Hmm, idk. I don't like their modeling approach. Their dependent variable is the number of crashes per hour, scaled by the "population average" for segments of

Hmm, idk. I don't like their modeling approach. Their dependent variable is the number of crashes per hour, scaled by the "population average" for segments of the same length. I guess this is fine so long as you look at the same areas. They give two tests to try and control. One is checking upstream crashes but I kind of suspect the signage is positioned where you're supposed to be extra careful due to increase risk, so that may not be relevant. The other is looking at the same areas before and after. This does seem like a fair argument. They also include specific effects for exact date-hour combos, but I feel like this is a bad approach as you would want to ideally capture higher level weekday + hour seasonality traits; not specific date hours. They say on the multi variate: > In particular, we show that clustering by segment-year-month reduces the standard error in half, clustering by just geography produces slightly smaller standard errors, controlling for rain more flexibly does not affect our results, not controlling at all for rain doubles our estimated treatment effect, not controlling for holidays increases our estimate slightly, and dropping hours immediately before and after campaign weeks (i.e., hours outside of campaign weeks that sometimes display fatality messages, see fig. S6) further increases our estimate Which implies a lot of it may be random sample bias of the test set all being on worse weather days, even if they try to control for it here. Regardless. I think it's a plausible analysis. The effect size is just 1-2% though.