Thread by @Jabaluck

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 I know you had more to say but it's past my bed time so let me quickly reply before I wake up tomorrow and 30 other people have replied and changed the subject.

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 You are making several distinct (but related) claims. I want to focus on your claim about economists not being able to solve toy problems like the example you sent in order to determine what empirical strategy will identify the causal effect of interest.

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 There is a separate but related claim about whether economists would gain by estimating structural models (incidentally, in my own work, I almost always estimate structural models. I don't think they're necessary for good work, but I often find them helpful).

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 I want to focus on whether, bereft of the ability to determine which of the 4 scenarios have a legitimate IV, "X-econs" estimate models which include too many or too few controls, or select variables as IVs despite the fact that they fail to satisfy the relevant properties.

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 I claim that, with the exception of some relatively rare complex cases, good work by "X-econs" gets these issues right. I include the above paper as an example. If you read it, especially Section 3, you will see how economists handle these issues.

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 In the time you've spent trying to rationalize not reading it, you could have just read the paper already. After you do, please explain whether the authors would use a different estimation strategy if they could better solve the problems in the link you sent.

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 Understanding whether they have failed to justify their IV properly is a precursor to discussing whether they could go further with structural models.

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 If you are interested in that, you could read @autoregress 's job market paper mit.edu/~hull/RAM_0220…, which relates the estimates in the paper above to a Roy selection model.

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 But please, read Section 3 of economics.mit.edu/files/10570 to see how economists discuss and justify exclusion restrictions in a real-world setting.

@yudapearl @fuzzydunlop123 @autoregress @MariaGlymour @PHuenermund @EpiEllie @causalinf @EpiDancer @metrics52 And keep in mind: there are always trade-offs in empirical work. There is an opportunity cost to time. No one paper can do everything. The question at hand is how being able to solve your 4-variable problems would have led them to do anything differently.