AIM: To assess the prevalence of publication bias in the radiology literature, data-mining techniques were used to extract p-values in abstracts published in key radiology journals over the past 20 years.
MATERIALS AND METHODS: A total of 34,699 abstracts published in Radiology, Investigative Radiology, European Radiology, American Journal of Roentgenology, and American Journal of Neuroradiology published between January 2000 and December 2019 were included in the analysis. Automated text mining using regular expressions was used to mine abstracts for p-values.
RESULTS: The text mining algorithm detected 43,489 p-values, the majority (82.4%) of which were reported as “significant”, i.e., p<0.05. There has also been an increased propensity to report more p-values over time. The distribution of p-values showed a step change at the conventional significance threshold of 0.05. The odds ratio of a “significant” p-value being reported in the abstract compared to the full text was calculated to be 2.52 (95% confidence interval 1.78–3.58; p<0.001). Taken together, these results provide strong evidence for selective reporting of significant p-values in abstracts.
CONCLUSION: Statistically significant p-values are preferentially reported in radiology journal abstracts.
- Radiology literature
- Data mining