The scientific method is the closest to perfection among implemented plans for mining new truth devised to date by the ingenuity of humans. All of the proposed better plans that I know of basically involve ways of running the scientific method faster (e.g. strong AI).
The utility of the scientific method is that it works even when run by self-interested, flawed, irrational humans. Technology moves ahead. People live rather than die, and become far wealthier than their ancestors. That this process kicked into high gear sometime around the time and place that the scientific method was finally formalized and that formalism successfully popularized, after thousands of years of ragged, slow, and erratic progress, does not seem to me to be a coincidence.
This is just an anecdote, but I consistently see the same effect when doing A/B testing. The initial hypothesis always does fantastic for the first week or so. But if I let it run long enough, it always drops to zero. Been scratching my head about it for a few years.
I think the underlying cause is similar to the publication bias they mention in the article.
For almost all A/B tests, A is actually not much different from B. But due to small sample sizes you will see after the first week, at random, a rather strong positive or negative result. And now the publication bias/selection bias kicks in. If you see strong negative results in the first week you will quickly give up and start a new A/B test. If the initial results show positive results you get excited and keep testing but then in most circumstances you get reversal to the mean at high sample sizes. This would most likely also have happened to the experiments you terminated early but you selected them away and in your memories only for the good initial results a reversal to the mean often seem to happen.
Most A/B tests, ran for a long enough time, will show insignificantly differences. Blogs may give a different impression but again explained by publication bias.
I always compare A/B testing to genetic mutations. Almost none have a strong impact on the fitness of an animal but once in a very long while you have a positive one. Luckily they accumulate and you can get some impressive results with A/B testing (aka natural selection)
I think the consensus on why this happens boils down to regression to the mean, selective experimentation (exciting things over boring things like attempting to duplicate results), and selective publishing (e.g. positive results are much more interesting to journals than hypotheses being rejected).
Plus if you have a million researches in your area of research a 1 per million chance your results are wrong is still no assurance. Should statistically happen to one of the researchers in your area.
I'd add bad math. Fisher significance testing is on shaky grounds[0] both theoretically (significance testing violates the likelihood principle) and practically (just by nature of having a small sample size, a lot of null hypotheses can be rejected, while with a larger sample size hardly none are rejected). Hypothesis testing is little better. The "Bayesian revolution" has barely begun and has yet to influence a majority of researchers.
I don't think this is quite right. If the null hypothesis is true and you conduct the same study an infinite number of times, you will reject the null hypothesis 5% of the time at alpha = 0.05 regardless of sample size, unless there is something wrong with your sampling procedure or hypothesis test. This is the point of null hypothesis significance testing.
In practice, if you fix p=0.05 and increase your n, the probability that you will find a statistically significant result often increases because your power increases, and in many situations, the probability that the null hypothesis is true is close to zero. (Andrew Gelman uses the example of asking whether there are significant differences between voting patterns of men and women.)
On the other hand, effect size estimates become more accurate as the sample size increases. This mitigates the above issue, provided you actually report your effect size. It also means that small sample studies that report statistically significant results are more likely to overestimate their effect size, which is especially problematic if you are applying null hypothesis significance testing when you know the null hypothesis is false.
For those that faintly recognize the author's name... this piece was written by Jonah Lehrer, whose writings were later discredited for a combination of plagiarism, inaccuracy, and methodological weakness. Interestingly, in the ensuing reassessment of Lehrer's writings, this article has survived basically unscathed, but I think this discussion (including some statements from the interviewed scientists) is very useful:
One of the nice quotes from that link by Rich Palmer, who was one of the authors of the scientific work referenced:
If there’s a lesson here, it’s about a widespread human failing. Most people would rather some other clever person distill down all the complex details into a good story for them, preferably in excellent prose. But those distilled stories should never be treated as a substitute for original research results. If anyone really wants ‘the truth’, they’re going to have to slog through an awful lot of turgid and arcane original research and draw their own conclusion.
Wouldn't it be great if educational institutions allowed "un-theses" to be accepted for degree requirements? That is, if you rigorously invalidated an existing theory. The university equivalent of the "Black Team".
Why do you think this wouldn't be accepted? I know in my field of biology if a person were to disprove an existing theory it would be an important work. My colleague was just scooped in Science on work that disproves the ribosome spacing hypothesis when it comes to translation initiation and codon optimization.
Isn't this the same sort of thing that was discussed in the recent A/B testing articles? A test that demonstrates amazing results is going to be more visible than the ones that demonstrate weak results. It's kind of like how if you stop your A/B test when you see the strongest evidence, you're missing the point.
The real problem with the FA is the ending, which strongly agrees with people epitomized in fiction by the likes of Sheldon Cooper's mother in her reesponse to her son's statement of a scientific law as a fact: "and THAT is YOUR opinion."
I think this article is interesting in that it mostly talks about experiments that are aiming to capture hard correlations in incredibly complex systems. The scientific method is an amazing tool that we've been using over the last 350 years to pretend that the world is a machine and thus glen incredibly insightful information. In the end though it probably isn't a machine and as systems become more and more complex they can become even harder to machinify. (Christopher Alexander has a lot more to say on the issue: http://www.katarxis3.com/Alexander.htm)
19 comments
[ 3.3 ms ] story [ 58.4 ms ] threadThe utility of the scientific method is that it works even when run by self-interested, flawed, irrational humans. Technology moves ahead. People live rather than die, and become far wealthier than their ancestors. That this process kicked into high gear sometime around the time and place that the scientific method was finally formalized and that formalism successfully popularized, after thousands of years of ragged, slow, and erratic progress, does not seem to me to be a coincidence.
For almost all A/B tests, A is actually not much different from B. But due to small sample sizes you will see after the first week, at random, a rather strong positive or negative result. And now the publication bias/selection bias kicks in. If you see strong negative results in the first week you will quickly give up and start a new A/B test. If the initial results show positive results you get excited and keep testing but then in most circumstances you get reversal to the mean at high sample sizes. This would most likely also have happened to the experiments you terminated early but you selected them away and in your memories only for the good initial results a reversal to the mean often seem to happen.
Most A/B tests, ran for a long enough time, will show insignificantly differences. Blogs may give a different impression but again explained by publication bias.
I always compare A/B testing to genetic mutations. Almost none have a strong impact on the fitness of an animal but once in a very long while you have a positive one. Luckily they accumulate and you can get some impressive results with A/B testing (aka natural selection)
Did I miss any reasons?
Very good article by Warren Buffet that touches on the same issue: http://www.tilsonfunds.com/superinvestors.html
[0] http://uncertainty.stat.cmu.edu/ Chapter 12.
In practice, if you fix p=0.05 and increase your n, the probability that you will find a statistically significant result often increases because your power increases, and in many situations, the probability that the null hypothesis is true is close to zero. (Andrew Gelman uses the example of asking whether there are significant differences between voting patterns of men and women.)
On the other hand, effect size estimates become more accurate as the sample size increases. This mitigates the above issue, provided you actually report your effect size. It also means that small sample studies that report statistically significant results are more likely to overestimate their effect size, which is especially problematic if you are applying null hypothesis significance testing when you know the null hypothesis is false.
http://www.lastwordonnothing.com/2012/11/05/jonah-lehrer-nat...
If there’s a lesson here, it’s about a widespread human failing. Most people would rather some other clever person distill down all the complex details into a good story for them, preferably in excellent prose. But those distilled stories should never be treated as a substitute for original research results. If anyone really wants ‘the truth’, they’re going to have to slog through an awful lot of turgid and arcane original research and draw their own conclusion.
God says... partake incorruptible sanctified endued strives entrance weighing demonstrate