The natural selection of bad science · labs as the unit of selection

Publish orperish

Labs investigate things and publish what they find. A lab that publishes more gets copied by the next generation. Nobody here fabricates data, nobody cuts a corner they know is a corner, and nobody is told there is a trade-off. Care collapses regardless.

The rules

Why quantity wins without anybody choosing it

Being careful buys you two things: a better chance of detecting an effect that is really there, and fewer reports of effects that are not. It costs you studies, because a careful study takes longer.

Add those up and the careless lab publishes several times as much. Most of what it publishes is false, but nothing in the system is counting that. What gets counted is publications, and what gets copied is whatever the lab with the most publications was doing.

Smaldino and McElreath (2016) called it the natural selection of bad science, and the point of the phrase is the mechanism: selection needs no intent. This page has a coda that belongs on a site like this one, and it is at the bottom.

Start from
Generation
0
not started
How careful
—
out of 100
Chance of finding a real effect
—
statistical power
Published findings that are false
—
of the whole record
The trade every lab faces
How careful the field is
Over time drag or arrow keys to replay
How careful the field is Published findings that are false
What the interventions actually do
Things worth trying

Four runs and a comparison

Press Nobody checks anybody
Care falls from about 42 to 1 within fifty generations, and the published record reaches 95% false. Watch the distribution slide left with nobody deciding anything.
Press A quarter checking
The record improves markedly. Care stays on the floor. Those are two different things and they are usually discussed as one.
Press Nulls publishable
Same shape. The record gets cleaner because honest negatives are now in it, and the incentive to run big careful studies has not changed at all.
Press Compare the interventions
Averaged over four seed families, deliberately. A single family reads about four times too kindly here, and the encouraging number nearly went into this page as a finding.
Drop hypotheses that are true to 2%
The record gets worse at every level of care. A field asking mostly wrong questions publishes mostly wrong answers however carefully it works, which is a point about what a field studies rather than how.
The argument

Why anyone cares

Every other page here models people choosing badly, or choosing well and getting a bad outcome. This one models an institution selecting for something nobody in it wants, and it is the only page whose subject is the activity that produced all the others.

What selection needs

The phrase doing the work is natural selection. It needs variation, heredity and differential reproduction, and nothing else. Labs vary in how careful they are. Practices are inherited, because students learn from supervisors and successful groups get imitated. Reproduction is differential, because publication drives hiring and funding.

That is the whole apparatus. No individual has to be dishonest, careless on purpose, or even aware that care and output trade against each other. Smaldino and McElreath are explicit that the dynamic requires no conscious strategising, and that is what makes it hard to argue with — the usual reply, that most scientists are honest and diligent, is granted from the start.

Nobody chose this. Publication counts got counted, and whatever produced them got copied.

The two interventions are not the same intervention

Replication and publishable null results are usually proposed together as the fix for the same problem. In this model they do the same thing as each other and it is not the thing they are credited with: both substantially clean the published record, taking false findings from over nine in ten down to around one in five, and neither gets care off the floor.

That distinction matters for what you would expect an intervention to achieve. A cleaner literature is worth having on its own terms. But if the concern is that the field is producing underpowered work, then cleaning up afterwards leaves the production untouched, and the model suggests the amount of checking that would change the production is more than any real field does.

The coda

This page belongs on this site for a reason beyond its subject. Kohrt and colleagues replicated Smaldino and McElreath's model in 2023 and found a coding error in the original: the implemented model deviated from the one described in the paper, and selection on replication frequency produced a burst of activity the paper never reported. A formal correction was issued. The conclusions were unaffected.

So a paper about how bad science gets selected for turned out to need replicating, was replicated, had an error found, and survived it. Whatever else it demonstrates, it demonstrates its own argument for replication rather well.

Worth adding in the same spirit: while building this page I read a single run as showing that heavy replication restores care to about 13 out of 100. Across four seed families the figure is under 5. The encouraging number was an artefact of one seed and it was in a draft of this page before it was checked. The comparison above averages over families for that reason.

What this model is not

One kind of study, one measure of success, and no theory. Labs here cannot specialise, collaborate, train anybody, or pick easier questions, and there are no journals with different standards, no peer review, no grants and no careers. Nothing here distinguishes a field where replication is cheap from one where it is impossible. The claim is narrow: incentives of this shape are sufficient to degrade methods, so observing degraded methods does not require anybody to have behaved badly.

References

Sources

  1. Heesen, R. (2018). Why the reward structure of science makes reproducibility problems inevitable. The Journal of Philosophy, 115(12), 661–674.
  2. Ioannidis, J. P. A. (2005). Why most published research findings are false. PLoS Medicine, 2(8), e124.
  3. Kohrt, F., Smaldino, P. E., McElreath, R., & Schönbrodt, F. (2023). Replication of the natural selection of bad science. Royal Society Open Science, 10(2), 221306.
  4. Nosek, B. A., Spies, J. R., & Motyl, M. (2012). Scientific utopia II: Restructuring incentives and practices to promote truth over publishability. Perspectives on Psychological Science, 7(6), 615–631.
  5. Romero, F. (2017). Novelty versus replicability: Virtues and vices in the reward system of science. Philosophy of Science, 84(5), 1031–1043.
  6. Smaldino, P. E., & McElreath, R. (2016). The natural selection of bad science. Royal Society Open Science, 3(9), 160384.
  7. Smaldino, P. E., & McElreath, R. (2023). Correction to: The natural selection of bad science. Royal Society Open Science, 10(9), 231026.
  8. Zollman, K. J. S. (2018). The credit economy and the economic rationality of science. The Journal of Philosophy, 115(1), 5–33.