# Bias does not imply discrimination

Many people have seen this quotation:

As a practical example, consider this equivalent quotation:

Bias does not imply discrimination

Under a certain (specific) definition of “bias” and “discrimination”, the second quote is just as accurate as the first quote. Under these definitions, there is a solution to the problem where the data is analyzed, a bias is located, and therefore discrimination is (incorrectly) charged. Under these definitions, “bias” is acknowledged as a real thing (just as “correlation” is a real thing), but it is disconnected from the automatic (incorrect) conclusion that there must be discrimination.

(Note: there is not agreement of any kind on using these definitions. An informal look at articles mentioning “bias” and “discrimination” shows many articles simply conflate the two terms to mean the same thing. Then, proceed to show why in their particular case, that correlation does imply causation.)

Definitions (from “Sex Bias in Graduate Admissions: Data from Berkeley“, Science, Bickel and Hammel, as referenced by “The Book of Why“, Judea Pearl, chapter 9, page 311):

- Bias: “… we mean here a pattern of association …”
- Discrimination: “… we mean the exercise of decision …”

As stated by Pearl, “association” aka “correlation” is on rung 1 of the ladder of causation, and “decision” aka “doing” or “intervening” is on rung 2. And nothing in the rung 1 toolset can lead to a rung 2 conclusion (aka “correlation does not imply causation” is not “sometimes correlation proves causation”).

Some further notes:

- Fun little article titled “Does Statistical Bias Equal Descrimination?”
- Table and weighted average treatment of Berkeley data.
- Vocabulary note: “statistical bias” is close, if you emphasize the meaning “bias as detected only by using the data”, aka rung 1 reasoning.