Resolving Non-identifiability Mitigates Bias in Models of Neural Tuning and Functional Coupling.

Area of Science:

• Computational Neuroscience
• Statistical Modeling
• Neurophysiology

Background:

• Neuronal activity is influenced by both recorded and unmeasured neurons, necessitating models that account for both.
• Statistical models are crucial for understanding neuronal influences, but their parameter estimates can be biased.
• Sources and evaluation of bias in neural data models are poorly understood.

Purpose of the Study:

• To identify and understand the causes of statistical biases in common models of neural data.
• To develop and demonstrate inference procedures that mitigate bias in neural data analysis.
• To evaluate the impact of these biases on scientific interpretation across diverse neurophysiology datasets.

Main Methods:

• Extensive numerical simulations and analytic calculations to evaluate inference procedures and models.
• Analysis of diverse neurophysiology datasets to assess parameter estimation biases.
• Investigated the role of model non-identifiability in statistical bias.

Main Results:

• Common inference procedures and statistical models for neural data are typically biased.
• Neuron-to-neuron coupling contributions are often overestimated, while tuning to external variables is underestimated.
• Model non-identifiability was found to contribute to bias, not variance, in parameter estimation.

Conclusions:

• Accurate parameter selection before estimation is critical for resolving model non-identifiability and mitigating bias.
• The study provides practical inference procedures to improve the accuracy of neural data analysis.
• Understanding and correcting statistical biases are essential for reliable scientific interpretation of brain function.