2
$\begingroup$

Suppose we have a dataset with many (~100) features and a binary outcome. I am interested in not only assessing whether any given feature is causally related to the outcome, but in actually being able to say that users X,Y,and Z got the outcome specifically because of some specific feature(s). Is this even possible? If so, could someone please give me some pointers as to go about it (math, codes, packages, etc.)?

$\endgroup$
  • 1
    $\begingroup$ Welcome to CV. There is no math or code to determine causality. You should have an intervention study to make any claims of a causal relationship. The closest you could get is probably estimating an undirected graph of partial correlations, and then using a PC algorithm to determine direction here and there, but this still does not demonstrate causality, that is up to the study design. $\endgroup$ – Frans Rodenburg Jul 11 '19 at 11:54
  • 1
    $\begingroup$ Here is a math pointer: Pearl's work on causality. ftp.cs.ucla.edu/pub/stat_ser/r350.pdf. Page 136 and onward discusses whether and when questions like "did Z cause Y?" can be answered. $\endgroup$ – CloseToC Jul 11 '19 at 12:02
1
$\begingroup$

Yes, this is the problem of "probability of causation." It is not a major topic in the causal inference literature, but there has been some written on it. A Google search led me to this CDC page on interpreting probability of causation and this paper (Dawid, Musio, Murtas, 2017) on contemporary perspectives in probability of causation. The concept is used when attempting assign blame or responsibility for an event to some causal actor.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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