Working men are p***ed

Unemployment at 4.9%. More than 75 straight months of continuous job growth. Per capita GDP at an all-time high. Summer’s here. But nobody’s dancing in the streets.

Here are a couple graphs that may help explain why:


First, note that even as real per capita GDP continues to reach new peaks, the typical American adult (i.e., the person at the 50th percentile) in 2014 was no better off than in the last two recessions. Median personal income was down about 5% from about a decade ago. The rising tide is clearly not lifting all boats. The more politicians crow about the improving economy and the more economists say we are at “full employment,” the greater the disconnect becomes.

Caveat: Ideally I’d have median personal income data from after 2014. The numbers have likely improved since then. Jared Bernstein notes that real weekly earnings have grown almost 4% since 2014. Whether personal income has had similar growth also depends on employment trends. Which brings me to:


The above is the US “prime-age” (25-54) employment/population ratio from 1990 to present. The employment/population ratio is about 2 points lower now than then, partly because the standard unemployment rate is about 1 point higher and also because more jobless people are “out of the labor force,” i.e., not actively seeking work. Would they take work if offered it? The Bureau of Labor Statistics report for June suggests many of them would — the alternative unemployment rate including jobless “discouraged workers” and “persons marginally attached to the labor force” is 6.0%. (Add in part-time workers who’d prefer to be full-time and the rate rises to 9.6%. It’s been improving for six years but is still no better than in mid 2008 when we were in a recession.)

What’s really striking to me, and what inspired the title, is the gender breakdown of those prime-age employment/population numbers. (Sorry, the separate BLS graphs for men and women are too messy to use here, but you can Google them.) For women, the employment/population ratio has regained its 1990 level of 71%; for men, the ratio is down about 5 points, to 85%. There’s a story here, probably several. The erosion of male privilege has been a big theme of this year’s political commentary. Without getting into the politics or ethics of that, let’s just note:

  • These graphs indicate that it by no means a new thing, at least in the labor market. The male employment/population ratio has been falling since 1969, when it was 95%. The female employment ratio has been rising steadily since at least the late 1940s, when it was less than half the current level.
  • During the boom decade of the 1990s, the male employment/population ratio fell by about half a point, while the female ratio rose sharply, from 71% to 74.5%.
  • Male employment was hit harder than female employment during the 2008-2009 recession (men’s employment fell from 89% to 81%, women’s from 74.5% to 69%).
  • Men’s employment has actually risen faster than women’s during the post-2009 recovery (ratio up by about 4 points vs. 2 points), but again, the male employment ratio is down about 5 points from 1990, whereas the female employment ratio is unchanged.


Tags: , ,

5 Responses to “Working men are p***ed”

  1. Richard W DiSalvo Says:

    So, you’re suggesting that support for Trump can be partly explained by a “widening gap” between male and female labor market outcomes, so, a regression like:

    Percent Trump (in republican primary) ~ Change in [LaborM – LaborF] in last k years

    should have a positive coefficient (maybe with a state effect)

    so if there is a widening gap between the labor market outcomes of males relative to females over the last k years that’ll explain why people support Trump in some places rather than others, maybe.

    Maybe the republican primary is not a good place to look for this evidence, what do you think?

    Do you propose the same argument for Bernie?

    Also what is a good k to pick? Your reference to 1990 suggests k = 25.

  2. Richard W DiSalvo Says:

    I’d never pass up the opportunity to use the CPS sooo let’s see what happens when you run that county-level reg.

    I obtained primary election data by county from Ben Hamner on Kaggle ( and got CPS data from IPUMs so I could calculate employment by gender and county.

    I worked on the period 2000 to 2015, county identifiers weren’t available in 1995 on my first data pull from ipums so I decided to play it safe. Then I defined two periods: period 1 is from 2000 thru 2003, period 2 is from 2012 thru 2015. So each are 4 years. I then tried to calculate employment rates (employed people/people surveyed) for each county-period.

    There are only 164 counties in common in both periods, on average 600 people (300 M 300 F) in each.

    Focusing just on period 2 where I have more power, 326 counties on average with 200 M 200 F in each. I regress proportion in republican primary voting for trump on (a) state fixed-effects, (b) proportion of women working and (c) proportion of men working. (I weight by RHS variable sample sizes because that’s where most of the measurement error is). Results:

    . reg fracTrump i.statefip isWorkingM isWorkingF [aw = N] if period==2, r
    (sum of wgt is 6.4017e+04)

    Linear regression Number of obs = 326
    F(31, 288) = .
    Prob > F = .
    R-squared = 0.8965
    Root MSE = .06607

    | Robust
    fracTrump | Coef. Std. Err. t P>|t| [95% Conf. Interval]
    statefip |

    … state fixed-effects …

    isWorkingM | -.2021733 .0742408 -2.72 0.007
    isWorkingF | -.3135516 .0885615 -3.54 0.000
    _cons | .7832353 .0759342 10.31 0.000

    WordPress doesn’t like regression output so if I dropped the confints.

    There isn’t immediate support cross-sectionally for a “Trump does well where men do worse than women” theory. For “effect sizes” we might be interested in the (weighted) standard deviation of these RHS variables across counties… well, it’s higher for females:

    . su isWorkingM [aw=N]

    Variable | Obs Weight Mean Std. Dev. Min Max
    isWorkingM | 585 139894 .8266353 .0588672

    . su isWorkingF [aw=N]

    Variable | Obs Weight Mean Std. Dev. Min Max
    isWorkingF | 585 139894 .6831753 .0631355

    But even if you standardize by considering the impact of 1 sd change,

    male “effect” = -0.20*0.059 = -0.0118
    female “effect” = -0.31*(0.063) = -0.0195

    the female point estimate is still larger.

    As you can probably tell from the standard errors, p = 0.38 on a test of equality, so, the “effects” by gender are pretty much the same.

    What does this mean? The counties where men are doing worse than women in the labor force are no more or less trump-supportive than the counties where women are doing worse. This holds if you measure “worse” in terms of weighted county gender ranks (that’s what that sd stuff was all about). There really seems to be just one constant effect across genders, where counties where people are working support trump less, doesn’t matter if those people are men or women.

    Anyway. About that change regression.

    Frankly it just doesn’t work at all, but I think it’s more because of lack of power than anything else. Point estimates look in your favor, where places that have had lower changes in male working prospects (e.g. deterioration) since the early 2000s are more likely to support trump, and female labor force changes seem unrelated.

    | Robust
    fracTrump | Coef. Std. Err. t P>|t| [95% Conf. Interval]
    chngIsWorkingMale | -.1026162 .1547598 -0.66
    chngIsWorkingFemale | -.0086968 .1147873 -0.08


    (Btw the changes over the two periods have similar sds across genders:

    . su chngIsWorkingMale [aw=Ntimeavg]

    Variable | Obs Weight Mean Std. Dev. Min Max
    chngIsW~Male | 164 49085.25 -.0419997 .0469276 -.2865979 .1268762

    . su chngIsWorkingFemale [aw=Ntimeavg]

    Variable | Obs Weight Mean Std. Dev. Min Max
    chngIsW~male | 164 49085.25 -.0299925 .0566993 -.25 .4924243


    So we might speculate that trump support has to do with angry men, but the county-level variation in trump support regressed on county-level variation in employment of men and women, just doesn’t support this theory well.

  3. Ranjit Says:

    Nice empirics there, Richard. I’ve actually been skeptical that Trump’s support has all that much to do with economic issues, ever since I saw a national exit poll result that the median income among his voters in the primaries was about $70K, higher than for Hillary’s or Bernie’s voters.

    You’ll note that I said “has been a big theme of this year’s political commentary” rather than agreeing with it.

  4. Democommie Says:

    Trump is a racist, misogynistic, p.o.s. serial bankrupt, snake-oil selling huckster. I do not require a explanation for why his like minded fanboy supporters have flocked to his NeoNuremburgian brown shirt fantasy

  5. Richard W DiSalvo Says:

    If you are interested in some better empirical work on Trump’s support, I have been informed of a rush job on the topic by Jonathan Rothwell at Gallup:

    Religious, self-employed, unemployed, veteran, married/was married, male, white, high-school or technical school-educated. These positively predict favorable views of Trump. (See probit reg table 3 page 18.)

    But, this analysis doesn’t provide evidence for or against what I think is Ranjit’s hypothesis, namely, that men who find themselves earning less relative to women are more likely to support Trump.

Leave a Reply to Richard W DiSalvo Cancel reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: