Introduction
In 2020, Neville et al. published a study exploring the computational underpinnings of "judgement bias" in rodents. The research aimed to understand how a rat's internal state, shaped by recent experiences with rewards or punishments, influences its interpretation of ambiguous signals. Recently, the authors identified a statistical oversight in their model comparison methodology and a coding error in their data visualization. This formal correction addresses these issues, specifically transitioning from the standard Bayes Information Criterion (BIC) to the more robust integrated BIC (iBIC) for hierarchical model selection. Despite these technical adjustments, the fundamental biological conclusions regarding rodent decision-making remain statistically sound.
The Judgement Bias Task
The experimental framework utilized a Judgement Bias Task (JBT). In this task, rodents are trained to associate specific auditory tones with distinct outcomes: a "positive" tone leads to a reward (e.g., sucrose), while a "negative" tone is associated with a punisher (e.g., air-puffs or aversive ultrasonic vocalizations). Once trained, the animals are presented with ambiguous "midpoint" tones. The proportion of "stay" responses—interpreted as an optimistic or pessimistic bias—serves as a proxy for the animal's underlying affective state or "mood."
Methodological Refinement: Transitioning to iBIC
The technical crux of this correction lies in the method used to compare different computational models of behavior. In the original analysis, the authors employed the standard Bayes Information Criterion (BIC). While widely used, standard BIC can be problematic for complex hierarchical models where individual animal behaviors are nested within group-level trends. Standard BIC penalizes model complexity based on the total number of parameters and data points but often fails to adequately account for the distribution of individual differences.
To rectify this, the authors adopted the integrated Bayes Information Criterion (iBIC). This approach is significantly more rigorous for models utilizing empirical priors. The iBIC approximates the marginal likelihood by sampling individual parameters from the group-level distribution and averaging them. This provides a more accurate assessment of which model best explains the observed data without the risk of over-fitting, which is a common challenge in hierarchical Bayesian modeling.
Model Parsimony and Selection
The shift to iBIC led to a change in which model was deemed the most "parsimonious"—the simplest model that provides a high-quality fit to the data. Under the new criteria, the most parsimonious model excluded the parameter $\delta$ (delta), which represents a specific component of decision bias.
However, the authors observed that in the "full" model (which includes all parameters), both $\delta$ and $C_{p/r}$ (representing the relative costs of rewards and punishers) remained statistically significant. This posed a choice: follow the strict dictates of parsimony or retain the more complex model that captures subtle treatment effects. The authors opted to maintain the full model for their final analysis, ensuring that the nuances of how different treatments (e.g., USVs vs. task-specific punishers) affect judgement were not lost. This decision underscores the importance of balancing statistical simplicity with biological relevance.
Results and Predictive Validity
The recalculation of model rankings using $\Delta iBIC$ confirmed that the model including parameters ($\delta, C_{p/r}, \lambda, \sigma, \omega, \beta$) remains a strong candidate for explaining the data. Crucially, the parameter estimates themselves remained significant even with the corrected statistical framework. For instance, the log-transformed estimates for the cost parameter $C_{p/r}$ were significantly less than zero (mean ± SE: $-1.103 \pm 0.023, p < 0.001$), as were the estimates for the bias parameter $\delta$ (mean ± SE: $-0.100 \pm 0.045, p = 0.026$).
Furthermore, the correction fixed a coding error that had previously misaligned the "Generated" data plots. The updated Figure 1 demonstrates a high degree of alignment between the "Observed" behavioral responses and the "Generated" data produced by the computational model.
This visual alignment serves as a critical validation of the model's predictive power. It confirms that the computational framework accurately captures the decision-making patterns observed in the rodents across the spectrum of tones, from positive to negative.
Conclusion
This correction serves as a methodological cautionary tale for the field of computational biology. It highlights how the choice of information criteria in hierarchical modeling can influence model selection and the interpretation of specific parameters. By adopting iBIC and correcting visualization errors, the authors have enhanced the statistical integrity of their work. Most importantly, the primary biological insight—that the history of reward and punisher experiences significantly and predictably alters rodent decision-making—is reaffirmed as a robust finding.
Summary updated