Suppose I see a hypothetical click-bait article that says: “One thing you can do that science says doubles your chance of living past a hundred.” I foolishly click on it, and from the first paragraph find out that supposedly that thing is a serious photography hobby. Now, the idea isn’t crazy: having a serious hobby that can be pursued over a lifetime and that involves artistic and intellectual skills could well increase your lifespan. But I am reasonably sceptical.
Suppose for simplicity that after reading the first paragraph, I assign a credence of 1/2 to the null hypothesis N that there is no correlation between photography and living to a hundred and a credence of 1/2 to the hypothesis H that serious photography doubles the chance of living past 100.
Suppose I now tell my mother about the article, and she says: “Your great-grandma Alice was an avid photographer and she lived past 100!”
This is paradigmatically anecdotal evidence. But let’s do a quick and dirty Bayesian analysis. I just learned the fact E1 at least one of my eight great-grandparents was both a serious photographer and lived past 100. Let’s say for simplicity that each of my great-grandparents was born more than 100 years ago and had a 1% chance of living past 100 (some actual data is here), and that 15% of people are serious photographers. Then the conditional probability of my evidence E1 on the null hypothesis N is 1 − (1 − 0.01 ⋅ 0.15)8 or 1%, but on the doubling-chance hypothesis H it is 1 − (1 − 2 ⋅ 0.01 ⋅ 0.15)8 or 2.4%. Plugging these into Bayes’ theorem, I get a 67% posterior probability of H, and now I have a significant degree of credence in a photography-centenarianism link.
But suppose that instead of talking to my mother, I read further on in the article, and find after reading whatever scientific study spawned the article, the author found and interviewed a centenarian Bob who has been an avid photographer for half his life. What does that piece of anecdotal data do to my credence in the link between photography and a long life? Nothing! To a first approximation, the relevant fact I learned from the interview is the fact E2 that there exists at least one person in the world who was an avid photographer and lived past 100. And the conditional probability of E2 is very close to 1 on both H and N, so by Bayes’ theorem it doesn’t change my credences in H and N. (To a second approximation, I learned that the there was one person accessible to the author who was an avid photographer and lived past 100. And that is presumably slightly more likely on H and N. So I should get a slight boost in H, but only a slight one, since in the modern world we have access to lots of people.)
Now consider an intermediate case. Instead of talking to relatives, I share the article with a hundred people, and one of them writes back: “Wow! My tennis partner’s great-uncle Carl was an avid photographer and lived past 100.” Let’s over simplify by supposing that each of my hundred correspondents read the article and on average contributes ten people born more than 100 years ago to the sample. So from the first response, I have learned the fact E3 that in this sample of 500, there is at least one person who is a centenarian and a serious photographer. The probability of this on H is 95% and on N it is 78%. Plugging these into Bayes’ theorem, my credence in the photography-centenarianism link is 55%, which is a rather modest boost over my initial 50%. (The crucial point was that in the initial grandparents sample, the probability on H was double than on N, but now as both probabilities approach 100%, the ratio gets less impressive.)
There are some lessons here: If we are careful with our reasoning, anecdotal data can actually be quite relevant. Moreover, while it’s presumably been ingrained in us since high school science classes that larger sample sizes are better, for certain kinds of anecdotal data, smaller sample sizes are better. This is because the relevant information given by certain kinds of anecdotal data is positive: it is information that some sample contains at least one instance of some sort that is rare on both the null hypothesis and the alternate hypothesis (say, a photographer centenarian). In those cases, once the sample size gets large enough, the probability of the evidence on either hypothesis gets close to 1, and the evidential force disappears.
What this means is that for certain kinds of anecdotal data it makes perfect sense to be more impressed by an anecdote about oneself (a sample of one) than by an anecdote about a relative, and by an anecdote about a relative than by an anecdote about a friend’s friend, and to be essentially unmoved by an anecdote about a stranger on the Internet. And that is, I suspect, how most people actually proceed, notwithstanding blanket condemnations of anecdotal reasoning.
How can we do even better? Well, we should try to enrich our positive anecdotal data with other kinds of anecdotal data: Did I have centenarian relatives who weren’t photographers or photographer relatives who weren’t centenarians? All that would ideally be taken into account. But, nonetheless, if all I have is one positive anecdote, Bayesianism requires me not to dismiss it.
5 comments:
Question: "How can we do even better?"
Answer: Pearson's chi-squared test
Make a 2×2 Contingency table with the two variables of age at death (eiter at least 100 years or under 100 years) and proclivity of doing photography (actually doing photography in some form, as a hobby or as a profession, or not doing photography at all) out of an empirically gained data with sufficient amount of sample size.
Then do the Pearson's chi-squared test. I would first check, if there is a statistical dependency between those two variables of age at death and proclivity of doing photography at all and test if one can reject the null hypothesis H0 of independency on 95% significance level. If so, then a further analysis might be interesting.
Certainly and hypothetically the first thing, which I would do given that hypothetical click-bait article that says: “One thing you can do that science says doubles your chance of living past a hundred.”, would be checking the references and provided links in that article.
Then I would further check, if the proper science has been done in order for that claim in that article being properly substantiated or if the proper science has not been done.
If not, then I would myself do a proper research - a proper empirical and scientific research with reasonable data with a reasonable sample size.
Do you, Alexander R. Pruss, possess any degree or trainee in science?
This blogpost of yours would rather suggest otherwise.
I mean, that given this blogpost of yours is boosting my credence of you not possessing any proper degree or trainee in science.
Of course, one can do all sorts of sophisticated statistical analysis (though I have to say that I am suspicious of non-Bayesian statistics) if one has serious non-anecdotal data, if one has access to the original science, etc. But my question towards the end was meant to be whether one can do better *given merely anecdotal data*.
One cool thing about Bayesian analysis is that you can actually get something out of a single data point. :-)
You, Alexander R. Pruss, are suspicious of any non-Bayesian statistics?
hahahahhahahahahahahahahahahahhahahah...
I'm also very suspicious of the usage of Bayesian statistics without any considerations of the occurrences of false positives and false negatives, which is quite often done by philosophers or in philosophy, but not so much done in proper statistical analysis. Besides that it also appears to me being quite suspicious, how any prior probability/credence are basically arbitrary - setted to a specific probability without much of a proper justification.
So then how much value has that Bayesian analysis from a single data point and how much is that worth actually, if that Bayesian analysis lacks so much proper justifications? And how much worth is that, if we can do proper statistical analysis - just a little bit more work than guessing and assuming things out of thin air?
I guess then in that case, that Bayesian analysis from a single data point is not that much worth.
Some worries:
Your estimated probability of at least 1 centenarian photographer among your eight grandparents is pretty low, only 2.4%, even on hypothesis H. This suggests other possibilities: maybe your family is somehow atypical, or maybe your estimates are off. This is sort of a catch-22: it’s only with low probabilities that you get good discrimination, but low probabilities raise doubts. If your group had been a random sample maybe you could dismiss such doubts, but with anecdata that’s not so easy.
Anecdata can be tricky. You didn’t explicitly ask your mother to consider only your eight great grandparents. Suppose your mother had volunteered that your great aunt Agatha had been a centenarian and a photographer. How many relatives would you have grouped Agatha with?
The hypotheses N and H not complementary. Your prior for either should be close to zero. The correlation between photography and living to 100 might be low, but it’s unlikely to be precisely zero. Similarly, photography might improve your chances of living to 100, but it’s unlikely that it precisely doubles them. You would need a good reason to limit your attention to these two hypotheses. I can’t see such a reason in the setup. I suggest that a Bayesian should have a prior distribution for the proportion of centenarian photographers. With the numbers in the post, this distribution should cover the range 0% (no centenarians were photographers) to 1% (all centenarians were photographers). I have no idea what shape it should be.
Ian:
These are all good points.
By the way, one way to handle the issue of N and H not being exhaustive is to suppose that one is lucky enough that one's priors allow one to divide up the hypotheses about correlations between centenarianism and photography into two chunks, N* and H*, such that conditioning on N*, there is no correlation (N* will include hypotheses where the correlation is negative, say due to the inhalation of development chemicals), and conditioning on H*, we get the doubling.
Post a Comment