Hacker News Reader

Comment by lm28469 | Hacker News Reader

> Is coffee good for you? What about wine or chocolate? Scientists trying to answer these questions

There is a virtually infinite amount of cofounding variables, genetics, meal timing, fitness level, sedentarity, &c. . It's a 80/20 type of problem, do the 80, forget about the 20, you'll never be able to get your answers anyways.

If you look and feel like shit you're most likely eating like shit. If you look and feel good a glass of wine every now and then or a bite of chocolate after dinner won't do much.

View on HN

You reduce the uncertainty of the remaining 20 by substantially increasing sample size across a randomly selected sample.

Unfortunately for these studies you have multiple selection criteria that are nonrandom:

(1) interest in the study

(2) adherence to protocol of the study

(3) reporting back in

If nutrition science wants to be serious, their N should not be in the 10s but rather the 10,000s.

That has an expense, but for important things it is absolutely the right thing to do.

Until they track absolutely everything including each trial subject microbiome, hormone profile, &co over time, I still feel it just won't cut it.

Plus it doesn't even matter what is true for the statistical average, given the infinite amount of variables and outcomes one glass of wine might be statistically beneficial but absolutely terrible for your own health because you have one specific gene combination or one specific microbiome mix. Which means you'd have to go through the same regimen of analysing and tracking all the parameters for yourself for it to be applicable

Actually, this is why stats exists in the first place. Larger samples (including metastudies) are so powerful -- you can measure and predict causal impact of test factors even if you can't control for unobservables. The goal is to minimize type 1 and type 2 error. So long as those unobservables are not driving a selection bias, you get wonderful things like the central limit theorem coming to the rescue.

No one can monitor or measure everything, whether philosophically (Heisenberg uncertainty principle) or prosaically (cost). But if something is true, we can often probe it enough to get at least a low-res idea of the nature of it. This moves us light years ahead of primarily using our personal experience, gut, and vibe to establish epistemologically sound assertions.

This somewhat feels like 2 layer neural networks are a universal predictor.

It is true in the limit but not useful in practice.

When it comes to studying food / diet studies really do need to be a lot more careful about trying to control their variables.

Nutrition science as a field seems to have few absolute truths and many many overturned papers / results.

Central limit theorem is, without qualification, useful in practice.

I suspect (I'm not an expert) that for subjects like nutrition, experimental psychology and so on the next big step forward isn't scientific but political: figuring out how to somehow get funders, researchers and others lined up behind a Big Science model where a very few organisations run experiments with those truly large participation numbers. There are obvious risks in switching to such a model, but if small or middling experiments simply can't answer the open questions then there may be no better alternative.

or you're sleeping like shit. or you have an autoimmune disease. or you're depressed. or you have an ongoing inflammatory state from a lingering virus. etc

Reply on news.ycombinator.com