In simple terms
A finding replicates when a different team asks the same question, gathers its own data, and reaches a consistent answer. The replication crisis is the concern that this happens less often than scientists assumed.
It is an argument about how much confidence a single published result deserves. It is not a claim that science as a whole is broken.
How it works
Two similar words carry different meanings, and the difference matters.
Reproducibility means getting the same answer from the original data and analysis steps. Replicability means getting a consistent answer from a fresh study that collects new data. The US National Academies drew that line in a 2019 report.
The best known test came in 2015, when a large collaboration repeated 100 psychology studies. Ninety seven percent of the original results had been statistically significant. Only 36 percent of the repeats were, and the effects they did find were on average about half the size.
Rates differ sharply between fields. Across the disciplines the National Academies examined, replication rates ran from fewer than one study in five to more than three in four.
Why it matters
Not every failed repeat is a scandal. The report separates helpful sources of non-replication, such as real variability in complex systems and reasonable differences in method, from unhelpful ones, such as weak design, thin reporting, poor statistical training and, rarely, misconduct.
Some non-replication is therefore normal and even useful. The National Academies declined to call the situation a crisis, and warned that the overall non-replication rate is a poor gauge of the health of science.
What has changed is practice. Journals and funders now push for shared data, published code, larger samples and analysis plans registered before the data arrive.
Where you’ll see it
- News reports that a famous psychology, nutrition or medicine finding did not hold up.
- Journal rules requiring authors to post data and code.
- Funding agency conditions on rigor, transparency and sample size.
- Textbooks quietly dropping effects that later work could not confirm.
Example
A striking result from one small study makes headlines. Three larger repeats find a much weaker effect, so the original is treated as provisional rather than established.
Often confused with
Fraud. Most non-replication comes from small samples, flexible analysis and selective publication, not from anyone faking data.
Key facts
- The US National Academies define reproducibility as consistent results from the same input data, code and analysis, and replicability as consistent results from studies that each gather their own data.1
- Across the disciplines its 2019 report examined, replication rates ranged from fewer than one study in five to more than three in four.1
- The report separates helpful sources of non-replicability, such as inherent uncertainty and legitimate methodological choices, from unhelpful ones, such as poor design, inadequate reporting, weak training and misconduct.1
- The report avoids declaring a crisis and holds that the overall extent of non-replication is an inadequate indicator of the health of science.1
- In a 2015 replication of 100 psychology studies, 97 percent of the original findings were statistically significant but only 36 percent of the replications were.2
- Effects measured in those replications were on average about half the magnitude of the originals.2
Quick checkWhat is the difference between reproducibility and replicability?Show answer
Reproducibility means the original data and analysis give the same answer again. Replicability means a new study with new data reaches a consistent answer.
Sources
- National Academies of Sciences, Engineering, and Medicine. Reproducibility and Replicability in Science: Summary (NCBI Bookshelf). 2019 (accessed 15 September 2026)
- Open Science Collaboration, Science (PubMed record). Estimating the reproducibility of psychological science. 2015 (accessed 15 September 2026)
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed September 16, 2026Researched and drafted with AI assistanceReport an issue