How does Naïve Bayes behave under likelihood misspecification in real-world datasets?

Naïve Bayes remains stable under mild likelihood misspecification but degrades when distributional assumptions are severely violated.

Likelihood misspecification occurs when assumed distributions (Gaussian, multinomial) do not match true data distributions. NB still often performs well due to its focus on decision boundaries rather than exact density estimation. However, probability calibration becomes unreliable, and decision boundaries may shift under severe mismatch.