🎯How can we learn to target interventions when we care about outcomes that are only observed after a delay?🗓️

This comes up in settings as varied as clinical trials (5 yr all-cause mortality) and marketing (customer lifetime value).

Our new paper:
arxiv.org/abs/2010.15835 We were motivated to work on this through a collaboration with The Boston Globe, helping them retain subscribers who might otherwise churn by targeting "thank you" emails and discount offers.

(Nice to do marketing research that helps journalism make a business model transition.)

Can Twitter's changes to add slightly more friction to retweeting have substantial effects?

Well in 2012 we randomized which Facebook posts are shown with collapsed vs. expanded comments to >900 million people.

Note that the number of clicks to like or comment remain the same. This then affected how much feedback (likes + comments) people received. This is on a log scale, so the effects are around 1%. Not nothing, but not huge either.

Part of the Cambridge Analytica scandal — or what @kmmunger calls the meta-scandal — is how much the media amplified dubious claims about the the effectiveness o what was a mix of standard practice, ill-gotten but not so useful data & hype
newstatesman.com/science-tech/s… But it is worth giving credit to journalists who got this right all along by, as @kmmunger notes, talking with social scientists

https://twitter.com/deaneckles/status/1314311734853873666

This NYT piece by @qdbui @ShaneGoldmacher says this is who the campaigns are "targeting" and "trying to reach". But this is who they actually reached, which depends as much on competition in the auction and FB's machine learning as anything else. Let me spell this out... @qdbui @ShaneGoldmacher Advertisers can specify which people they are willing to pay for impressions and clicks from. But then whether those people see the ads depends on who else is trying to show them ads and FB's estimation of, e.g., whether they will click. This is sometimes called "optimization".

What's the best PhD-level introduction to counterfactual policy learning, dynamic treatment regimes, and the like? Some options I've considered follow... Chapter 10 of Manski's Identification for Prediction and Decision
amazon.com/Identification…