You Can't Replicate What You Can't See: A Call for Researchers to Share Their Data and Avoid "Fragile" Correlations.

We suggest that a measure of a correlation's fragility is the minimum number of items that, when replaced with the group median, result in a nonsignificant correlation on reanalysis. Between January 2000 and July 2021, there were 1769 significant correlations reported in 142 papers published in this journal, and only 51 correlations (2.9%) had available data (scatter plots from which we could digitize the raw data). Twenty-six of these 51 correlations were fragile at 4 or fewer replacements. Five of the reported significant correlations were not significant when we replicated the analysis from the extracted data. .