The day in conversation for 2026-11-08

discussion

Transcript

The discussion: Tom: So this work on decomposing p-values into components independent under the null hypothesis sounds powerful, but what's the practical implication for testing hypotheses when we already have selection bias inherent in how those p-values are generated? Jane: It’s about isolating the part of the test statistic that isn't inflated by that selection process. I was thinking back to our work on sequential analysis; did this selective inference framework offer a cleaner way to condition the error rates than the traditional bootstrap approaches we used last month? Lu: From my perspective, the technical elegance is in defining those components independent under the null—it moves us closer to a rigorous framework for Type I error control in complex sampling scenarios. But how robust is this decomposition when the underlying selection mechanism itself is highly non-linear? Meng: That’s a fair point on non-linearity. I wonder if we can actually compute these components practically, or if the complexity of identifying the "independent" part makes it computationally intractable for real-world application compared to simpler methods? Lalam: We need to look closely at how they define that independence. Does it rely purely on the structure of randomization, or does it require assumptions about the distribution of those p-values conditional on selection? That dependence is where I think the practical hurdle lies. Tom: Exactly. If we lean too heavily on structural assumptions, we risk losing accuracy. Speaking of accuracy, what about that GMSE metric from the multisource register work? It seems like a way to quantify global error beyond just point estimates, which addresses something different than conditional error rates. Jane: The

More episodes

← Home