Do you know: the students who skip classes regularly get better grades? Attractive people are more likely to be mean, and nonsmokers have more chance of getting Covid!
The instances described above are examples of what is known as collider bias or Berkson’s paradox. These typically happen in empirical studies, such as surveys. And it happens when we derive conclusions from a dataset that over-represents some groups or under-represents others.
Take the case of grades. There are four possible categories: students, 1) who attend classes and get good grades, 2) who attend and get bad grades, 3) who do not attend and get good grades, and 4) who do not attend and get bad grades. The question here is: who are those pupils less likely to participate in a survey? The last group. And the result is a bias of the sample in which, among non-attenders, the percentage of good graders dominates.
Attraction is the second example. Assume that people are classified as attractive vs not attractive, kind vs mean. Of the four possible combinations, the chances of having a date, for example, with a not attractive and mean are low, suggesting an under-representation of that combination from the list.