I mean USC can do whatever they want, and it makes sense they don't want to risk a loss, but more teams have been helped by a win over ND than hurt by a loss. If ND and Miami replace their h2h with a cupcake this year, ND is like, 8th?, and Miami is out of the playoff.

In the particular scenario you gave above, the committee would just bump USC up/ND down to whatever positions are required to get the h2h winner in. Maybe some other team is collateral damage (eg they move USC up to 10 and some random team at 11 gets hosed if ND stays at 12) but not USC.

Last 3 years I think it was:
-Texas over FSU (About 6 points apart)
-SMU over Ole Miss (8)
-Bama over ND (14)

Maybe if you do some as hoc smaller penalty for ccg losses it's a bit less dramatic past two years but wouldn't come close to flipping things

What are the biggest ever resume SP+ gaps between an at large CFP team and a team that didnt make it?

Ah ok, didn't think about how the average scoring level based on the other scores this week would play in. Lotta moving parts! Thanks

What's the intuition for why the ND-Miami statistical tie (against a pre-game expectation of a statistical tie) was so much worse for ND (6 point rating drop) than it was good for the Canes (basically unchanged rating)?

Not arguing with the model that basically sets the spread, just wondering

Garbage time is the extreme version of this--teams stop trying altogether--so I know you throw that out.

But it seems like game states could matter beyond garbage time. IIRC from your 2017 post, teams up <14 in Q4 have ~ the lowest success rate of any game state? Does PGWE adjust for that?

--That means their predictive stats were weaker than you'd expect for a team that won by 13.
--But it doesn't necessarily indicate that you'd expect them to lose with those stats *given the game states in which those stats arose*.

How does PGWE handle game states?

An arg you could make about the ND-UGA game might be something like:

--ND got in the lead due to fluky things happening.
--Once they were in the lead, they sacrificed EV to limit variance.
--This suppressed their explosive rate, success rate, etc.

Figuring out how much to scale is pretty hard though (as you say, it probably differs by disease...probably also differs by location/context as well).

Cool, will be on the lookout (don't think i can send DMs your way but should be able to receive).

We often use cause-specific burden ests when estimating effects of an intervention targeting a particular disease, then scale that up to account for the wedge you highlighted in the sepsis example.

That makes sense--and whoa on the sepsis numbers.

Would love to chat more. I have been thinking about related Qs on the research team at GiveWell. DMs open!

Cool thread!

Any thoughts on the cause vs. risk factor distinction in epidemiology / how much nuance is lost by IHME-style one-death-one-cause models? Have always heard it's a simplification, but never known how much of one.

Hey Bill, related to HFA:

I was curious about timezone effects now that we have national conferences. Are teams producing above expected HFA when the opponent is traveling 3 time zones?