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

• DOI: 10.1101/2023.07.11.548615
• Publication Date: 2023-07-15T16:15:10Z
• Authors: Pratik Sachdeva; Ji Hyun Bak; Jesse Livezey; Christoph Kirst; Loren Frank; Sharmodeep Bhatta...; Kristofer E. Bouc...
• Keywords: Identifiability; Statistical Inference

In the brain, all neurons are driven by activity of other neurons, some which maybe simultaneously recorded, but most not. As such, models neuronal need to account for recorded and influences unmeasured neurons. This can be done through inclusion model terms observed external variables (e.g., tuning stimuli) as well latent sources variability. Determining influence groups on each relative is important understand brain functioning. The parameters statistical fit data commonly used gain insight into importance those influences. Scientific interpretation hinge upon unbiased parameter estimates. However, evaluation biased inference rarely performed bias poorly understood. Through extensive numerical study analytic calculation, we show that common procedures typically biased. We demonstrate accurate selection before estimation resolves non-identifiability mitigates bias. diverse neurophysiology sets, found contributions coupling often overestimated while exogenous underestimated in methods. explain heterogeneity biases across sets statistics. Finally, counter intuition, contributes bias, not variance, making it a particularly insidious form error. Together, our results identify causes neural data, provide mitigate reveal impact sets.