3 ms·

This is not a random experiment, so statistical methods (Bayesian or other) do not apply. I really wish that this meme of throwing around big words like "Bayesi

by fsh 5y ago

This is not a random experiment, so statistical methods (Bayesian or other) do not apply. I really wish that this meme of throwing around big words like "Bayesian reasoning" or "Priors" completely out of context will disappear at some point.

shoto_io 5y ago
Sorry, if I used it out of context. I was under the impression that virus formation (copying, mutations and all that) was indeed a statistical process. But maybe I am wrong?
- fsh 5y ago
  Statistical modelling assumes that the observations can in some way be described as random samples drawn from an (unknown) probability distribution. One can then try to estimate properties of that distribution using the observed data. For example, one can assume a gaussian distribution and estimate its mean and width using the mean and standard deviation of the data. Bayesian inference aims to directly extract the underlying probability distribution without assuming a fixed functional form. In any case, one should be extremely skeptical when someone tries to apply such methods to sparse or cherry-picked data. Including SARS-CoV-2, there are only seven coronaviruses that are known to infect humans. This is an extremely small number to do statistics with. Of course one can make educated guesses based on experiences with previous similar viruses (such as SARS-CoV). But using the language of Bayesian inference for this is just cargo cult science in my opinion.
  - rubatuga 5y ago
    Having 6 previous coronaviruses allows you to construct a prior with mean and variance. Gaussian distributions assume the least about a distribution. I'm not sure why you are so against using information/bits that we already know to be true in trying to make predictions about SARS-CoV-2. Humans do bayesian reasoning all the time, albeit in a much less formal fashion. There is no cargo culting here.
    - fsh 5y ago
      Of course you can construct all kinds of interesting distributions with any random piece of data. This doesn't mean that it makes any sense to do so. Pretending that human reasoning is somehow equivalent to Bayesian inference (and abusing domain-specific terminology like "prior") is exactly the kind of cargo culting that Feynman wrote about. If there is no actual mathematical model and software implementation, you are not doing Bayesian inference (and the airplanes will just not land).
      - shoto_io 5y ago
        I am unsure what you are suggesting instead. Should we ignore everything we know from the past?

by fsh 5y ago

shoto_io 5y ago

Sorry, if I used it out of context. I was under the impression that virus formation (copying, mutations and all that) was indeed a statistical process. But maybe I am wrong?

fsh 5y ago
Statistical modelling assumes that the observations can in some way be described as random samples drawn from an (unknown) probability distribution. One can then try to estimate properties of that distribution using the observed data. For example, one can assume a gaussian distribution and estimate its mean and width using the mean and standard deviation of the data. Bayesian inference aims to directly extract the underlying probability distribution without assuming a fixed functional form. In any case, one should be extremely skeptical when someone tries to apply such methods to sparse or cherry-picked data. Including SARS-CoV-2, there are only seven coronaviruses that are known to infect humans. This is an extremely small number to do statistics with. Of course one can make educated guesses based on experiences with previous similar viruses (such as SARS-CoV). But using the language of Bayesian inference for this is just cargo cult science in my opinion.
- rubatuga 5y ago
  Having 6 previous coronaviruses allows you to construct a prior with mean and variance. Gaussian distributions assume the least about a distribution. I'm not sure why you are so against using information/bits that we already know to be true in trying to make predictions about SARS-CoV-2. Humans do bayesian reasoning all the time, albeit in a much less formal fashion. There is no cargo culting here.
  - fsh 5y ago
    Of course you can construct all kinds of interesting distributions with any random piece of data. This doesn't mean that it makes any sense to do so. Pretending that human reasoning is somehow equivalent to Bayesian inference (and abusing domain-specific terminology like "prior") is exactly the kind of cargo culting that Feynman wrote about. If there is no actual mathematical model and software implementation, you are not doing Bayesian inference (and the airplanes will just not land).
    - shoto_io 5y ago
      I am unsure what you are suggesting instead. Should we ignore everything we know from the past?