False negative rate (type II error, upper panel) and false positive rate (type I error, lower panel) for the analyses of 10000 simulated datasets in each case.

The datasets were generated by use of three different distributions (Poisson, Negative binomial and lognormal Poisson). The best model (Poisson, negative binomial and zero-inflated negative binomial for count data) with and without effect was identified and compared to each other using BML. Datasets with an effect were simulated so as to yield approximately 10% false negative for the count data analysis. Each cell contains the percentage of false negatives/positives for the data type (count of frequency) followed by the ratio between them (frequency score divided by count data score). Note that a ratio >1 is favoring the count data approach while ratio < 1 is favoring the frequency approach. These are indicated in bold. Ratios larger than 4 in favor of count data is indicated with bold italic characters.