When do we end up with a "crap ton" of instrumental variables suitable for use with ML causal inference methods? A list — and a call for other suggestions... https://twitter.com/Andrew___Baker/status/1277740175859179520

When we have many randomized experiments and are doing an integrative meta-analysis
https://dl.acm.org/doi/10.1145/3178876.3186151 @alex_peys

When we want to allow for heterogenous effects of the low-dimensional instruments on the endogenous treatment, but stick with a linear model (e.g. with lasso). Though in this application it didn't make much difference. https://www.pnas.org/content/113/27/7316

When we have a lot of candidate instruments (e.g., genes), but many might be invalid. You can even allow that only a plurality are valid. This comes up in so-called Mendelian randomization https://rss.onlinelibrary.wiley.com/doi/abs/10.1111/rssb.12275

When there are 743 credit limit discontinuities https://twitter.com/arpitrage/status/1277744015232897031

When we are using weather as an instrument (this is like 70% of all IVs, right?), an instrument for every NAICS industry, and aren't sure how much different levels of (eg) rain matters and how that interacts with other variables https://osf.io/b9psy/