Hacker News Reader

justindomkeon UCLA professor refuses to cover for Dan Ariely in issue of data provenance

I've long been suspicious of this study from Ariely from 2003. It suggests that if you "prime" students by having them write down the last 2 digits of their student ID, then students with numbers like "93" will then be willing to pay 3.3x as much for a bottle of wine as students with numbers like "04". At first, some follow-up research seemed to replicate this, but then later work found no effect whatsoever. Very odd. https://justindomke.wordpress.com/2021/10/14/is-anchoring-a-...

justindomkeon Confidence Games

I agree that often in practice a confidence interval will end up being similar to a Bayesian credible interval. However, having a flat prior is not enough to guarantee this: the post gives an example with a uniform prior and a valid 70% confidence set that only captures 37.5% of posterior mass, and even an example with a uniform prior and a valid 70% confidence interval that captures 0% of posterior mass (because, similarly to your example, the set is empty).

justindomkeon The Human Regression Ensemble

In higher dimensions, you could do a series of 1d regressions, something like this:

1. Regress x1 against y. Call your curve f1(x1).

2. Regress x2 against (y - f1(x1)). Call your curve f2(x2).

3. Regress x3 against (y - f1(x1) - f2(x2)). Call your curve f3(x3).

4. etc.

You'd have to use your computer to read the curves and recompute the residuals each iteration, but there still isn't any algorithmic machine learning.

I suspect this would work pretty well, but it would struggle when there are strong interactions between the different input variables.

justindomkeon Markov Chain Monte Carlo: Burn-In Is Unnecessary

A helpful way to think about burn-in is that you probably don't care about bias as such but rather error which is a combination of bias and variance. Burn-in reduces bias at the cost of variance (since you have fewer points in your final estimates).

My intuition is you want to be somewhere in the middle of the tradeoff. You want to tolerate some bias -- use burn-in to get partway but only partway to the stationary distribution. Here's a paper that advances this way of thinking: http://web.stanford.edu/~ngoodman/papers/LiederGriffithsGood...