I am reading
Barocas, S., & Selbst, A. D. (2016). Big Data’s Disparate Impact. California Law Review. https://doi.org/10.2139/ssrn.2477899
Among other things, the article talks about the difference ways in which big data can introduce bias in decisions taken by firm, for example, hiring whites instead of blacks just, ultimately, due to skin color.
One of this way is sample bias: suppose the manager is biased against blacks. He/she may monitor black more than white. By spending more time monitoring blacks than white, of course the manager will spot more errors / misbehavior committed by blacks than whites. Then, by relying on past data for future hiring decision, the manager will judge black as unreliable, and hire whites. So far so good, this is sampling bias.
Then at page 687 there is written:
In the employment context, even where a company performs an analysis of the data from its entire population of employees—avoiding the apparent problem of even having to select a sample—the organization must assume that its future applicant pool will have the same degree of variance as its current employee base. An organization’s tendency, however, to perform such analyses in order to change the composition of their employee base should put the validity of this assumption into immediate doubt. The potential effect of this assumption is the future mistreatment of individuals predicted to behave in accordance with the skewed findings derived from the biased sample
Although I am not understanding it. If I have the entire population, then I have no sample bias by definition. But what does it mean "perform such analyses in order to change the composition of their employee base"? How using the entire population can still introduce bias? Can you make me an example, please?