Effect of Roadside Bias on the Accuracy of Predictive Maps Produced by Bioclimatic Models

Abstract

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.

Completeness of Digital Accessible Knowledge of Plants of Benin and Priorities for Future Inventory and Data Discovery

“…These results coincide with those of Kadmon et al (2004), who affirmed that bias in distributional data is common in the form of high concentrations of collection sites along roads. Souza-Baena et al (2013) also found road bias effects on plant data completeness across Brazil, and Hijmans et al (2000) reported that most gene bank accessions for wild potatoes in Bolivia were collected within 2 km of roads, 3-fold greater than random expectations.…”

A methodological framework to quantify the spatial quality of biological databases

“…Although Kadmon et al (2004) found significant differences between the distribution of collection localities and that of the rainfall conditions based on a random selection of localities in the study area, they demonstrated that predictions of habitat suitability were not biased, since the statistical difference was weak (although significant). In this study, strong statistical differences were found and therefore we conclude that model predictions based on the current database are likely to produce biased and misleading estimates of species range predictions.…”

“…We used a similar approach to identify factors influencing spatial bias in the database and estimating environmental bias as implemented earlier by Kadmon et al (2004) and Loiselle et al (2008). Although Kadmon et al (2004) found significant differences between the distribution of collection localities and that of the rainfall conditions based on a random selection of localities in the study area, they demonstrated that predictions of habitat suitability were not biased, since the statistical difference was weak (although significant).…”

“…Some research has been done to deal with this issue (e.g. Kadmon et al 2004, Allouche et al 2008, De Marco et al 2008, Phillips et al 2009) and the methodologies recommended can potentially be also applied for the database (see Dormann et al 2007 for a review of methods to deal with spatial autocorrelation). Algar et al (2009) made a comparison of the two main approaches described above to estimate species richness patterns and predictions to the future and found that after dealing with the spatial autocorrelation issue the first approach (i. e. empirical diversity theory approaches) fared significantly better.…”

Assessing the completeness of bryophytes inventories: an oceanic island as a case study (Terceira, Azorean archipelago)

“…There are few records of where species were looked for but not found, i.e., sites with recorded and reliable absences (but see standardized ecological studies). Numerous examples exist where the maps of diversity generated with census information do not show the real distribution of species but, instead, they represent the maps of collections that were conducted according to their cultural or landscape interest, or for accessibility reasons (Kadmon et al 2004;Pautasso and McKinney 2007;Sánchez-Fernández et al 2008).…”