Musings on Discrimination & Affirmative Action

What is discrimination? Naively, it's something like "treating people unfairly due to categories they belong to". But this definition is incredibly vague. What makes treatment "unfair"? What is a "category"?

To see the problem, consider the sale of life insurance.

At the simplest level, the average man pays more for the same life insurance coverage than the average woman. Why? Simple: they don't live as long.

Is this discrimination?

Suppose we decide this is discrimination and we outlaw the use of sex as a way to determine life insurance cost. Moreover, suppose the actuaries at these companies have no motive or intention of discriminating against men; they just run their logistic regressions with the data they have and use the unbiased statistical estimates to set policy rates.

It is all but certain that men will still end up paying more for life insurance. Plenty of variables let the logistic models statistically guess someone's sex (e.g. height and weight).

Is this discrimination?

Let's table that for a bit and turn towards hiring.

It is (I think) generally accepted that you should hire the most qualified candidate. But what does this mean?

From a naive Econ 101 perspective, the most qualified candidate is the one that brings in the most additional accounting profit on the margin.

However, we can quickly see a discrepancy between this definition and the popular intuition. For instance, a women in her late 20s is much more likely to go on maternity leave which cuts into the profits of the firm if they offer paid maternity leave. This means that, all else being equal, the expected value from hiring a women in her late 20s is statistically lower than the value of hiring an equally qualified man or older woman.

Does this expected value differential make a preference against late-20s women "unfair"?

Needless to say, most people would consider such disparate treatment "unfair" even though this clearly impacts the ability of the woman to contribute to the firm's bottom line (her job).

The only thing clear from the above examples is that common notions of "fairness" don't have a simple definition - if they have a coherent definition at all.

The Naive Social Welfare Perspective

As is tradition, then, I will ignore conventional ethical wisdom and figure out what maximizes social welfare.

Let's start with anti-discriminatory restrictions on insurance premiums such as life insurance, health insurance, car insurance, etc.. The first thing to note is that the average cost of insurance can only increase with such restrictions since it encourages higher risk people to get insurance while driving low-risk people from the market. Therefore, absent distributional concerns, outlawing discrimination along any axis will result in a net-welfare loss.

Of course, the key phrase there is "absent distributional concerns". In an ideal society, all these variables can be used as tags in the tax-and-transfer system. This would eliminate any distributional benefit from restricting insurance companies.

Unfortunately, in the real world, we woefully underuse tagging as a redistribution mechanism because it violates widespread intuitions regarding "fairness".

Given this, the question of whether to forbid insurance discrimination along certain characteristics depends a great deal on (1) how much welfare is lost from market distortions and (2) how much welfare is gained via improved resource distribution.

When the demographic currently being discriminated against tends to have higher earnings (men, asians/whites, the educated), this makes the analysis trivial: there is no reason to forbid insurance discrimination since such a law would cause a market distortion for no distributional benefits. When the demographic tends to have lower earnings (women, blacks/hispanics, the uneducated), the answer becomes less clear.

(A similar analysis holds for credit ratings, but with interest rates standing in for insurance premiums.)

What about hiring?

The basic framework doesn't change: it's still a question of efficiency losses balancing against distribution gains. I've messed around with some models and the policy conclusions really depend on the model and parameters chosen - armchair musings won't meaningfully prevail here.

Second-Order Considerations

There are obvious second-order considerations here. For complex social reasons, methods of discernment can cause very different psychological effects.

For instance, affirmative action can make some people feel that the people who received help are systemically less qualified or less deserving. Conversely, there are conceivable social benefits to cause more intermixing of various social groups.

Returning to insurance/hiring, some forms of efficient market "discernment" strike many people as intrinsically "unfair", which is not a good state of mind to leave people in. Conversely, other groups will (in turn) see legal attempts to "correct" the free market as biased/discriminatory.

Sometimes, these debates can get quite subtle. For instance, some companies use the same bar to join regardless of category, but they make an extra effort to reach out to members of marginalized groups. Is this "questionable"?

At first it doesn't seem so, but it seems reasonable to suspect that applicant quality drops with increased outreach - this paired with an imperfect hiring filter will (statistically) cause the average employee from these groups to be less able than the average employee from other groups. In other words, the end results are qualitatively similar to more outright affirmative action - though it probably results in a better distribution-improvement-to-efficiency ratio than raw affirmative action.

All of these factors add nasty complications to an already complicated set ethical quandaries.