Keywords

Pillai-score reliability, merger underreporting, data-size effects

Abstract

Pillai Scores

Are they good measures?

  • Often supported in meta-analyses (Nycz & Hall-Lew 2013, Kelley & Tucker 2020)
  • Appear to be better than Euclidean distance, mixed-effects regression, spectral overlap, SOAM, VOACH, APP, etc.
  • However, Bhattacharyya’s Affinity may be better suited for vowel data since it’s robust to skewed data. (Johnson 2015)

Implications:

Mergers are probably underreported and separation is probably overreported

  • It takes a lot of data (more than many studies use) to get reliably low Pillai scores
  • Mergers may be more common previously reported

Statistical significance should be reported

  • Reporting p -values from MANOVA tests removes the need for ad hoc thresholds. puts less weight on Pillai scores, and makes interpretation more objective.

Comparison across studies

  • Analyses of speakers with less data will look less merged
  • Analysis of speakers with more data will look more merged

Comparison within studies

  • Speakers with less data will have inflated Pillai scores.
  • Reading tasks will have higher scores than conversational data (which will be interpreted as style differences)

Original Publication Citation

Joseph A. Stanley & Betsy Sneller. “Sample size matters when calculating Pillai scores.” Poster presentation at the 181st Meeting of the Acoustical Society of America (ASA). Seattle, WA. November 29, 2021

Document Type

Poster

Publication Date

2021

Publisher

181st Meeting of the Acoustical Society of America

Language

English

College

Humanities

Department

Linguistics

University Standing at Time of Publication

Assistant Professor

Download

Included in

Linguistics Commons

Share

COinS