4 ms·
At best, only approximately Gaussian. Height cannot be Gaussian if only because height is never negative but every Gaussian density is positive for all real nu
by graycat 5y ago
At best, only approximately Gaussian.
Height cannot be Gaussian if only because height is never negative but every Gaussian density is positive for all real numbers, including negative real numbers.
"interactions of many genes that are measureable"
The interactions of many is commonly used to justify a Gaussian assumption, but there is no such theorem. The main theorem to get a Gaussian density is the central limit theorem which usually assumes an infinite sequence of independent identically distributed random variables (plus some), i.e., i.i.d. The i.i.d. is asking a LOT!!! Asking so much that in practice it is essentially unrealistic.
What people often mean by Gaussian is just a central peak, some symmetry, and long tails, but an actual Gaussian distribution has a lot more, e.g., sufficient statistics (as in a classic paper by Halmos and Savage, with a derivation in M. Loeve's two volumes, from Springer, Probability) which, as I recall, E. Dynkin showed are quite sensitive to deviations from actual Gaussian.
Gaussian is important in a lot of derivations -- a LOT is known. But in practice Gaussian has some utility but only as an approximation where don't need to be very careful about accuracy.
- dekhn 5y agoI am not really disagreeing with you, but for a first order approximation, treating human height as a bunch of iids is the simplest approach. https://www.nature.com/articles/s41431-021-00836-7 https://www.nature.com/articles/s41431-021-00836-7. "At the same time, it is well-understood that additivity of effects and normal distribution of residuals is only an approximation to the distribution of height in human populations. .... However, up until now, no evidence for non-additive genetic interactions was found for height [6, 18]. Moreover, if non-additive effects were to exist for polygenic traits, very large sample sizes would be required to detect them [9]."
- graycat 5y ago> treating human height as a bunch of iids is the simplest approach. Sounds fine. If you treat heights as some i.i.d.s, then can use the central limit theorem to argue that in the limit for large samples (we get convergence in distribution) the probability density distribution of the sum of heights expressed as a z-score (mean 0, standard deviation 1) has Gaussian probability density distribution. But that does not mean that the probability density distribution of heights is Gaussian. Indeed, if include both males and females, then for the density likely get two peaks instead of just one. If also include children, then get a left tail longer than the right one. So, net, we cannot expect that heights are Gaussian. And we will have a tough time finding a large population that is Gaussian. z-scores in the limit for large samples Gaussian -- sure, can get that. A large population Gaussian, not much chance.