CICO is a weird reference class
People tend to form what I guess I’m going to call default-hypothesis-reference-classes. This is what they see as the “default” hypothesis that new hypotheses have to compete against. Obviously this isn’t how a perfect bayesian reasoner would act, but we’re only human.
These weird reference classes can be tricky, because you can get “sucked into them” and stop noticing that they’re weird.
I started thinking of this when thinking about the CICO model of weight gain/loss. Calories-in-calories-out sounds obviously true under thermodynamics. There’s about 2 × 1010 Calories in a gram of matter, and humans need about 2000 Calories per day, so make sure not to eat more than a tiny fraction of a gram per day and you’re sure to lose weight.
Well, not quite. In my opinion, the most “natural” default reference class for weight gain or loss is mass-in-mass-out (MIMO), which is obviously true, if not super useful. This has to be true—everything we know about physics tells us it’s impossible for your body to gain or lose mass unless it comes from or goes somewhere.
What people mean when they say CICO is that different foods have different amounts of bioavailable energy, and weight gain is dependent on difference between the bioavailable energy you eat and the amount of heat your body produces. This doesn’t seem like it necessarily has to be true! People gain weight when they grow, and that’s partially because they’re eating actual minerals that are deposited into their bones. (And these minerals largely don’t go anywhere once they’re deposited.) So that’s a straightforward counterexample to CICO.
We could imagine the same thing happening with fat gain or loss. Your body could store energy by accumulating fat, but not remove the actual fat molecules from your body when their chemical energy is consumed. (I’m not saying that this actually happens, just that it’s not forbidden by thermodynamics.) Or our model for what kinds of energy are bioavailable could be wrong, and if so we’d give wrong predictions when trying to apply CICO even if the theory itself were true.
What I’m trying to say is, if your previous “default” reference class for models of weight gain was MIMO, CICO would seem weird and not that special to you. CICO is very likely more useful than MIMO, but it wouldn’t intuitively seem like it should be the default theory all other theories are at best a special case of.
Once I started looking for weird reference classes, I started seeing them all the time. For example, I once saw some people discussing a study that examined the effect of new housing on “the asking price for 3-bedroom condos”. What seemed weird to me is that the study chose a refencence class, “the asking price”, which intuitively seems like not the thing we usually want to talk about, which is the sale price.
I think this one was discussed more broadly, but at the beginning of the pandemic I saw a lot of people saying “this country had good outcomes, but their population density is very low” or things like that. But population density is an imperfect metric for what we really care about, which is how clustered a population is. Canada’s population density is very low, but you could delete 90% of canada and probably only lose 10% of their population. This fictional decimated-canada would probably have had similar covid outcomes to real-canada, but with a 9x higher population density. So the reference class of “the difficulty of managing covid in a country is dependent on the country’s population density” seems like a sketchy choice for a default reference class of covid-management-difficulty prediction generators.
Population density is very easy to calculate, and as a result very easy to look up. For population clustering, it isn’t even obvious how to reduce it to a number (or manageable set of numbers) given perfect knowledge. I suspect this is why it would occur to someone to compare COVID outcomes based on population density—one might expect it to at least be correlated with the thing about which one actually cares, and it’s available.
This kind of looking under the lamppost for data is I suspect the major reason for the phenomenon you’ve identified, but not the only one. For your title example, calories in and mass in (neglecting any mass (or calories!) coming from air) are both reasonably easily available, but getting precise data on calories out seems quite difficult, and not much (maybe not at all) easier than precise data on mass out. (Mass out might actually be substantially easier if it’s okay to ignore exhaled air, sweat, etc., but I have no idea whether that’s okay; we’re typically talking about quite small changes relative to the time taken.)
I don’t have a good story for what’s going on with CICO’s appeal. Maybe something like “assuming ‘weight gain’ is ‘accumulation of matter storing bioavailable energy’, and that the body is efficient at extracting bioavailable energy from food, you get CICO”. Those assumptions might be further justified (probably in most cases without ever being explicitly considered) by common sense observations like “people who eat a lot tend to be fat” and “people who exercise a lot tend to be thin”. Perhaps the tie to common sense / shared experience allows a theory like this to spread. Even among people with generally decent epistemics, not everyone has time or interest to be a nutritionist.
This feels speculative though. I’m pretty sure the phenomena I named make sense, and they probably happen, but something else could be equally or more important. I think my main takeaway from this article is just to notice choice of reference class as a decision point, and consider alternatives rather than jumping into an assessment of where the case in question falls within the reference class.
When you lose weight, most of it exits through the breath. In the very short term, fluctuations are dominated by the matter visibly entering at one end and departing out the other, but that is noise.
Thanks! That’s a great example of how weird this topic is and how inadequate common sense is to the task of understanding it.
Anthropic clustering. I used to use this for forecasting at the beginning of the pandemic.
Sure, I didn’t mean it was impossible to put a useful number on it. But you (or whoever calculated the metric; I don’t know if that’s you) had to make a decision that 1⁄4 mile was the best distance around which to calculate, and it still apparently took days of computer time and required making up random positions for people within their census blocks. It’s absurdly complicated compared to population density, which could probably be calculated reasonably accurately thousands of years ago, if anyone cared to do so. So (I contend) population density is more available both in one’s mind, as a potential explanatory factor, and outside of one’s mind, as a statistic to try to explain some phenomenon (such as COVID results).