The transience of health science

One of the challenges with 3 Resources/Concepts/Health is that the science seems to change constantly. Animal fats were fine, then evil, and now okay again (read: the holy of holies if you're in the keto cult). Breakfast was the most important meal of the day until the intermittent fast took its place. Coffee caused cancer, cured cancer, and who knows what it does now. We were supposed to avoid exercise to preserve energy. Now, if you don't exercise, you're pretty much committing suicide by attrition.

So too with Sleep. Neuroscientist Matthew Walker's Why We Sleep is one of the more recent examples of popular-health-science-gone-viral. The thesis is that we should be getting more sleep. Sounds reasonable enough.


Only it turns out that the book may contain a host of erroneous claims. Most interestingly, the sacred eight hours a night may be a myth. You might be able to get by fine with seven hours or even six hours. Now, if you've been sticking to the recommendations like I have, two additional productive hours a day would be life-changing. It could mean years of being awake added to an already short existence.

But they'd better be "productive". And I don't want to be the fool hopping from trend to trend at the whims of popular opinion only to have to swing back to eight hours when that again becomes fashionable. Especially when health science in particular is so fickle. How are we outsiders, too busy to go into the primary literature ourselves, to judge these claims?

Null hypothesis

First, the evidence of absence deserves greater credibility than the evidence of presence. In complex systems (like the human body) it's difficult to come up with anything that looks like a clear one-sided correlation.

So a scientific1 claim that the amount we sleep has a definitive causal relationship with mortality and cancer risk warrants greater skepticism than the scientific claim that sleep has no consistent effect in any yet-observable way (source). At least within some reasonable bounds (say 6 to 9 hours).

Walker, true to scientific heritage, took the time to respond to his critics. He reviews a host of studies that point to 6 hours really being too little: it's bad for your reaction time, your risk of diabetes, certain cancers, cardiovascular disease, your all-cause mortality, etc. Sleep is important. Still, his book has enough sloppy mistakes that we should somewhat discount the bolder claims.

Ultimately, we don't know what the ideal amount of sleep is per individual. And that's really where the fundamental challenge with health science lies. Individuals vary. A lot. What works for me need not work for you. And when the scientific method requires us to study large groups, we work with averages that may not be representative of any one individual.2

The solution is to treat yourself as test-subject. Try a bunch of different things and see what works. Space your experiments across time to recover a semblance of the scientific cornerstone of repeatability. It will never be as rigorous as the experiments in a particle accelerator or gravitational wave detector, but we don't need the same 5-sigma significance for our investigations to yield a positive effect.

The promise of the rather hype-y fields of "behavioral health" and "personalized medicine" is to make this tinkering more accessible. To make it easier to track important metrics and automate the analysis so that everyone can become the experts of themselves. If we can deliver on that promise, well, then good health need not remain the sole dominion of the elite and wealthy.

Don't mindlessly chase all the latest health trends. But also don't let the variability of health science become an excuse not to listen to new findings. Let scientific literature, old and new, serve as a starting point for your own investigations into yourself. Find what works for you, and keep on experimenting.

At the end of it all, I've learned to be less of an 8-hours-a-night zealot. And to explore a wider range of times to see what works. For now, that's enough of a lesson.



  1. I.e.: backed by experimental and observational evidence.

  2. For example, flip an unbiased coin, and average what you see. Heads = 0; Tails = 1. In the large number limit you'll approach 0.5, which is not equal to any individual sample. You can take the mode, but this will alternate between 0 and 1 infinitely many times. And if you have more options or continuous options (as is the case with most health metrics), the mode becomes uninformative or ill-defined.