I have a database which has individuals in different months and a target variable which indicates wheter an event happened or not, let's say:

Id: 1, Month: 1, Event: 0
Id: 1, Month: 2, Event: 0
Id: 1, Month: 3, Event: 1
Id: 2, Month: 1, Event: 0
Id: 2, Month: 2, Event: 0

As my sample is so big (think of the number of combinations), I would like to choose a sample n much lesser than N (one month per individual).

For the individuals who had the event, the selected month is always the one which had the event. Always. There is no randomness.

Instead, for individuals which did not have the event, any month could be randomly selected.

These facts...

  • Selecting one month per individual.
  • Difference between selection methods for positives and negatives (I think is called Berkson's bias).

...affect the performance of the classification model?

  • 1
    $\begingroup$ Can you clarify what you are trying to achieve ? If you are trying to predict the event selecting the month with the event for those who had the event might be a bad thing. $\endgroup$ Commented Jan 18, 2020 at 9:52
  • $\begingroup$ I am trying to predict the probability of a person having or not an event in a month. You say is a bad thing to select the month with the event for the people with the event. Why is it? $\endgroup$ Commented Jan 18, 2020 at 17:33

1 Answer 1


It would be better to frame your problem as survival analysis, rather than classification. The goal of survival analysis is predict the duration until an event happens.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.