TwoAndFour

UPDATE: The man in the video is ID'd as Eric F. Vinson. 30 prior assauIt arrests.

I just had an $8 milkshake and realized the one they get in Pulp Fiction was $5 and that was considered ridiculous in 1995.

We’ve just put our latest issue to bed. Christopher Caldwell has written the cover. He argues that the Iran war will likely mark the end of Trumpism:

Analysis on the Illinois congressional district MeToo accusations: *Everyone on Bluesky sympathizes with Megan Wachspress, the 42-year-old accuser, who is a lecturer at Stanford Law School. They call her brave for coming forward and talking about how she was harmed. *Everyone on X is exasperated by what they see as MeToo over-reach and sympathizes with Daniel Biss, the 48-year-old mayor of Evanston, Illinois, and a leading candidate in today's congressional primary in the Chicago area. Wachspress wrote on Bluesky and in a Substack that when she was a 20-year-old junior at U. of Chicago, Biss was her professor in a math course. He was then an assistant professor a the university and was 26. He was very friendly to her during office hours and then, after the term ended, sough a romantic course of action with Wachspress. The two dated for a little while and did some making out, she asserts. But then each of them decided this wasn't appropriate and they cut it off. Biss' campaign has since confirmed this account. Neither has reported that there was any sex between them. Wachspress says she found this, among other elements of what she described as demeaning sexism, so demoralizing that despite her mathematical prowess, she pursued a different field in her graduate studies.

I *think* I've figured out most of what Horvath did to get his results that he didn't explain. I'll explain by using his Grade 4 Mathematics scores. I took his graph and digitized it, which introduces a small amount of error because his graph's resolution was a bit low. The amount of error here is small. Then, I programmatically searched for his specification. I got pretty close by cycling through possibilities. I matched the slope and my RMSE on the points was small before the alleged digital inflection point. Afterwards, it was a bit larger, because it's less clear what he did exactly. But to get as close as I did, I had to: - Bin scores by 2-year event-time relative to his stated inflection years - Exclude 2022 (which he only theoretically, but not statistically justified) - Cut Florida's outmost bins - Use event-time as x instead of bin midpoint This delivers his exact slope: 1.08 before digital inflection and -0.28 after. But what if we undo all of the arbitrary decisions we had to use to get here? If we use bin midpoints instead of mean event-time, we move to 1.06/-0.27. Not a big effect. If we include Florida post bins in the slope, we go to +1.08/-0.33. If we include 2022, we go to +1.08/-0.76. If we include Florida 1992/1996, we get +1.08/-0.28. If we keep N = 1 bins instead of dropping them (which we shouldn't do, because they are reliable since they're whole states!), we get to +1.05/-0.25. And if we bin midpoints and all post bins, we get +1.06/-0.33. If we combine these things and conduct the most sensible analysis available to us given the data, we get +1.04/+0.10, a reduction in slope for sure, but what were we expecting anyway? If the pre-trend had held, the extrapolation would be to 256.3 points, which is higher than any individual state has ever scored. The pre-trend of +1.08/year is not a plausible counterfactual. Rather, it's a catch-up trend from the 1990s-2000s that was already decelerating (just look!) before any state had adopted digital teaching. Extrapolating it forward and treating that as real evidence of a gap or a 'loss' is attributing a ceiling effect to EdTech. Moreover, the real key to Horvath's error is this: He fumbled around with his data until he found a design that just recapitulates national trends, regardless of causality! 76% of states have inflection years in 2014-16, and he cut out the furthest entries (FL/TX), so centering on the inflection year and averaging across states is nearly-identical to centering on 2015 and averaging. The pre-trend is the national NAEP improvement from the 1990s through the mid-2010s, and the post-trend is the national stagnation and decline, and the permutation tests confirm this: randomly shuffling inflection years across states give the same pattern! In fact, Horvath made decisions that effectively guaranteed that his result would just end up being a replot of the national trends by cutting down on his power with the exclusion of FL and TX, and the consequent elimination of 36% of the variance in inflection dates. This is not a credible analysis by any means. The only really credible analysis is the synthetic control study of Florida. It's the only analysis with identifying variation because Florida adopted in 2011, four years before the bulk of the nation. And Florida *outscored* its synthetic control post-adoption! Every other analysis you can do here, from a TWFE, to permutation tests, to dose-response analyses of all kinds, is fundamentally limited by having just 49/51 jurisdictions within a 4-year window. (This is a big issue for all sorts of analyses that just use America's 50 states + D.C. They're very underpowered!) The non-fiddly construction I ran, where everything is sensible and up-front and not ad hoc (always test!) returns nulls and possibly even positive results. And Horvath's analytic methods? Well, they just weren't able to support his conclusions even if he hadn't fiddled. Advice for other analysts: never construct your way to a result. Just stick to basic data cleaning and the simplest tests you can run. You don't need to get fancy, or you might end up so confused you replot a transformed version of your observed data and call illusory trend-changes 'causal evidence.'

The 1995 Cleveland Indians lineup was loaded