The subject of this morning's RI is this paper as cited in this article. The thesis of the article is in the title: "As Women Take Over a Male-Dominated Field, the Pay Drops".

The quote we'll investigate is this one from sociologist Paula England:

Once women start doing a job, “It just doesn’t look like it’s as
important to the bottom line or requires as much skill,”

Which is a conclusion supposed to be drawn from England (2012) . This is what's called the "devaluation theory" in the wage gap literature, the view that male centric work is seen as "more valuable", basically the pure discrimination view.

Now we'll leave the obvious criticism of the quote aside for today. This criticism is basically supply and demand -- when women start doing a traditionally man dominated job, aggregate supply of worker increases and hence market price lowers. Then there is the issue of drawing causal inference in these sorts of statistical exercises which was talked about here some time ago and Levine(2003) addresses as such:

Whether women limit their fields of work in order to fulfill family responsibilities or discrimination confines
women to certain jobs, the market’s wage response is the same: the abundant supply of women to certain jobs 
relative to employer demand will prompt firms to use more women than otherwise and, in hiring each 
new worker, spread complementary inputs (e.g.,  capital)  more  thinly;  this,  in  turn,  decreases  the  
productivity  of  workers  in female-dominated jobs, and consequently, their earnings.

Basically, you can't imply any sort of causality when observing an effect in this context.

But let's leave all of these completely valid criticisms aside. Instead, let us focus on Ms. England's body of work supporting the quote. To do this, we need to go back to 2007, 5 years before the 2012 article supporting the quote. In this article titled "Does Bad Pay Cause Occupations to Feminize", England first does a literature review on wage gap, citing 5 articles, all using longitudinal data, then uses a panel data model to differentiate itself from previous methodology.

The idea is to test whether the devaluation theory is plausible, or if lower paying jobs attracted women (which remember from above, we can't know using these methods). She briefly contemplates the competing theory (that wage differential is partly explainable by non-wage compensation, like time flexibility), but notes it is espoused by "neoclassical economists".

Now a few things are to be noted here. First, she uses the words "causal effects" in situations where this is distressingly not applicable. More importantly, all of the longitudinal studies all state findings of evidence for the devaluation theory from studying the evolution of wage and the man/woman proportion in the industry overtime.

Also, let's call the "longitudinal methods" they use what they are: distributed lag models. As any undergrad will tell you, you should test for stationarity whenever you have an AR(p) component in a model, yet, none of this literature ever does this. Not once is are the words "station", "unit root", or "difference" mentionned in any of these papers.

Now England (2007), using panel data does not find any relevant effects supporting devaluation theory, and notes that the previous studies were in all likelihood suffering from omitted variable bias. Note that England (2007)'s panel model also uses an AR(p) component not testing for non-stationarity, so I'd take all of this lightly, but maybe I'm asking too much. She notes:

One interpretation is simply that the reported association between occupations’ sex composition and wages has always been spurious rather than causal, due to some unidentified omitted variable. [...] This is not our interpretation, however.

She then proceeds to tell a 3 page story (founded in no empirics whatsoever) explaining away her results. And, in 2012, 5 years later, she publishes the paper linked above, where she uses the same bad time series methodology as the papers she critiqued herself in her literature review, and finds the same results as the stated literature. This is what's quoted in the NYT article published this week.

And this, my friends, is "Deeply Bayesian" methodology in it's full glory.