I mostly only write statistics stuff on this blog. In part it is understandable why sparsity is desired – such that we only have statistics in our life, and no more worries about how to...

I came up with this random thought when I was sitting in Andrew’s class this afternoon (as a TA). In order to emphasize why Bayesian is different, Andrew listed a few alternatives for data analysis,...

I saw a visualization of wealth distribution in the 2019 Credit Suisse Global Wealth Report (https://www.credit-suisse.com/about-us/en/reports-research/global-wealth-report.html)

This is trivial math question but it once bothers me in the construction of Dirichlet process using stick breaking process. The context is that if $V \sim beta(1, M)$ and $\theta \sim Dirichlet (\bar \alpha)$...

That said, xenophobism is xenophobism, no matter coated with a caucasian supremacism, or a seemingly-progressive popularism, or even when it is companied with a chauvinistic salute towards another group of straw men.

No you cannot

This is a jitt question in Andrew’s class: to (by hand) approximate the standard deviation of X/Y if X is N(5,1) and Y is N(1,3).

What if I run 4 chains and find $\hat R >>1$

Like importance sampling, ABC is in principle immune to metastability that MCMC has to suffer, but ABC is also problematic as we have no idea what it will converge to when the model is misspecified.

When I talked to someone about the old proof the invariance of odds ratio in a respropective sampling, I mentioned the estimation of $q(x)$ is achieved by its non-parametric MLE or its empirical distribution (see...

Some loss functions are invariant under domain adaption, which suggests we can indeed learn the optimal model of the population from a non-representative sample without sacrifice from the extrapolation.

The odds ratio from a case-control study is exactly the same as in a cohort study, therefore I could fit a retrospective logistic regression as if it is prospective and report its MLE or Bayesian posterior distribution. But considering the sampling distribution shift, should I reweight it regardless?

We can evaluate causal inference methods through simulations and based on predictive performance of ATE, but is it enough?

Most statisticians-orientated introduction to optimal transport theory starts with the motivation: we can always obtain a one-D distribution $f(\theta)$ through its inverse cdf transformation

We wrote a new paper on approximate Bayesian inference in deep network, with a similiar computation cost as point estimation.