Earlier today I spent some time shitting on this paper without a broader discussion of why I was doing so I shall hold forth on this matter now as I think it is revealing of the absolute state of empirical social science and microeconometrics in particularhttps://twitter.com/RRHDr/status/1295488617125687297 …
-
Show this thread
-
I am not going to sink into a detailed methodological critique because 1) while I hate their antique methods this is unnecessary to make my point, and 2) I am writing for a general audience and want to demonstrate that any idiot can reject papers that uses such methods badly
1 reply 3 retweets 83 likesShow this thread -
heres some text from the methods section along with some of the figures from the paper depending on your background you may find your eyes glazing over before opening it or you may be shouting imprecations at your phone good news if your former! it doesnt fucking matterpic.twitter.com/Kowy5rCDSn
13 replies 0 retweets 61 likesShow this thread -
Replying to @eigenrobot
do they cluster @ hospital level later on? seems like you should be clustering at that level, not physician, the fact that physician fixed effects halves the effect is also sus
2 replies 0 retweets 4 likes -
Replying to @decadantism @eigenrobot
I will defend the use of OLS here, that in itself is not particularly objectionable ... no idea where that 3x number is coming from though
1 reply 0 retweets 3 likes -
Replying to @decadantism
Partly an error (real but insignificant to my broader point imo) from relying on the cnn article rather than actually computing the coefficient ratios myself (Im extremely lazy) You really don't have a problem with an LPM for events with a probability < 0.01? (Actually curious)
1 reply 0 retweets 1 like -
Replying to @eigenrobot
my intuition is that the lpm should be better if the event is quite rare + sample sizes r small, but hadn't thought about that – here's a table from a paper I found that argues LPM > logit for larger samples: https://pdfs.semanticscholar.org/afe3/bb3575d8d5a493cfac36e190788350ef85f1.pdf …pic.twitter.com/eTofDwQQWp
2 replies 0 retweets 1 like
Honestly I haven't run anything but a forest for classification in years
-
-
Replying to @eigenrobot
fully convinced that forest/etc > ols for lots of cases, mostly b/c I don't know enough! but hand-wringing over non-linearity in itself re: probit/logit I feel is overblown
0 replies 0 retweets 1 likeThanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.