The findings in this group are about evidence itself: the patterns that make data look like it shows something it does not. Some are statistical, some come from how samples are gathered, and some come from how science is published. Together they amount to a list of reasons to distrust a striking result before checking where it came from.
Statistical traps
The most important entry is Galton's Regression to the Mean: extreme measurements tend to be followed by more ordinary ones, so any intervention applied at an extreme will appear to work. Littlewood's Law observes that with enough events, a one-in-a-million coincidence happens to each of us about once a month. The Will Rogers Phenomenon shows that moving cases between groups can raise the average of both, which is how better diagnosis can make every stage of a disease look more survivable without anyone living longer.
Who gets counted
A second cluster concerns samples. Survivorship Bias takes its famous example from wartime aircraft: armour belonged where the returning planes had not been hit, since planes hit there did not return. Selection Bias generalises the point, and the Rashomon Effect adds that sincere witnesses to the same event can give irreconcilable accounts.
How science publishes itself
The third cluster concerns the literature. Publication Bias favours positive and novel results, so the published record overstates effects. The Experimenter Expectancy Effect shows researchers' hopes leaking into their procedures. The Woozle Effect describes claims gaining authority through repeated citation without anyone checking the original. The Decline Effect and the Replication Crisis record the result: effects that shrink or vanish when others try to reproduce them. Lysenkoism is the extreme case, in which Soviet doctrine displaced genetics inside the scientific institutions themselves.
Regularities in data
A smaller set records patterns that appear in data of many kinds. Benford's Law finds leading digits following a logarithmic distribution in many naturally occurring datasets, which makes it a tool for detecting fabricated figures. Zipf's Law, Bradford's Law of Scattering and Heaps's Law describe skewed distributions in language and literature. Lindy's Law holds that for ideas and technologies, remaining life expectancy grows with age.
Where it applies
These findings are the basic defences of anyone who reads research, news or business reports. Regression to the mean should be the first explanation considered for any improvement following an extreme result, whether a pupil's test score or a company's quarter. Survivorship bias applies to every study of successful firms, funds or people. Auditors and tax authorities use Benford's Law to flag suspicious accounts. The publication and replication entries tell a reader to prefer meta-analyses, pre-registered studies and replications over single striking papers.
Where it is argued over
The statistical entries are mathematics and not in dispute; the disputes concern scale. How much of the published literature is false, and in which fields, is argued over, and estimates vary widely with the field and the method. The decline effect's causes are mostly mundane, selective reporting and regression, though the name can suggest something more mysterious. Zipf's law holds as a shape rather than an exact formula, and random processes can reproduce it, which limits what it explains.
Why it is worth knowing
Every empirical claim in this ledger depends on evidence that could be distorted in the ways this group describes. Knowing these patterns is what makes the rest of the ledger usable: it explains why the entries report replication, sample and source, and why a finding with a single dramatic study behind it is filed differently from one that has been reproduced for a century.
The entries behind this essay: Evidence in Science & Epistemology. Every subject has its essays on the essays page.