The day in conversation for 2026-11-06

discussion

Transcript

The discussion: Ruth: That paper on model selection in high-dimensional regression sounds like a significant step for practical econometrics; what specifically about integrating forward stepwise selection with the multiple testing framework makes it distinct from standard penalized regression methods? Iris: I'm more interested in the oracle properties mentioned; those properties are usually conditional, so what are the exact assumptions regarding the sparsity and correlation structure of those high-dimensional predictors that allow you to guarantee that property holds consistently across different datasets? Warren: It seems like a necessary trade-off between screening speed and greediness, which is something I've been thinking about when we look at prior work on variable selection; does this method offer a clearer path toward identifying truly causal predictors versus just strong correlators in these complex settings? Ruth: Exactly, and connecting that back to earlier days, if we look at the sensitivity analyses from last week on time-varying coefficients, I wonder if this new framework handles the non-stationarity of those proxy variables better than the static assumptions underlying the original model selection literature. Iris: If we assume strong screening properties are maintained, how does that method compare practically to methods focusing purely on information criteria like AIC or BIC when dealing with hundreds of potential predictors where false positives are a major concern? Warren: The synthetic control work mentioned earlier—the one using the information-theoretic measure—suggests a different angle for causal inference, but I'm curious if there's any overlap in the underlying assumptions regarding the constraints. Is the relaxation approach to synthetic control leveraging similar affine inequality structures found

More episodes

← Home