Discover more from Samstack
Notes on the Gender Wage Gap
About a month ago, I wrote that ‘most of the reporting on the relationship between happiness and income is pretty bad’. Compared to what has been written about the Gender Wage Gap, the reporting about happiness/income is Pulitzer-worthy. Discussions about the Gender Wage Gap (GWG) are really bad. The first type of bad reporting on the GWG are the claims made about the amount of money that women make compared to men - UN Women claims, for example, that ‘Worldwide, women make only 77 cents for every dollar earned by men’. This is trivially true, but isn’t very interesting, and doesn’t contextualise the figure. Readers of this blog probably already know why this isn’t a useful figure, but to elaborate slightly - we don’t know why women earn less than men, and we certainly can’t attribute the entire gap to discrimination. It could be that women have a preference for working fewer hours than men do, it could be that women prefer the types of jobs that don’t scale and are likely to earn less money at the highest levels, and so on. It could also be largely to do with discrimination, but the figure itself doesn’t get us very close to figuring out why the gap exists.
But then you get the bad arguments by the people trying to debunk the GWG. The main bad argument made by those arguing that the GWG is not to do with discrimination at all is one that looks roughly like this:
To get the true answer as to whether there is a Gender Wage Gap, a good social scientist can’t just compare the wages of women to the wages of men, they need to control for occupation, control for education, and so on. And when we we control for those things, the GWG becomes very small indeed.
The problem here though, is that controlling for occupation is a bad control, because occupation comes after gender in the causal chain. To illustrate: suppose that we’re interested in the question of whether going to Oxford for your undergraduate degree has a causal effect on your salary compared to going to the University of Manchester. We look at the crude difference in wages between Oxford grads and UoM grads, and see that there is a big difference. But then imagine that someone objects that we haven’t controlled for occupation. And then we find that after controlling for occupation, the salary difference becomes much smaller. The problem is that there is a causal connection between going to Oxford and getting a good job - it may be the case that investment bankers who went to Oxford don’t earn much more than investment bankers who went to UoM, but people who went to Oxford are more likely to get investment banking jobs, so we’re controlling away the causal mechanism.
A similar thing is likely to be happening when we introduce controls into our regression for the GWG - if, for instance, women have lower wages than men partially because high-paying jobs are less likely to hire them, controlling for occupation is a bad control, for the same reasons as our Oxford/UoM example. So, what can we do? Well, this seems like a good example where regression + controls simply isn’t a good way to figure out how large the wage gap is, and how much of that gap is due to discrimination. It almost certainly isn’t the case that the entire wage gap is due to discrimination, so we can’t just use the crude figure provided by UN Women. But we also can’t just control for everything we can think of that has an effect on wages, because if there is discrimination, there’s a good chance we’re controlling for the causal mechanism. So we need to use experiments or some quasi-experimental technique here instead.
There is a pretty big literature on the GWG, so I’m not going to try and do a comprehensive literature review here. But the studies we want to look at generally are Audit Studies that send identical CVs (resumes) out for similar jobs, changing only whether the name is a female-sounding name or a male-sounding name. Take, for instance, Neumark (1996) - he sent out identical CVs (except for the names) to various high-pay restaurants to isolate the effect of discrimination. The finding was that:
In high-price restaurants, job applications from women had an estimated probability of receiving a job offer that was lower by about 0.4, and an estimated probability of receiving an interview that was lower by about 0.35.
It should be noted though, that it isn’t clear where the discrimination comes in. It could be that customers who go to high-end restaurants prefer male waitstaff, sommeliers, and so on. Or it could be that employers think that women are likely to be less competent or quit earlier, and so they prefer hiring men to women. But we should also note that this is just one example of hiring practices in high-end restaurants in the mid 90s, so we can hardly extrapolate to all industries in all countries at all times! This is just being used as an example of a good study that isolates the effects of discrimination.
Neumark also did a pretty comprehensive literature review called Experimental Research on Labor Market Discrimination on all sorts of discrimination among job-seekers, which I recommend perusing. It’s difficult to summarise the findings on gender specifically in a blog post, but basically the conclusions are a bit all over the place - some studies find that young women are discriminated against (possibly because employers think they’re going to take time off when they have children) but older women face less discrimination, other studies find that this effect holds only for women with children rather than young women specifically (e.g. the callback rate for childless women seems to be 1.8 times that of the callback rate for mothers), others find that there is basically no discrimination against women in the labor market (call-back rates appear to be basically the same), and on it goes!
The main thing I think that is worth taking away from the literature on the GWG is that figuring on the effect of gender discrimination on wages is hard! You can’t just compare the amount of money that women earn to the amount of money that men earn, and you can’t just use regression with a ton of controls that are correlated with income. Experiments work a lot better to isolate the effects of discrimination, but their external validity might be pretty low (finding there is discrimination against young mothers with STEM degrees in Mexico doesn’t tell you much about discrimination against childless women in the UK in their 40s who want to work in PR). So, basically: *shrugs*.