Companies aren't transforming their entire strategies around the use of JIRA and Notion. The hype and FOMO has a lot of places getting far over their skis.

Not to brag, but I called this almost a decade ago. It's been wild to see the shift towards DS as quant soc sci happen.

Not an early performance, but I don't think people realize Devo could fucking rock. Solo at 2:30 is absolutely sick. youtu.be/8uK-50VmYYQ?...

I hate conferences, but I don’t hate posit:conf.

TBF, Whether the population mean is a number or a random variable is also a sticking point for a lot of people.

So reading it I think the paper is fine but a little strange. They’re doing an FDR type thing. Basically if your prior is that most A/B tests aren’t improvements, then if you get an SS result it’s more likely to be a FP. Which, like, yeah. But just go Bayes already instead of twisting like this.

I need to read it in full since I don’t even follow the motivation of the paper (how is FP different from Type I and if it isn’t what are we doing?). But I’m not surprised. In everything of his I’ve read, his statistical perspective has been somewhere between elementary and just plain wrong.

“Given our sample size the true treatment effect is not inconsistent with H0” is not a sensible statement. Neither the true treatment effect nor H0 depend on sample size. So I don’t know what he’s trying to say here. Unless this is a typo where “true” should be “estimated.”

This is how I use take-homes for hiring. It’s almost never a screen, it’s just a way for them to prep for an onsite round by getting some familiarity with the problem and coming up with initial ideas in a low-stress setting. If AI is a crutch, we find out in the onsite.

Ditto. Because there’s so much telemetry data in this domain, people are sure they can model their way to answers, but the causal issues and core data quality problems are near insurmountable. Even the incrementality/quasi-experimental stuff is extremely limited. It’s just all voodoo.

My experience w/ Marketing Science was that it was like 1 part respectable modeling and causal inference to 5 parts pseudo-quantitative bs. There’s so much money at stake w/ so much uncertainty around it that people are willing to invest a lot in tools for better decisions but mostly nothing works.

Discord is good for this.

Does wrapping it in TRANSPOSE do what you want for that case?

If your strategy is just “accurate prediction of the pretreatment trajectory implies accurate prediction of the post-treatment counterfactual trajectory” then why restrict yourself to SCM at all?