[Turneyboy]

You are assuming that the underreporting will be uniform. In reality people may be underrporting things they are embarrassed about and maybe even overreporting the opposite.

This is a flaw in the data that is much harder to account for.

[BoxFour]

Why would that be a problem for reporting relative results if everyone is under-reporting things they're embarrassed about and over-reporting the opposite?

[OkayPhysicist]

Different people are embarrassed by different things. A frat student's probably going to overstate their alcohol consumption, a Morman understate.

People with bigger appetites underestimate their food consumption, people with smaller appetites overstate.

Not to mention the degree of over/under statement will vary wildly. "A big meal" might be 300 calories for somebody with an eating disorder, or 3000+ for somebody on the opposite end of the spectrum.

[Sohcahtoa82]

> "A big meal" might be 300 calories for somebody with an eating disorder

I knew a guy that complained that he "ate like a lion" and yet couldn't gain weight.

Turns out, his breakfast was typically a single egg and a slice of toast. Lunch would be half a sandwich and a bag of chips that he wouldn't finish. Dinner of course varied, but basically was like 4-6 oz of meat of some sort and a small side of veggies.

Overall, his daily calorie intake was probably only around 1,000 calories.

I don't know if this qualified as an eating disorder, or what, considering when we hear about someone undereating, it's because they're trying to lose weight. He was trying to GAIN weight and yet was still horrendously undereating.

[BoxFour]

Sure, but in a representative sample size this is largely irrelevant. The fraternity brothers and the Mormons cancel each other out, and regardless both are dwarfed by the large middle of the population that likely systematically and reliably under-reports their drinking by a few units.

The idea of outliers and systematic biases isn’t new to statistics, relative comparisons are still useful.

[constantcrying]

>Sure, but in a representative sample size this is largely irrelevant.

There is no way to know whether your sample size is representative. What amount of fraternity brothers and Mormons cancel each other out?

>and regardless both are dwarfed by the large middle of the population that likely systematically and reliably under-reports their drinking by a few units.

And? That does not prevent spurious correlations.