Sampling bias is a common phenomenon in records of plant and animal distribution. Yet, models based on such records usually ignore the potential implications of bias in data collection on the accuracy of model predictions. This study was designed to investigate the effect of roadside bias, one of the most common sources of bias in biodiversity databases, on the accuracy of predictive maps produced by bioclimatic models. Using data on the distribution of 129 species of woody plants in Israel, we tested the following hypotheses: (1) that data collected on woody plant distribution in Israel suffer from roadside bias, (2) that such bias affects the accuracy of model predictions, (3) that the road network of Israel is biased with respect to climatic conditions, and (4) that the impact of roadside bias on model predictions depends on the magnitude of climatic bias in the geographic distribution of the road network.As expected, the frequency of plant observations near roads was consistently greater than that expected from a spatially random distribution. This bias was most pronounced at distances of 500-2000 m from roads, but it was statistically significant also at larger scales. Predictive maps based on near-road observations were less accurate than those based on off-road or ''rectified'' observations (observations corrected for roadside bias). However, the magnitude of these differences was extremely low, indicating that even a strong bias in the distribution of species observations does not necessarily deteriorate the accuracy of predictive maps generated by bioclimatic models. Further analysis of the data indicated that the road network of Israel is relatively unbiased in terms of temperature, and only weakly biased in terms of rainfall conditions. The overall results are consistent with the hypothesis that the impact of roadside bias on model predictions depends on the magnitude of climatic bias in the geographic distribution of the road network. We discuss some theoretical and practical considerations of bias correction in biodiversity databases.

No abstract

Effective application of species distribution models requires some knowledge concerning the accuracy of model predictions. Yet very few studies have attempted to systematically analyze factors affecting the predictive power of distribution models. This study fills this gap for Climatic Envelope Models, which have been applied extensively for a variety of conservation and management purposes. We hypothesized that model predictions are influenced by properties of the data (both quantity and quality) and distribution properties of the modeled species. Hypotheses concerning the effects of both types of factors were tested by analyzing distribution patterns of 192 species of woody plants in Israel. Analyses were based on Monte Carlo simulations and standard statistical tests. The total number of observations had a strong positive effect on model performance; but on average, 50-75 observations were sufficient to obtain the maximal accuracy. Climatic bias (the degree of sampling bias with respect to climatic conditions) had a significant negative effect on predictive accuracy. Climatic completeness (the degree to which the climatic range occupied by the species is covered by the observations) had a negative effect on model performance-a result contradicting our original hypothesis. Among the species properties, commonness had a positive effect while niche width had a negative one. Niche position with respect to rainfall and temperature was also important in determining the accuracy of model predictions. The overall results are discussed with respect to trade-offs between commission and omission errors and the potential implications of scale dependency.