Distortion that occurs when the data you analyze are not representative of the population you want to draw conclusions about, because of how units ended up in the sample. Survivorship bias, non-response bias and self-selection are common forms.
Card 1 of 244. Answer: Distortion that occurs when the data you analyze are not representative of the population you want to draw conclusions about, because of how units ended up in the sample. Survivorship bias, non-response bias and self-selection are common forms.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Selection bias comes from the process that decides which observations you get to see. If that process is related to the outcome, any estimate from the sample is skewed, and more data of the same kind doesn't fix it. Common forms: self-selection (volunteers differ from everyone else), non-response, attrition from a study, and survivorship (only the winners are left to observe).
Example: in 1936 the Literary Digest polled its readers plus people from phone directories and car registrations, and got about 2.4 million replies predicting that Alf Landon would beat Franklin Roosevelt. Roosevelt won about 61% of the vote. Phone and car owners were wealthier than average in the Depression, and the people who chose to reply differed again. Size couldn't rescue a biased sample.
In machine learning it hides in the labels. A credit model trained only on approved loans never sees how rejected applicants would have repaid, and a recommender trained on clicks only learns about items it chose to show. Remedies include random sampling or a small random holdout, reweighting by the inverse probability of selection, and modelling the selection itself, as in James Heckman's 1979 correction.
