Sample sizes and model comparison metrics for species distribution models

Species distribution models use small samples to produce continuous distribution maps. The question of how small a sample can be to produce an accurate model generally has been answered based on comparisons to maximum sample sizes of 200 observations or fewer. In addition, model comparisons often are made with the kappa statistic, which has become controversial. Therefore, we used sample sizes ranging from 30 to 2500 individuals to model 16 tree species or species groups in Minnesota's Laurentian Mixed Forest. We compared all smaller sample sizes to models for 2500 records and then 1000 records using Cohen's kappa, Pearson's r, Cronbach's alpha, and two intraclass correlation coefficients. We then began confirmation of our findings by repeating the process using a smaller extent in a different area, a portion of Missouri's Central Hardwoods. Although there are disadvantages to using the kappa statistic and intraclass correlation coefficients, due to conversion to categories or computation limitations respectively, the model comparison metrics produced similar results. Comparison values depend on the maximum sample size, and at sample sizes roughly around 10–20% of the maximum sample size, values will begin to decrease more rapidly. Models may not be very accurate below a sample size of 200, for our study areas, extents, and grains. Nonetheless, models based on small sample sizes still may provide information for rare species. We recommend using the full sample available for modeling, after using a partial sample for accuracy assessment. Future research is needed to confirm our findings for different areas, extents, grains, and species.

► We compared species distribution models for samples sizes ranging from 30 to 2500 records. ► Cohen's kappa, Pearson's r, Cronbach's alpha, and two intraclass correlation coefficients produced similar results. ► Comparison metric values depended on the maximum sample size. ► At sample sizes around 10–20% of the maximum sample size, metric values will begin to decrease more rapidly. ► Models may not be very accurate below a sample size of 200, for our study areas, extents, and grains.