[dr_dshiv]

Seriously, what about that kind of publication bias: A/B tests don’t get published.

If you run a useful system where it would be meaningful and interesting to know whether a social science theory actually applied, you might run an A/B test to see if it works. If it works, it is adopted—but it is almost never published. And that is for two reasons: 1. no incentive to publish and 2. major incentive not to publish. #2 is recent (post Facebook experiment) and it is specifically because a large portion of the educated public accepts invisible A/B testing but recoils with moral indignation at the use of A/B testing results in published science. Too bad: Facebook keeps testing social science theories, but no longer publishes the results.

[MereInterest]

The standards of selecting a result of an A/B test are less stringent than those of publication for the advancement of knowledge. For publication, the goal is to determine whether a model is accurate. For A/B testing, the goal is to select the best design/intervention. The difference is that for scientific testing "inconclusive" means that there isn't enough evidence to consider it a solved problem and it should have more research, while in A/B testing "inconclusive" means that any effect is small so you should pick an option and move on.

As an example, suppose I flip a coin 1000 times and get heads 525 times. The 95% confidence interval for the probability of heads is [0.494, 0.556], so from a scientific standpoint I cannot conclude that the coin is biased. If, however, I am performing an A/B test, I would conclude that I'll bet on heads, because it is at worst equivalent to tails.

[dr_dshiv]

I think you are missing the point. With academic publication bias, sometimes an unbiased coin gets heads 600 times by chance. Those studies get published. But, if you ran the test again, you might only get 525. That study won’t get published.

And, in opposition to your assumption: there is nothing to prevent A/B tests being published with high academic standards— like a low p value and tons of n. In an academic context, that’s just fine— it’s a small but significant effect.

A/B tests are simply controlled experiments—which are the gold standard of scientific evidence generation in psychology. My point is that the main generators of this evidence are only permitted to use this evidence to inform commerce not public knowledge. That is a loss for science and public policy, in my opinion.