Reproducibility and Replication

The two words name different tests. Reproducibility, in the sense methodologists now use, is computational: given the original data and code, an independent analyst recovers the published numbers, and failure indicates an error or a description too vague to follow. Replication is a fresh study, new participants or samples, asking the same question, and failure there indicates the finding may not exist. A result can be perfectly reproducible and entirely unreplicable, since reproducing the analysis of a false positive returns the same false positive.

Two bodies of work fixed the scale. The Open Science Collaboration repeated 100 psychology experiments in 2015 and found significant results in roughly a third of replications against 97 percent of originals, with replication effect sizes about half the published ones. In preclinical biology, Begley and Ellis reported in 2012 that an Amgen team confirmed the findings of only 6 of 53 landmark cancer papers, and a Bayer group had earlier reproduced well under half of the target-validation studies it tried.

The evidential unit is therefore the replicated finding rather than the study, particularly where the original was small or the result surprising enough to be interesting. Independent replication also probes what a repeat in the same laboratory cannot: different reagent lots, animal suppliers, operators and assay calibrations.

The failure mode in peptide marketing is the immortal citation. A single in-vitro or rodent result from one laboratory, published decades ago and never independently repeated, is cited endlessly as though repeating the citation repeated the experiment. A related sleight calls a result reproducible on the strength of technical replicates, three wells in the same hands on the same day, which measures assay noise rather than generalisability.