Towards a Cumulative Science of Prosody in ASD
Objectives: To overcome these challenges, we need to identify the barriers to data-sharing across research groups and highlight the potential of multi-study analysis across corpora (i.e. what is known as mega-analysis in genetic studies).
Methods: We updated the pool of studies under consideration (now 51 studies). We contacted all corresponding authors to inquire about shareability of the data and where possible requested participant level data. We performed a Bayesian multilevel mega-analysis combining participant-level and population-level data (when the former was not available), explicitly modeling age, gender and language of the participants, as well as relations between acoustic features.
Results: 30% of corresponding authors answered and 10% provided at least some of the requested data (some answers still pending). The mega-analysis identified: i) strong effects of age, and genderbeyond what approximate group-level matching accounted for; ii) strong interdependence between acoustic features, which needs to be accounted for, and iii) larger effect sizes than the previous meta-analysis. Interdependence is partly due to the common and inappropriate use of mean and standard deviation to describe long-tailed distributions of (within-subject) acoustic features. In addition, we identify key concerns: confidentiality and ethical concerns; lack of consent from co-authors; temporal and economic costs of retrieving the data.
Conclusions: Mega-analyses allow for more nuanced and powerful analyses than meta-analyses. However, they require more sophisticated statistical techniques and access to more detailed data than currently shared either within scientific articles or directly by their authors. We outline recommendations to reduce concerns in sharing data, avoid statistical confounds between acoustic features and provide examples to run analogous Bayesian mega-analyses.