Lightnews — Scholar-powered news

Hmm, what is the gap in the go for 2 down 8? This old article by Walder says it's 59% win percentage if going for 2/47% for the XP. Using Walder's analysis, it's 50%/50% for this other case. These numbers are not precise enough to be different.

Going for 2 down 8 points: Packers pull off come-from-behind win thanks to this strategy

The Packers profited from being the first team this season to try this unconventional 2-point conversion attempt. Here's why others should follow their lead.

www.espn.com

Ben Recht @beenwrekt.bsky.social · 2d

Why do you take psychology into account here but not when computing the odds of winning when going for 2 down 8?

1

Ben Recht @beenwrekt.bsky.social · 2d

2

Ben Recht @beenwrekt.bsky.social · 3d

The majority of Koopman's article argues for why "at fault" metrics are deeply flawed. It's a compelling argument and worth a full read.

1 2

Ben Recht @beenwrekt.bsky.social · 3d

In non-LLM AI news, Waymo has quietly been involved in two fatal car crashes in 2025. Phil Koopman details that this fatality rate is now comparable to humans.

People: 7/billion miles in all conditions
Waymo: 5/billion miles in geofenced conditions

Blame-Free Robotaxi Crashes Are Still Crashes

The tricky logic behind robotaxi crash metrics

philkoopman.substack.com

3 12 42

Ben Recht @beenwrekt.bsky.social · 3d

he's so tiny!

1

Ben Recht @beenwrekt.bsky.social · 3d

Yes, and though I'm not inclined, it behooves technocrats to embrace this position and make the case for bureaucracy as a public good.

(But also reiterating: there are non-bureaucratic uses of statistics. They just don't try to reduce everything to RCTs or counterfactual reasoning.)

1 3

Ben Recht @beenwrekt.bsky.social · 3d

Yes, I highly recommend this paper. A masterpiece.

In contrast to much potential outcomes work, Dawid is upfront that we only do experimentation with the expectation that there is some form of external validity, which necessitates modeling how the sample relates to future individuals.

1 8

Ben Recht @beenwrekt.bsky.social · 3d

Yes, you can, but this is not necessary.

Ben Recht @beenwrekt.bsky.social · 3d

we don't even have to think about infinite samples. Imagine you have 52 cards laid out face down on a table, and you want to estimate the prevalence of hearts. You sample 10 of them. It's not hard to imagine checking your answer on the other 42. 1/2

Ben Recht @beenwrekt.bsky.social · 3d

But what good is lower variance if the estimand isn't real?

3 6

Ben Recht @beenwrekt.bsky.social · 3d

My issue is that George Box ruined the brains of all subsequent statisticians who torture his quote to equate "useful" with "usable."

2

Ben Recht @beenwrekt.bsky.social · 3d

In an RCT, I can either give someone a drug or not. There is no world where I can observe both. 2/2

1 3

Ben Recht @beenwrekt.bsky.social · 3d

we don't even have to think about infinite samples. Imagine you have 52 cards laid out face down on a table, and you want to estimate the prevalence of hearts. You sample 10 of them. It's not hard to imagine checking your answer on the other 42. 1/2

1 4

Ben Recht @beenwrekt.bsky.social · 4d

Sure, but it behooves us to convincingly argue why we should take regression coefficient seriously. Too often, people compute them because it's what they are trained to compute.

3

Ben Recht @beenwrekt.bsky.social · 4d

In ITEs, you can never apply both treatments. You are reasoning about something that can't exist. 2/2

3 2

Ben Recht @beenwrekt.bsky.social · 4d

There is a fundamental difference in the missing data problems of superpopulations vs ITEs.

You can imagine superpopulations as drawing cards from a shuffled deck. It is possible to draw every card if you sample long enough. 1/2

2 4

Ben Recht @beenwrekt.bsky.social · 4d

I'm ok with this split, to be honest. In the first camp, would you agree that "gold standard" means "all causal inference is an RCT, either real or imagined?"

FWIW, I have more gripes with the missing data interpretation because ITEs are purely metaphysical.

1 2

Ben Recht @beenwrekt.bsky.social · 4d

There are many ways the gold standard thing gets tossed around, but the Rubin school wants you to imagine that *all* observational studies are simulated RCTs.

It's why their whole system is broken.

2 3

Ben Recht @beenwrekt.bsky.social · 5d

Tying together many previous threads on the blog, I respond to Ben Chugg’s critique of “A Bureaucratic Theory of Statistics.”

How do know so much about swallows?

A dialogue about the use and meaning of statistics with Ben Chugg

www.argmin.net

6