4 ms·

I agree with this broadly with a caveat I'm curious about: What do we do when these factors also have factors? Does it increase the contribution of a factor if

I agree with this broadly with a caveat I'm curious about: What do we do when these factors also have factors? Does it increase the contribution of a factor if it applies to both the measured factor's dependencies and also the dependencies' dependencies? Let's say pay gap is 5% "anti-black racism" when it comes to a black man being paid in software development. Pay gap is also 40% job performance, 20% visibility, 15% negotiations, 10% unknown, 5% underpaying due to being black. What happens if visibility is 10% antiblackness, due to a manager have just a small amount of bias when delegating visible work? Does this increase the 5% antiblackness by the 10% of 40% (to like 5.4%?) EDIT: I'm not trying to be socially just in this analysis, so we can also replace being black with being openly conservative for example, or being from a different college, or not liking the same sport or same beer as one's peers. Pretty much anything can fit into this, I'm just trying to analyze this viewpoint.

wenc 7y ago
So the way I think of it in my head (and I was really trying hard to avoid being all mathematical because most people don't need this level of detail) is really to think of things, as a first approximation, as a conceptual regression problem. Let's say y = gender pay gap and x_1, x_2, .., x_n are factors. You can think of y as follows: y ~ x_1 + x_2 + ... + x_1 * x_2 + x_2 * x_3 ... etc. where you have interaction terms like x_1 * x_2 (say x_1 = being part of a non-dominant-culture, x_2 = gender) and so on. Then you can see stacking issues show up, e.g. if you're a woman you're -40% and if you're a woman of color you're -45% etc. (numbers are made up of course). It's possible to get into more rigorous modeling methods like Bayesian networks (DAGs) and hierarchical regression models (where the coefficients are dependent) but I think the added rigor/refinement in most cases is unnecessary because it often washes out due to the inherent uncertainty in the data and in most cases does not add value to the goal at hand: that is to understand the major contributors to an issue in order to design an intervention. If more folks could bring themselves to think in simple percentages, we would have moved the needle toward that goal.
bena 7y ago
The answer is we don't know what to do. Because it's already hard to figure what's a factor of factors. It's why we have to constantly be vigilant of our own subconscious biases. For instance we know that fields dominated by women typically pay less than fields dominated by men. It's a factor in the pay gap overall. Now, we could say that women chose those fields and those fields are just less desirable to the market and all that jazz. However, there is something else that has been noted. When a field becomes dominated by women, the average pay decreases and when a field becomes dominated by men, the average pay increases. So the average pay of a field is also influenced by the group that dominates it. Women aren't getting paid less because they take jobs like teaching and nursing. Teaching and nursing pay less because women dominate those fields. So there are always factors upon factors upon factors and nothing exists in a vacuum. Visibility, negotiation, perceived job performance, and even the unknown are all influenced by the person being black.

wenc 7y ago

So the way I think of it in my head (and I was really trying hard to avoid being all mathematical because most people don't need this level of detail) is really to think of things, as a first approximation, as a conceptual regression problem. Let's say y = gender pay gap and x_1, x_2, .., x_n are factors. You can think of y as follows: y ~ x_1 + x_2 + ... + x_1 * x_2 + x_2 * x_3 ... etc. where you have interaction terms like x_1 * x_2 (say x_1 = being part of a non-dominant-culture, x_2 = gender) and so on. Then you can see stacking issues show up, e.g. if you're a woman you're -40% and if you're a woman of color you're -45% etc. (numbers are made up of course). It's possible to get into more rigorous modeling methods like Bayesian networks (DAGs) and hierarchical regression models (where the coefficients are dependent) but I think the added rigor/refinement in most cases is unnecessary because it often washes out due to the inherent uncertainty in the data and in most cases does not add value to the goal at hand: that is to understand the major contributors to an issue in order to design an intervention. If more folks could bring themselves to think in simple percentages, we would have moved the needle toward that goal.

bena 7y ago

The answer is we don't know what to do. Because it's already hard to figure what's a factor of factors. It's why we have to constantly be vigilant of our own subconscious biases. For instance we know that fields dominated by women typically pay less than fields dominated by men. It's a factor in the pay gap overall. Now, we could say that women chose those fields and those fields are just less desirable to the market and all that jazz. However, there is something else that has been noted. When a field becomes dominated by women, the average pay decreases and when a field becomes dominated by men, the average pay increases. So the average pay of a field is also influenced by the group that dominates it. Women aren't getting paid less because they take jobs like teaching and nursing. Teaching and nursing pay less because women dominate those fields. So there are always factors upon factors upon factors and nothing exists in a vacuum. Visibility, negotiation, perceived job performance, and even the unknown are all influenced by the person being black.