Hacker News Reader

Code review is not the way to validate statistical inferences.

Even during internal review of statistical inference code, committers argue correctness through a combination of readable code, readable tests, and statistical validation on a combination of synthetic and real data. While I agree the TrueAllele source should be provided, the "zero-defects" bar is neither sufficient nor necessary for correctness of the inferences made.

View on HN

> Code review is not the way to validate statistical inferences.

But it is a way to invalidate the inference, which is what the defense is interested in.

How much uncertainty in the system are you willing to live with if you were falsely accused of murder?

That is a question for the judge or jury. The statistician's job is to provide a probability distribution, and to argue for the statistical correctness of that distribution based on similar inferences made on validation data.

In that case, at least with respect to the death penalty, "zero-defects" is precisely what the courts have moved towards.

Reply on news.ycombinator.com