Only Half of Social Science Findings Held Up in a Giant Retest

Key Points

  • A Nature paper published April 1, 2026 retested 274 positive claims from 164 social and behavioural science papers published from 2009 to 2018.
  • Weighted by paper, 49.3 percent replicated with a significant result in the original pattern; counted claim by claim, 55.1 percent did.
  • The median effect size fell from a correlation of 0.25 to 0.10, and 80.3 percent of papers showed a smaller effect on replication.
  • Rates ran from 42.5 percent in economics to 63.1 percent in education, and the differences between disciplines were not statistically significant.

Published .

A red brick building with a white dome and a white columned portico stands at the far end of a wide green lawn, framed by bare trees and long white colonnades on both sides, with a few people near its steps.
The Rotunda and Lawn at the University of Virginia in Charlottesville, Virginia, photographed in March 2011. Photo: Phil Roeder, CC BY 2.0, via Wikimedia Commons, resized, photo source.

Take 164 published social science papers, test their key claims again with independent evidence, and roughly half come back the same way. That is the central result of "Investigating the replicability of the social and behavioural sciences," published in Nature on April 1, 2026, with Andrew H. Tyner first in the author list and Brian A. Nosek and Timothy M. Errington last1.

What the team tested

The replications were part of SCORE, a program funded by the US Defense Advanced Research Projects Agency (DARPA)2. Tyner works at the Center for Open Science, Nosek is the center's executive director and a professor at the University of Virginia, and Nosek told PsyPost the program began in 20193.

The targets were 274 claims of positive results from 164 quantitative papers published from 2009 to 2018 in 54 journals1. Those journals were grouped into six disciplines (business, economics, education, political science, psychology and sociology)2. In all, 296 replications were run. Some claims were tested more than once, and combining those left 274 replications of unique claims2.

Teams used original materials when they could get them, had their designs peer reviewed in advance (an author of the original study was invited to review), and preregistered their plans before running them2. Some collected new data, and others used existing data that the original research had not used2. The median statistical power to detect the original effect size was 99.6 percent1.

What held up

The headline figure depends on how the claims are counted. Counted one by one, 151 of 274 claims (55.1 percent) produced a statistically significant result in the same pattern as the original. Weighted so that each paper counts once, the figure was 80.8 of 164 papers, or 49.3 percent1. Another 40.4 percent of papers showed no significant effect, and a smaller share produced a significant result in the opposite direction2.

Effects shrank even when they survived. For claims whose effects could be converted to a correlation, the median Pearson's r was 0.25 in the original studies and 0.10 in the replications, a 58.1 percent drop2. On replication, 126 of 157 papers (80.3 percent) and 175 of 249 claims (70.3 percent) showed a smaller effect than the original2.

Statistical significance is only one yardstick. The team applied 13 methods for judging replication success, and their estimates ran from 28.6 percent to 74.8 percent, with a median of 49.3 percent1.

Field by field

Weighted by paper, replication rates were 42.5 percent in economics, 47.2 percent in business, 49.0 percent in psychology, 51.1 percent in sociology, 52.0 percent in political science and 63.1 percent in education3. The differences between disciplines were not statistically significant (a chi-square p-value of 0.76), and publication year showed no significant effect either2.

The top of that range rests on few papers. Education contributed 13 papers to the tested set and political science 15, against 58 for psychology3, and the authors note that some discipline estimates had high uncertainty because of small sample sizes2. Replications that collected new data and those that used existing data succeeded at similar rates, 47.8 and 51.4 percent of papers2.

Why a failed replication is not a verdict

Some decline was expected before any data came in. The authors write that some drop in effect size and significance follows from the power to detect original effects and from regression to the mean, since only positive results were picked for replication1. In a check that took every outcome from 200 papers, significant or not, an estimated 89.6 percent of the claims had significant results2, which the paper reads as the published record's tilt toward significant findings.

The paper also says one failed replication is not enough to conclude that the original research was wrong, and a single success does not prove it right2. Nosek made the same point to PsyPost: "Failing to replicate a finding does not mean that the original finding is wrong." He also said that "establishing the replicability of published results cannot be taken for granted."3

The sample has limits of its own. It left out qualitative research and papers claiming null results, and it drew on relatively prominent journals, which could push replicability higher or lower than in the wider literature2. Errington, the center's senior director of research, summed it up in its April 1, 2026 release: "The main message of SCORE is a simple one: research is hard."4

Trivia

  1. What does the acronym SCORE stand for? Systematizing Confidence in Open Research and Evidence2.
  2. How many researchers took part in the SCORE program, according to the Center for Open Science? 8654.
  3. How large was SCORE's starting sample of papers, and from how many journals? 3,900 papers from 62 journals2.
  4. Which of the 13 success measures gave the highest replication estimate? The meta-analytic combination of the original and replication evidence2.
  5. Weighted by paper, what share of replications found a significant effect in the opposite direction? 9.7 percent2.
  6. What were the median sample sizes of the original and replication studies, weighted by paper? 206 for the originals and 545 for the replications3.
  7. In SCORE's separate reproducibility check, what share of 600 assessed papers had data available? 24 percent4.

Sources

  1. Andrew H. Tyner, Brian A. Nosek, Timothy M. Errington and others, Investigating the replicability of the social and behavioural sciences, Nature, vol. 652, issue 8108, pp. 143 to 150, published online April 1, 2026 (issue dated April 2, 2026). Accessed November 14, 2026.
  2. Andrew H. Tyner and others, Investigating the replicability of the social and behavioral sciences, accepted manuscript for Nature, University of Essex Research Repository. Accessed November 14, 2026.
  3. Eric W. Dolan, Only about half of social science findings held up when retested, major replication project finds, PsyPost, October 29, 2026. Accessed November 14, 2026.
  4. Center for Open Science, Large-Scale Collaboration Releases New Findings on Research Credibility, April 1, 2026. Accessed November 14, 2026.