In the Research Article, “Cross-ethnicity/race generalization failure of behavioral prediction from resting-state functional connectivity,” by Li et al., some numbers regarding the behavioral measures showing significant accuracy difference between African Americans (AA) and white Americans (WA) were rerun and have been slightly changed by +1/-1. A summary of the changes to the article and supplementary materials following a correction of this error are noted below.
• In the Results section under “ABCD dataset,” the number of behavioral variables out of 36 that showed predictability and achieved >0.15 correlation accuracy across the test set changed from nine to ten.
• In the “Full-dataset model (linear ridge regression) yielded higher prediction error in AA than in WA” subsection, “seven” behavioral measures was changed to “six” and “Four of these seven measures” was changed to “Three of these six measures.” In the same paragraph, “crystalized cognition” was removed.
• Under the “Training the model specifically on AA increases its performance for this population” subsection, “19” of the 36 behavioral measures was changed to “18”; “No significant difference was observed for nine behavioral measures” has been updated to “No significant difference was observed for ten behavioral measures”; “26” behavioral measures observed in higher predictive COD in WA than AA was updated to “25”; WA-trained model only performed better in AA than WA for “two” behavioral measures was changed to “one”; and Differences in predictive COD between AA and WA were not significant for “eight” behavioral measures has been updated to “ten.”
• Fig. 3 has been corrected to show the updated sets of behavioral measures in panels B and D.
• In Table 1, in the first row, “Overall cognition” was removed from the first column; “Long delay recall” was removed from the second column and “BAS—drive” was added in the same column; “BAS—drive” was removed from the third column, and “Overall cognition” and “Long delay recall” were added in the third column. In the second row, “BAS —reward responsiveness”, “Visuospatial accuracy”, and “BAS —fun seeking” were removed from the first column, and “Somatic complaints” and “Mania” were added in the same column; “Anxious/depressed” was removed from the second column; “Somatic complaints” and “Mania” were removed from the third column, and “Visuospatial accuracy”, “BAS —fun seeking”, BAS —reward responsiveness” and “Anxious/depressed” were added in the third column.
• Fig. 5 has been updated to show the changed number of behavioral measures with significant accuracy difference between AA and WA in panel A and correlations between X- and Y-axes have changed in panels B and C.
• The numbers in each scatter plot of Fig. 6 have been updated to reflect the changed demeaning method. For example, in panel A, COD of AA changes from “-0.086” to “-0.085,” and correlation accuracy of WA changes from “0.048” to “0.049.”
• In the Supplementary Materials PDF, the order of a few behavioral measures has been updated in Figure S3, S4, S6, S10.
o In Figure S6, the number of behavioral measures with significant higher accuracy in WA than AA has been updated. In panels A and B of Figure S6, “Openness (NEO)” did not show significant difference in the original version but show significant difference in the corrected version; in (B), “Visuospatial accuracy” showed significant difference in the original version, but not in the corrected version.
o In Figure S10, the number of behavioral measures with significantly higher accuracy in AA than WA in panel B changes from 14 to 15: “Cognitive control, Attention (Flanker)” did not show significant AA-WA accuracy difference in the original version but show significant difference in the corrected version.
o In Figure S11, the number of behavioral measures with significant accuracy difference between AA and WA have been updated. For example, when training only on AA, 11 behavioral measures showed higher accuracy in test WA than test AA in the original version. In the corrected version, only nine behavioral measures showed higher accuracy in WA than AA. In panels B and C, correlations between X- and Y-axes are slightly changed.
The errors do not impact the conclusions of the paper. The Results section, Figs. 3, 5, 6, Table 1, and the supplementary materials have been updated in the HTML and online PDF versions to reflect these changes.
