Fitting N-mixture models to count data with unmodeled heterogeneity: Bias, diagnostics, and alternative approaches