Everything, altogether, all at once: Addressing data challenges when measuring speech intelligibility through entropy...

Area of Science:

• Statistical modeling
• Speech science
• Psychometrics

Background:

• Investigating unobservable traits presents data challenges like boundedness, measurement error, and heteroscedasticity.
• These data features can impede hypothesis testing for complex traits, necessitating advanced statistical approaches.
• Speech intelligibility research often encounters such data complexities, requiring robust modeling techniques.

Purpose of the Study:

• To demonstrate the efficacy of the Bayesian beta-proportion generalized linear latent and mixed model (beta-proportion GLLAMM) in managing complex data features.
• To compare the predictive accuracy of beta-proportion GLLAMM against the normal linear mixed model (LMM) for speech intelligibility data.
• To explore the model's capacity for estimating latent intelligibility and investigating speaker-related factors.

Main Methods:

• Reanalyzed aggregated entropy scores from spontaneous speech samples.
• Applied the Bayesian beta-proportion GLLAMM, requiring assumptions about data generation and probabilistic programming.
• Compared prediction accuracy and latent variable estimation with the normal linear mixed model (LMM).

Main Results:

• The beta-proportion GLLAMM demonstrated superior predictive accuracy compared to the normal LMM.
• The model successfully quantified a latent measure of speech intelligibility from entropy scores.
• The proposed model facilitated hypothesis exploration regarding speaker-related factors influencing intelligibility.

Conclusions:

• The Bayesian beta-proportion GLLAMM is effective in addressing data complexities inherent in unobservable trait research.
• This model offers enhanced predictive capabilities and robust estimation of latent constructs like speech intelligibility.
• The findings have significant implications for researchers and analysts needing to quantitatively measure intricate constructs.