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.