To create the latent variable for cognitive ability, allowing us to separately examine social skills and musical sophistication, data from the D-KEFS, Trails, WAIS-IV, CVLT-3, and WMS-IV were in the following analyses. Table 2 displays the PCA, EFA, and CFA loadings.

The first four principal components (PCs) accounted for 88.8% of the variance, indicating that all predictor variables made a meaningful contribution to the PCs (Verma et al. 2019). Based on Kaiser’s (1960) criterion, which retains PCs with eigenvalues greater than 1, we assessed the first two PCs. PC1 was associated with executive function and intelligence. PC2 represented executive function and memory. All assessments were retained following PCA.

Parallel analysis determined that a single factor should be retained, with factor loadings that exceeding 0.3, suggesting an association with the factor (Tavakol and Wetzel 2020). As a result, the Trails data was excluded from the final model. The fit indices indicated a very good fit to the cognitive ability factor (χ² = 1.64, p < 0.90, RMSR = 0.04, RMSEA = 0, 90% CI [0, 0.068], TLI = 1.328, BIC = −20.33). A final EFA with all assessments confirmed that the Gold-MSI and SSI contributed to separate factors.

Given the conflicting guidance on whether variables with loadings < 0.3 or < 0.4 should be retained (Ondé and Alvarado 2020), the factor loading cut-off was set at 0.35. The CFA confirmed that the WAIS-IV, DKEFS, and CVLT-3 should be retained in the final model, therefore the WMS-IV was removed.

A predefined model was developed to examine how musical sophistication, social skills, and cognitive ability influence BrainAGE, as shown in Fig. 1. Gold-MSI and SSI were treated as independent variables to assess their direct effects on BrainAGEs. Cognitive ability was represented as a latent variable, incorporating data from the D-KEFS, CVLT-3, and WAIS-IV, to evaluate its impact on BrainAGE scores. Residual covariances were added between SSI, cognitive ability, and Gold-MSI to explore shared variance among predictors. Based on prior literature and modification indices, a residual covariance between D-KEFS and BrainAGE was included to account for shared variance likely driven by overlapping neural regions not fully captured by the latent cognitive ability factor (Haas et al. 2022). The final model showed a good fit to the data (χ²(6) = 7.189, p = 0.304, RMSEA = 0.049 (90% CI [0.000, 0.159], p = 0.428 for RMSEA ≤ 0.050), CFI = 0.959, TLI = 0.898, and SRMR = 0.061).

Final structural equation model (SEM)

SEM results, including factor loadings, estimates, and significance values, can be seen in Table 3. The results demonstrate that D-KEFS was a strong contributor to the latent construct, while WAIS-IV and CVLT-3 were moderate and weak contributors, respectively. There were no significant predictors for BrainAGE, demonstrating the predictors in the model did not explain a substantial amount of the variability in BrainAGE scores. Further, there were no significant interactions. The D-KEFS explained most of the residual variance in the model, however, only 8% of the variance in the BrainAGE scores was explained by the predictors, leaving 92% unexplained. We conducted exploratory models using Gold-MSI and SSI subscales, as well as individual cognitive assessments, to test whether summary scores may have obscured potential effects. However, no subscales or individual assessments emerged as significant predictors.