Information transfer analysis: a first look at estimation bias.

Area of Science:

• Speech perception research
• Auditory neuroscience
• Psychoacoustics

Background:

• Information transfer analysis quantifies speech feature transmission (e.g., voicing, place of articulation).
• A 100% transfer indicates no phoneme confusions between feature categories.
• Miller and Nicely (1955) noted maximum-likelihood estimates are biased with small sample sizes.

Purpose of the Study:

• To examine the small-sample bias in information transfer estimates.
• To assess the extent of overestimation across different data scenarios.

Main Methods:

• Analysis of a random performance model with pseudorandom data.
• Examination of a dataset from Miller and Nicely (1955).
• Review of published speech perception data from hearing-impaired listeners.

Main Results:

• The maximum-likelihood estimate for information transfer is indeed biased, overestimating the true value with small sample sizes.
• The degree of overestimation is substantial and depends on sample size, confusion matrix dimensions, and data partitioning.
• Bias was observed across all examined cases, including hearing-impaired listener data.

Conclusions:

• Small-sample bias in information transfer analysis can significantly inflate results.
• Researchers must be cautious when interpreting information transfer values derived from limited data.
• This bias has implications for understanding speech perception, particularly in clinical populations.