原题http://www.glassdoor.com/Interview/You-have-results-from-an-A-B-experiment-You-applied-some-treatment-X-and-want-to-measure-the-conversion-rate-Y-How-do-you-QTN_802647.htm
You have results from an A/B experiment. You applied some
treatment X and want to measure the conversion rate Y. How do you control for confounding variable Z? Compute P(Y|X) after adjusting for Z.
Given that a confounding variable throws a wrinkle in the relationship between our dependent and independent variables, we have to take Z (the confounding var) into consideration.
Pretend Z didn't exist, our probability would be straight forward: P(Y|X), which is P(Y,X)/P(X).
Since Z is causing trouble our new formula will be: Sum P(Y| X,Z)*P(Z)
So a hypothetical: Let's say Y = Revenue Dollars, X = Ad, Z = Gender (Male/Female)
Our equation would be: P(Revenue | Ad , Male) * P(Male) + P(Revenue | Ad, Female) * P(Female)--plug in hypothetical numbers if you'd like and done!
plus http://venturebeat.com/2014/12/18/4-questions-that-reveal-whether-you-have-the-hottest-job-skill-of-2014/
According to wage data startup, Glassdoor, this was an actual interview question for a Facebook data scientist position. The most popular answer on the forum was “Sum P(Y| X,Z)*P(Z)”. To translate the question into English, a “confounding variable” is just something that makes it seem like y causes x, when really it doesn’t.
So, for instance, if there is a correlation between murder rate and ice cream sales, it’s not that selling ice cream inspires violent rampage. It could just be that warmer weather causes both. To answer this question, an interviewee could also say that they would “control” for the Z variable in the common statistical procedure known as a “regression.”