Figure 5

Examining model predictions. Predicted values from the four models can be compared to the data distribution to assess their suitability. The normal model is clearly inappropriate as 19% of the values are negative and the shape of the distribution looks nothing like the data. The Poisson model is better because only positive values are predicted, but it underpredicts the number of zeros (predicted = 8%, actual = 17%) and one counts, and slightly overpredicts counts between two and seven. The quasi-Poisson model is similar to the Poisson, and the distribution from the negative binomial model is the closest to the actual data.