Statcheck

Statcheck is an R package designed to detect statistical errors in peer-reviewed psychology articles by searching papers for statistical results, redoing the calculations described in each paper, and comparing the two values to see if they match. It takes advantage of the fact that psychological research papers tend to report their results in accordance with the guidelines published by the American Psychological Association (APA). This leads to several disadvantages: it can only detect results reported completely and in exact accordance with the APA's guidelines, and it cannot detect statistics that are only included in tables in the paper. Another limitation is that Statcheck cannot deal with statistical corrections to test statistics, like Greenhouse–Geisser or Bonferroni corrections, which actually make tests more conservative. Some journals have begun piloting Statcheck as part of their peer review process. Statcheck is free software published under the GNU GPL v3. In 2017, Statcheck's developers published a preprint paper concluding that the program accurately identified statistical errors over 95% of the time. This validity study comprised more than 1,000 hand-checked tests among which 5.00% turned out to be inconsistent. The study found that Statcheck recognized 60% of all statistical tests. A reanalysis of these data found that if the program flagged a test as inconsistent, it was correct in 60.4% of cases. Reversely, if a test was truly inconsistent, Statcheck flagged it in an estimated 51.8% of cases (this estimate included the undetected tests and assumed that they had the same rate of inconsistencies as the detected tests). Overall, Statcheck's accuracy was 95.9%, half a percentage point higher than the chance level of 95.4% expected when all tests are simply taken at face value. Statcheck was conservatively biased (by about one standard deviation) against flagging tests. More recent research has used Statcheck on papers published in Canadian psychology journals, finding similar rates of statistical reporting errors as the original authors based on a 30-year sample of such articles.