Sources of bias in single-trial normalization procedures

Area of Science:

• Neuroscience
• Data Analysis
• Signal Processing

Background:

• Baseline normalization is crucial for analyzing brain activity, using a reference period to adjust data across trials.
• Common methods like pseudo z-scoring, applied to spectral power, can introduce bias, particularly with skewed data distributions.
• Existing bias in normalization techniques stems from dividing correlated terms, not just outliers.

Purpose of the Study:

• To identify the root cause of bias in common brain activity normalization techniques.
• To develop a general, effective method for reducing or eliminating normalization bias.
• To demonstrate the superiority of the proposed method over existing complex techniques.

Main Methods:

• Investigated the mathematical origins of bias in normalization, focusing on the correlation between numerator and denominator terms.
• Developed and tested a novel 'baseline fusing' method, which combines baseline periods from multiple trials.
• Compared the performance of baseline fusing against traditional methods like pseudo z-scoring and other normalization techniques.

Main Results:

• Bias in normalization techniques arises from the division of correlated variables, influenced by data distribution and method properties.
• Pseudo z-scoring introduces bias with skewed data but is unbiased with symmetric data.
• Methods like dF/F inherently exhibit bias due to correlated numerator and denominator.
• The proposed baseline fusing method effectively reduces and can eliminate bias, providing accurate estimates with sufficient baseline data.

Conclusions:

• The division of correlated terms is the fundamental cause of bias in many normalization procedures for brain activity.
• Baseline fusing offers a simple, fast, and general solution to mitigate normalization bias.
• This novel method outperforms more complex existing techniques in providing bias-free estimates for neural data analysis.