algorithmwatch / 2020-monitoring-instagram-analysis Goto Github PK

Did you get a chance to control for the number of likes a post had when your donors encountered it? If I understood correctly, you successfully proved that racy/nude posts are displayed more often. However, Facebook and Instagram are known to feature posts with higher interaction rates. Thus this might be a case of a "chicken or egg" problem: If posts showing bare skin are more successful in generating likes/interaction among the users who first encounter the post, said posts will consequently most likely be more heavily featured in other users' newsfeed. Note that the reason for this would then not be that they show bare skin but that they have proven to be more popular/"relevant" to other users so far.

Here https://github.com/algorithmwatch/monitoringinstagram/tree/master/analysis#question-2-label-analysis a very significant effect is found. I think this is due to assuming that the created and encountered posts are independent and identically distributed (IID). As shown later, the encounters differ per donor, thus the IID assumption is violated. Violating the IID assumption makes the statistical test unreliable. I think it would be more meaningful to perform the test on data aggregated at the donor level.

PS: Thank you for making this analysis transparant!