Death rates for middle-aged white (wo)men in the US

From Andrew Gelman (this year's guest speaker at 3rd IAST-TSE conference in political economy/political science):

In a much-discussed recent paper, economists Anne Case and Angus Deaton reported “a marked increase in the all-cause mortality of middle-aged white non-Hispanic men and women in the United States between 1999 and 2013. This change reversed decades of progress in mortality and was unique to the United States; no other rich country saw a similar turnaround.”


Gelman then shows that:

(i) part of this effect is due to a composition effect: the 45 to 54 year-old in the US have become on average 5 months older between 1999 and 2013, which explains part of the increase in mortality among this group in between the two dates;

(ii) that the mortality within this group has increased between 1998 and 2005, but has remained roughly constant since;

(iii) most importantly, that it is mainly the mortality of women that has increased, and not that of men!

His conclusions:

First, post-publication review is a wonderful thing. A blog commenter alerted me to the possibility of age-aggregation bias, Angus Deaton pointed me to the relevant CDC website, and I was able to dive into the data, perform some calculations, and make some graphs. The classical peer-review system is painfully inefficient: Once an article appears in a journal, I could submit a letter of correction, that letter would have to go through a review process and would be severely limited in length, then the original authors could reply, and so on. All at the speed of the U.S. mail circa 1775. Real-time feedback gets us there much faster.

Second, when studying a time series, graph the whole thing, don’t just compare the beginning to the end. A simple comparison of 1999 to 2013 shows an increase in death rates at most ages. But the time series shows an increase since 2005 and then stasis—a much different picture.

Third, break up the data. That post-2005 stasis turned out to mask an increase for women and a simultaneous decrease in death rates for men.

Fourth, spot a potential bias, then estimate its size. The “pig in a python” image of the baby boom moving through the age distribution suggested that raw death rates in 10-year age bins might be biased. But by how much? To see, I first made an order-of-magnitude calculation and then followed Deaton’s suggestion and went to the raw data.

Finally, science progresses by continual revision. Case and Deaton made a mistake by not adjusting their numbers for changes in age distribution—but had their paper never been published, we never would’ve been having this discussion. Meanwhile, their main finding holds up and is clearly worth further exploration, and researchers can also look into the diverging patterns since 2005 for men and women. While this is happening, I’m pretty sure some people will find major problems in my analysis. That’s how it goes, two steps forward and, if we’re lucky, only one step back.