At a cursory glance, that site you linked does not appear to give any information on how it’s generating those correlations, but the term “spurious correlation” actually has a specific meaning. Essentially, one can make even statistically uncorrelated variables appear to be correlated by introducing a third variable and taking the respective ratios and finding those to be correlated instead. It should go without saying that you should make sure your correlations are actual correlations rather than mere artifacts of your analysis method. As it is, the first thing I’d do is question the validity of those correlations.
However, if the correlations actually are real, then I’d argue that they actually do constitute Bayesian evidence. The problem is that said evidence will likely be “drowned out” in a sea of much more convincing evidence. That being said, the evidence still exists; you just happen to also be updating on other pieces of evidence, potentially much more convincing evidence. So “You are not required to be stupid about it” is just the observation that you should take into account other forms of evidence when performing a Bayesian update, specifically (in this case) the plausibility of the claim (because plausibility correlates semi-strongly with truth). And to that I have but one thing to say: duh!
Bayes is definitely about statistical correlation. You can call it “updating your estimates on the basis of new data points” if you want, but it’s still all probabilities—and you need correlations for those. For example: if you don’t know how much phenomenon A correlates with phenomenon B, how are you supposed to calculate the conditional probabilities P(A|B) and P(B|A)?
Bayes is definitely about statistical correlation.
No, I strongly disagree.
it’s still all probabilities—and you need correlations for those
I do not need correlations for probabilities—where did you get that strange idea?
To make a simple observation, “correlation” is a linear relationship and there are many things in this world that are dependent in more complex ways. Are you familiar with the Anscombe’s quartet, by the way?
I do not need correlations for probabilities—where did you get that strange idea?
In that case, I’ll repeat my earlier question:
if you don’t know how much phenomenon A correlates with phenomenon B, how are you supposed to calculate the conditional probabilities P(A|B) and P(B|A)?
There is no general answer—this question goes to why do you consider a particular data point to be evidence suitable for updating your prior. Ideally you have causal (structural) knowledge about the relationship between A & B, but lacking that you probably should have some model (implicit or explicit) about that relationship. The relationship does not have to be linear and does not have to show up as correlation (though it, of course, might).
At a cursory glance, that site you linked does not appear to give any information on how it’s generating those correlations, but the term “spurious correlation” actually has a specific meaning. Essentially, one can make even statistically uncorrelated variables appear to be correlated by introducing a third variable and taking the respective ratios and finding those to be correlated instead. It should go without saying that you should make sure your correlations are actual correlations rather than mere artifacts of your analysis method. As it is, the first thing I’d do is question the validity of those correlations.
However, if the correlations actually are real, then I’d argue that they actually do constitute Bayesian evidence. The problem is that said evidence will likely be “drowned out” in a sea of much more convincing evidence. That being said, the evidence still exists; you just happen to also be updating on other pieces of evidence, potentially much more convincing evidence. So “You are not required to be stupid about it” is just the observation that you should take into account other forms of evidence when performing a Bayesian update, specifically (in this case) the plausibility of the claim (because plausibility correlates semi-strongly with truth). And to that I have but one thing to say: duh!
Bayes is definitely about statistical correlation. You can call it “updating your estimates on the basis of new data points” if you want, but it’s still all probabilities—and you need correlations for those. For example: if you don’t know how much phenomenon A correlates with phenomenon B, how are you supposed to calculate the conditional probabilities P(A|B) and P(B|A)?
No, I strongly disagree.
I do not need correlations for probabilities—where did you get that strange idea?
To make a simple observation, “correlation” is a linear relationship and there are many things in this world that are dependent in more complex ways. Are you familiar with the Anscombe’s quartet, by the way?
In that case, I’ll repeat my earlier question:
There is no general answer—this question goes to why do you consider a particular data point to be evidence suitable for updating your prior. Ideally you have causal (structural) knowledge about the relationship between A & B, but lacking that you probably should have some model (implicit or explicit) about that relationship. The relationship does not have to be linear and does not have to show up as correlation (though it, of course, might).