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

1Experimental data of interacting cells under the influence of external as well as unobserved factors are ubiquitous. Parametric models are often used to gain understanding of the processes that generated such data. As such, biological understanding hinges upon accurate inference of model parameters. Whether and how systemic parameter bias manifests in such models is poorly understood. We study this issue in the specific context of estimating the static and dynamic interactions of simultaneously recorded neurons influenced by stimuli and unobserved neurons. Through extensive numerical study and analytic calculations, we identify and mitigate bias in such models. When applied to diverse neural data sets, we found that common models and inference procedures often overestimate the importance of coupling and underestimate tuning. In contrast to common intuition, we find that model non-identifiability contributes to estimation bias, not variance, making it a particularly insidious form of statistical error. As the experimental and statistical issues examined here are common, the insights and solutions we developed will likely impact many fields of biology.