with reported race at death for people who aren’t known well or by uninformed family e.g. children. Also that white is probably the default race unless indicated otherwise by appearance for people without informants.

    • I was wondering about the effects of hormone therapy for women during menopause and post-menopause. There has been much controversy about that in recent years. Fortunately, the single year graphs can shed some light on that: It appears that overall death rates for women are down from age 53 up.

  14. Despair – Columbia University style. Thanks so much for breaking out all the data. It certainly helps put some constraints on the story-telling. I’m struck by the substantial decreases in mortality among some groups – for instance African-Americans. So part of the dramatic relative difference in the curve shapes may also be that there was less room for improvement in some of the larger mortality culprits (heart disease, cancer etc.) across groups? The group contrasts are likely a mix of differences in benefits due to relative health and healthcare improvements, as well as differences in relative lifestyle risks (a la despair american style).

    • A quarter of a century or so ago during the Crack and AIDS Era, African Americans were dying in large numbers due to homicide, overdoses, and AIDS. I presume all of those causes of death have gotten less prevalent than in 1991.

  15. “of course the absolute death rates remain much higher for men than for women, that’s just how things always are”

    I’m not sure this is a true statement… If you are fitting your statement to persons born after 1880, I’d say it is correct. If by saying “things always are [that way]” instead means a longer time period than ~140 years, then I don’t think your statement holds.

    I thought I had the underlying document but I can only seem to find the abstract. The paper I am thinking of is called “Twentieth century surge of excess adult male mortality” which was published in the middle of this year via USC. I don’t think the paper has actually has CDFs in it, but to the extent they did, the age you’d be interested in here is ~ 50 years old, and whether there were asymmetries in mortality rates between the sexes at say ~ 50 years old. I believe the assymetry did not exist in developed countries for individuals born prior to 1880. Then again it’s a fairly recent paper and perhaps there is an underlying bit of statistical bias in their calculations ala Deaton / Case that no one has caught yet?

    Food for thought.

  16. Great graphs.

    By the way, the single year of age graphs could be replotted with year of birth on the horizontal axis. The shape of the curve wouldn’t change, just the axis labels.

    I bring this up because I hypothesized last week from looking at earlier graphs that white people who were 18 years old during the Drug Years of the late 1960s into the early 1980s were more susceptible to dying in middle age due to overdoses, liver failure, or suicide:

    http://www.unz.com/isteve/is-there-a-generational-explanation-for-rising-white-death-rates/

    I may just be generalizing from a grade school classmate of mine who turned 18 in 1976. Even though he was born a little late for the Grateful Dead, he became a Deadhead, and eventually killed himself at about the same age that Jerry Garcia died of an overdose. But i don’t think it’s implausible that famous historical trends among Baby Boomers might have long term impacts.

  17. Why in God’s name is it the relative death rates of men vs women that is of concern here, not the rise in the absolute death rates?

    Men die from suicide at roughly 4 times the rate of women, and from alcohol poisoning at roughly 3 times the rate. If, say, men rise by a factor of 1.5, and women by a factor of 2 in suicide, then, of course, men rise by 50% and women by 100% — which seems like women are taking a much bigger hit. But, of course, the rise in the absolute number of men dying by suicide is double that of women. Really, who cares in the end about the relative rates? Why shouldn’t we care about the number of additional deaths?

    • Just to amplify on my point, if we regard the additional deaths as deaths which should be preventable, given better circumstances, isn’t the obvious moral imperative to focus on the increase in absolute number of deaths, rather than the relative rates? Is it OK to treat men as somehow more expendable merely because they started out with much higher rates?

    • Both are important and convey different information to researchers. Understanding that there are distinct changes for women will help researchers look for the underlying causes. This does not mean that researchers should stop focusing on overall death rates and their causes, but it does mean there might be something worth spending a bit more time and money on that might be unique to the lives and experience of women.

  18. Prof. Gelman: I actually worked in a SocSec disability law firm, and the worst stories I heard were from middle-aged white women addicted to painkillers – and for good reason. Their lives were miserable. They were truly “worn out workers.” Just saying.

  19. The authors’ analyses show severe education differentiation of mortality, your reanalysis of original data do not

    Would you consider an educational level disaggregation?

    The authors in their 1999 study discuss at length education as proxy for economic strata, so maybe we are talking about poor white non-HS women, which is a narrow demographic, even if they have this concentrated pathology, the impact and generalizability, is narrow

    Might I add, from Table 1 – Black all cause mortality is 581.9 (/100,000; and White is 415.5/ 100,000, so yeah in this demographic Blacks die younger (higher mortality in a younger cohort)

    BUT Hispanic mortality is 269.6/100,000 meaning that Hispanics are less likely to die than Whites

    This is counter intuitive

    Also if we look at mortality among Whites, contrasting any college, with none, the ratios are very big

    147.7 (none), vs 59.9 (any), and 36.8 (BA or more)

    Meaning that college, any, enormously immunizes these people, and presumably ONLY women, against these pathologies

    It is unlikely that going to college immunizes against early death, although it might. More likely people (women) who DO go to college are different, for many reasons, from those women who don’t

    Thoughts?

  20. Great example!

    Prior: I have absolutely no experience in this area.

    Currently, there is some moderate press activity about birth control pill health effects in Germany. That might help explaining some parts of this diverging M/F trends. If effect size comes close and pill-popularity development in the US fits the timeline here…

  21. Good thing you are also looking at the younger ages. I did the same for the age group 25-34 a while ago for whites, and found a 20%(!!) relative increase in mortality rates: economicssendil.blogspot.com/2015/11/deaton-and-case-are-wrong-mortality.html

    The numbers seen rather out of place, but since it’s the same source used by Case and Deaton I do not think it should be less believable than their numbers on the age group 45-54.

  22. Andrew, did I understand correctly that your student deserves credit for the work on this? You wrote:

    “Yair did the beautiful figures (as well as the beautiful modeling). My main contribution was general-purpose nagging.”

    If so, blogs like this one are misleadingly attributing the work exclusively to you:

    http://junkcharts.typepad.com/numbersruleyourworld/

    If I were Yair, I would be pretty unhappy to get zero credit in this blog post on junkcharts.

  23. You’ve left out a crucial variable. From Deaton’s paper:

    “The three numbered rows of Table 1 show that the turnaround in mortality for white non-Hispanics was driven primarily by increasing death rates for those with a high school degree or less. All-cause mortality for this group increased by 134 per 100,000 between 1999 and 2013. Those with college education less than a BA saw little change in all-cause mortality over this period.”

    What is going on among the non-HS degree group specifically?

  24. Is it possible that as women statistically come to live more like men – smoking, drinking, job stress, etc. – that their mortality rates will more closely resemble those of men? Maybe they are just catching up. The affluent and educated have resources to handle it, but others do not.

    • Zbicyclist:

      Yah, I noticed that too. I was going to mention it, but I wasn’t quite sure, because there’s also a convention in econ that authorship is alphabetical almost no matter what. So it’s possible that Case and Deaton made it clear that Deaton was the primary author, in which case it could make sense to think of it as his paper first. I have no idea so I want with Case and Deaton which is how the authors are listed in the paper.

    • Almost nobody was paying attention to increasing death rates among white people — after all, they have White Privilege — until a brand new (quasi) Nobel Laureate brought it up.

      Without Deaton’s Nobel Prize charisma, this would probably have not gotten anywhere near as much attention as it did, so emphasizing the Nobelist over his wife makes sense.

      • Steve:

        I dunno about that. Papers in PNAS can get a lot of attention even with no Nobel prize floating around. I’m guessing that even had Deaton not received the award, this paper still would’ve got lots of press. Not as much, sure, but still a lot. I think it was the prestige of PNAS more than the Nobel that pushed this into the limelight.

  25. Hi Andrew,
    Nice paper. Odd no one commented on the role of relatively small numbers of events in these subgroups. It is possible that what is happening is an artifact compounded by the use of the ACS for denominators.
    Dan

  26. When I first saw this study, I was very skeptical. But there are some curious numbers in there.

    First, among Whites 45-54(*), the increase in deaths from 1999-2013 due to *accidents* was greater than the total of the increase in deaths caused by alcohol, suicide, and lung disease. I have no idea what to make of that. Could it be a methodological issue? For other ethnic groups, the trend is flat. Did hundreds of thousands of middle-aged White people take up dangerous sports, or get jobs in hazardous industries? Per Steve Sailer’s comment earlier (http://statmodeling.stat.columbia.edu/2015/11/10/death-rates-have-been-increasing-for-middle-aged-white-women-decreasing-for-men/#comment-251541), one would expect this to be in constant fall simply because of the improvement in workplace safety and the changing nature of work.

    I made my numbers available here: http://meta-systems.eu/dev/stuff/US%20deaths%20age%2045-54%201999-2013.xls — feel free to download and play with them. There are some interesting things on the per-state tab in that file, such as the enormous increase in the correlation among Whites (none for Hispanics and not much for Blacks) between per-state income and per-state death rates from 1999-2013, cf. cells C59 and E60. (Of course, ecological fallacies lurk everywhere on that tab, so caution is required.)

    (*) Not age-adjusted, unfortunately. I couldn’t work out how to get CDC Wonder to give me age-adjusted rates for just one ten-year group, and in any case, the age adjustment doesn’t make much of a difference to the trend.

    • Nick:

      I did the age adjustment by getting data by single year of age and then doing the averaging myself. But there is some data that are only available by 10-year age category so then things get trickier.

    • Roger:

      Yeah, the web is full of these content-stealing sites. I guess their goal is to get clicks on Google searches.

      What really disturbs me, though, is this in today’s New York Times: a whole series of letters in which “readers offer their theories to explain why middle-aged whites are dying earlier,” with lines such as “Until we recognize that our distinctly American privileging of the self over the community is killing those of us who are not part of a loving and tight-knit community, the mortality rates will continue to increase.”

      This is particularly frustrating given that I’ve been in contact with two different New York Times reporters about these mortality trends. I guess there’s a division of labor between the letters-to-the-editors page and the science reporters.

      • P.S. On the plus side, the same section of the newspaper has an interview with Deaton where he says, “There’s been a tremendous amount of recalculating of our numbers in the media and blogs.” It’s good to be recognized, even if only in this indirect way.

      • I would guess that “division of labor” is probably underestimating the lack of contact between the letters-to-the-editors editors and the science reporters.

Leave a Reply to thebriang Cancel reply

Your email address will not be published. Required fields are marked *