Statistical Tests with MUSHRA Data

Mendonça C, Delikaris-Manias S. Statistical Tests with MUSHRA Data. In: AES Convention 144. Audio Engineering Society; 2018. Paper 10006. Available from: https://aes.org/publications/elibrary-page/?id=19402

@inproceedings{Mendonca2018_19402,
  author = {Mendonça, Catarina and Delikaris-Manias, Symeon},
  title = {{Statistical Tests with MUSHRA Data}},
  booktitle = {AES Convention 144},
  note = {Paper 10006},
  year = {2018},
  month = may,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=19402}
}

TY  - CPAPER
TI  - Statistical Tests with MUSHRA Data
AU  - Mendonça, Catarina
AU  - Delikaris-Manias, Symeon
T2  - AES Convention 144
M1  - Paper 10006
PY  - 2018
DA  - 2018/05/06
UR  - https://aes.org/publications/elibrary-page/?id=19402
PB  - Audio Engineering Society
LA  - en
AB  - This work raises concerns regarding the statistical analysis of data obtained with the MUSHRA method. There is a widespread tendency to prefer the ANOVA test, which is supported by the recommendation. This work analyzes four assumptions underlying the ANOVA tests: interval scale, normality, equal variances, and independence. Data were collected from one experiment and one questionnaire. It is found that MUSHRA data tend to violate all of the above assumptions. The consequences of each violation are debated. The violation of multiple assumptions is of concern. The violation of independence of observations leads to the most serious concern. In light of these findings, it is concluded that ANOVA tests have a high likelihood of resulting in type 1 error (false positives) with MUSHRA data and should therefore never be used with this type of data. The paper finishes with a section devoted to statistical recommendations. It is recommended that when using the MUSHRA method, the Wilcoxon or Friedman tests be used. Alternatively, statistical tests based on resampling methods are also appropriate.
ER  -