Advanced brain imaging is not enough for studies of brain development to be useful to understanding developmental psychopathology. RBC responds to a common obstacle in aggregating data across studies – different questionnaires can only be compared if harmonized. To this end, we harmonized instruments across studies to derive comparable dimensions of psychopathology.

Measures in the RBC

The RBC integrates baseline psychiatric phenotyping data from six important developmental studies. We focused on one of the most widely-used assessments the Child and Behavioral Checklist (CBCL). The CBCL is a 120-item parent-report assessment of emotional and behavioral symptoms over the past 6 months, answered on a 3-point Likert-type scale. However – as it is common in the field – different measures were used across these studies. Four studies (HBN, NKI-RS, BHRCS and CCNP) used the latest version of the CBCL, while PNC used the GOASSESS – a highly structured psychiatric interview.

Bifactor models of psychopathology

We tested 12 models of psychopathology (Hoffman et al., 2022, Hoffmann et al., 2023), focusing on bifactor models to disentangle general and specific dimensions of psychopathology. In a bifactor model, a general factor captures overall trans-diagnostic psychopathology (the “p” factor), while specific factors model distinct dimensions of psychopathology. One advantage of this framework is that the general and specific factors are uncorrelated.

Harmonization Procedure

We conducted a sequence of analyses to generate scores for a bifactor model that was harmonized across datasets. First, we tested the impact of various bifactor models to see which one worked best in terms of fit, internal consistency, invariance and convergent validity (Hoffman et al., 2022). Next, we examined different item-matching strategies to harmonize CBCL and GOASSESS and selected the expert-based 1-to-1 item matching as the best strategy (Hoffman et al., 2024). Finally, we used the results from the first two steps to select a model that harmonizes our dataset well (Hoffmann et al., 2023).

Main RBC Model

We found that the model, described by McElroy et al (2018), was the best option. This model fit the data very well. More importantly, it also displayed measurement invariance across questionnaires and retained the best proportion of the original ones. This model included 22 harmonized items, which contribute to a model with one general factor as well as internalizing, externalizing and attention/hyperactivity specific factors.