Refining MLU-like metrics in parent-child speech: Measurement properties and decision aids for language-inclusive measures of speech complexity

Mean Length Utterance (MLU) has two problems: High analytic flexibility, and a lack of theoretical justification for data handling choices (e.g., selection of metrics; approach to multilingual speech). We evaluate multiple MLU-like metrics in a large corpus of child-directed speech from multilingual Singapore. Using an existing dataset of wordless picture-book sharing between parent-child dyads with frequent translanguaging (N = 146: English, Mandarin Chinese, Malay, Tamil), we compare metrics across the full sample, and common subsamples. Total speech, central tendency and extreme bounds of complexity all show significant developmental trajectories (p<.05), with larger effect sizes when larger subsamples are included in analysis. The metrics perform well when samples are maximally inclusive for languages spoken and types of speech. As some metrics performed better in certain contexts, we present a decision-tree as a reproducible, pre-registerable pathway for deciding which measurement of MLU-like syntactic complexity is best for data of different types.