Out Of Control

Controlling for gender/race, there is no gender/race wage gap

live maia reaction to controlling for a collider

The Economist @TheEconomist

Control for factors like credit scores and troubling racial gaps almost disappear

econ.stRacial discrimination in mortgage lending has declined sharply in AmericaControl for factors like credit scores and troubling racial gaps almost disappear

Recently, The Economist published an article about how, if you control for credit scores and other factors, such as debt-to-income ratios, racial gaps in credit allocation disappear. Now, the point was that explicit discrimination disappeared, and the kind of research done by the Fed on such topics is relevant - after all, whether or not mortgage lenders reject Black people at double the rate they do whites for being Black or for being poorer is an important question.

What’s the issue here? The main issue is that it’s econometrically unsound to derive certain types of conclusions from that kind of regression. In econometrics, this is known as collider bias - if the link between two variables is mediated by a third variable they’re both correlated to, your regression is gonna be all screwed up. For example, if wages were highly correlated with height, then you could say that there is no gender wage gap when controlling for height, but it only disappears because men are taller than women - this doesn’t tell you anything about actual gender dynamics.

So the interpretation of the findings that claimed there was no discrimination, so Black people simply had lower credit scores for wholly colorblind reasons, were kind of baffling. I think everyone recognizes that, but a lot of people just fail to acknowledge the exact same logic for the gender wage gap in particular.

So I’m gonna try to look into the issue of systematic disparities by focusing on income, and specifically on racial and gender pay gaps. Seems like a breeze!

Fair or weak

The wage gap

So is there a wage gap between men and women? Yes. In the United States, women made 18% less than men on average, which is less than the 41% gap that existed in the early 70s. This gap is larger for both non-white women, except Asian women. At the same time, women make roughly as much as men for most of their early careers, but then earnings diverge around the early 30s and never converge again. Lastly, the largest income gap is for the richest women - women in the 95th percentile (i.e. who make more than 95% of people) earn a quarter less than men.

What gives? Women are more educated than men, so college attendance isn’t really an explanation. However, career choice might be - nursing is overwhelmingly female and low pay, while computer programming is overwhelmingly male and highly paid. In fact, there is a negative correlation between share of women and earnings in different industries. Some of it is caused by women being less interested in those fields than men (though how much is related to social factors and education is hard to say - there isn’t a performance gap in math between men and women until halfway through middle school). But the main issue is that only a third of the earnings gap is due to career choice - i.e. women do make less than men because they choose lower paying fields more frequently, but also make less than men in their own field, and this effect is twice the size of the first one.

The "The Two Income Trap" Trap

Where does this within occupations gap come from? Primarily, from hours worked and experience. Case solved, right? Not quite. Why would women systematically work fewer hours and for fewer years than men? If you’re extremely lazy you just say “preferences”, but of course there’s basically no way of verifying people’s preferences. In reality (and I’ve written about this recently!) it’s mostly because the burden of housework and childcare falls overwhelmingly and disproportionately on women - men do a third to a tenth of work around the house, and it’s primarily things that involve being out of the house (be it grocery shopping or just being outdoors).

Thus, the gender pay gap is just a motherhood pay gap - women who don’t have children make the same as men, ceteris paribus, and even in very progressive countries such as Denmark. This is due to both a drop in participation at the extensive margin (i.e. women stop working altogether - because you can raise a family on a single income), and at the intensive margin; they also work fewer hours to take care of the kids; this can be seen in Chilean data for women in the informal sector, who have more control over the hours they work. Another example comes from Mexico, where grandmothers provide a large chunk of childcare services: women’s labor force participation (at the intensive and extensive margin) suffers when their mothers die provided there are not other childcare alternatives. Additionally, women have different career tracks after childbirth, primarily because they shoulder all the (unpaid) domestic labor that men won’t, with clear implications for future earnings growth. Given the importance of parenting on society, childcare can be understood as a public good that is beeing free-ridden by fathers.

Race to the top

When talking about race, the Economist article attributed the difference mainly to credit scores, but the issue is that Black people have systematically lower credit scores than white people, and even if it’s not due to outright discrimination, it’s pretty evident that a lot of race-related stuff pops up: for example, Black people have lower wealth than whites mainly due to lower rates of homeownership, something that can be traced to zoning regulations being explicitly racist and US housing policy being systematically deployed to preserve or create racial segregation for most of the 20th century.

The thing about the racial wage gap specifically is that it’s very poorly understood by experts. It hasn’t shown a clear trend in the past (increased and then decreased and then increased again), is different by gender (used to be very small for Black women and then widened to match the overall gender wage gap), and almost entirely persists even after controlling for hours worked, experience, tenure, etc. There is, of course, a significant component for both industry and education, but it only accounts for half the gap between the two groups.

Occupation is the dominant factor for Black women, while educational attainment is for Black men - although educational quality itself might not be a big factor, and simply less education seems to dominate. Another major factor is occupation itself: Black men specifically have a much lower rate of labor force participation than whites, and a much higher unemployment rate, mainly driven by a higher exit rate. A proposed explanation for this are the disparities in incarceration rates between the races; and in fact, the decline in the racial earnings gap during the 80s and 90s might have been entirely an artifact of selection bias due to of mass incarceration.

The main factor that can explain the Black-White income gap is, ironically enough, unexplained factors, which have been rising in importance over the past 50 years. This could account, for instance, for why the racial wage gap is similar in magnitude for all levels of education for both genders. Exactly how to disentangle the unobservable component of the racial wealth gap is complicated, but some of it (about a third) appears to be accounted for by overall income inequality (given how Black people are overrepresented amongst the lowest levels of income), and the rest is an unknown unknown - it is likely that discrimination plays a role.

Birds of a feather stick together

What role does discrimination play in economic inequality? A complicated question indeed. There is some (empirical, experimental) evidence of discrimination: for example, a study found that blind auditions for orchestra positions increased the chances that women get hired, while another found that, for similar positions and identical resumes, those with stereotypically “white” or “neutral” names got more call-backs than those with “Black” names. Evidence from European countries also showed that identical resumes for women with different names (“Muslim” and “local”) and different pictures (none, without a veil, and with a veil); and while “Muslim” names overall were disfavored for subsequent interviews, the discrimination was worse for Muslim women who were wearing veils in their pictures.

The most traditional model of discrimination in economics comes from Nobel Laureate Gary Becker. In this model, people have heterogenous preferences for racial segregation, from “none” to “extreme racism”. When it comes to hiring, assuming ceteris paribus between candidates, the decision between a white and non-white applicant comes down to the racial preferences of the employer. However, minority applicants can get an edge by asking for lower pay, which gives rise to a wage gap unexplained by any observables. Given that the employment pool of racial minorities is determined by the total amount of discrimination, then how racist the average employer is, not how racist the most racist employers are, is what determines the overall racial wage gap.

The main problem with this model, to paraphrase Nobel Laureate Kenneth Arrow, is that it simply assumes that discrimination disappears on its own: if firms aim to maximize profits, there are gains to be made by employing minorities, so that the most racist employers would make smaller profits and get outcompeted by “liberal” firms over time. This might explain why big companies seem to have gotten anecdotally “woke” lately, but does not explain why the racial wage gap has persisted over time. However, under imperfectly competitive competition (for instance, due to different frictions in job search, or more imperfect information) then some racial discrimination is allowed to persist; and moreover, the profit penalty to discrimination might not be sufficiently large to eliminate racism in hiring decisions. A study evaluating Becker’s claims found them accurate, but also couldn’t explain three quarters of the impact of racial prejudice on earnings.

Another potential source of disparities is what’s called “statistical discrimination”: if securing information about a potential employee is difficult, then employers might resort to stereotypes about their gender, ethnicity, religion, or nationality. This should respond to differences in skills observed, but also in differences expected given overall cultural norms and expectations - which would not be easy to undo regardless of skills, education, or mass incarceration being undone.

Conclusion

Regarding gender, there’s nothing here I haven’t said in previous posts about the two income trap, the gender wage gap, women’s workforce history in the US, and women’s participation across the developing world. Real ones read those.

Regarding race, discrimination is bad m’kay.

And regarding the econometrics of controlling for racially-sensitive colliders:

maia @maiamindel

Controlling for height, Taylor Swift is as tall as Bruno Mars

Edit notice (Jan 5 2023): added the sentence about grandmothers in Mexico. (Poor tab management led to me leaving it out accidentally).