Discussion
As described in the results section, the highest performing model was Model 1, which incorporated the factor weights determined by the literature review and Analytical Hierarchy Process, as well as the MSLarge membership model for all three input variables. However, the worst performing model used the same factor weights, but with the Linear membership model for each anthropogenic variable. For this difference in data transformation to have such a significant impact on the model’s overall performance suggests that the membership model has more of an effect on accuracy than do the factor weights.
Despite these differences, all models performed within 3.1% of each other, and none of the models reached 50% of the possible overlap with the validation data, so although these MCE frameworks may provide a good starting point for predicting terrestrial mammalian extinction risk, additional refinement is clearly required to build a reliable model.
Specifically, because several key areas of high extinction risk according to the IUCN validation data were not predicted in any of the models, we can conclude that other variables not included in this analysis are influencing these risk zones. Additionally, several regions were consistently over-predicted by models, including the United States and India, again suggesting that other factors are influencing the reality of extinction risk in these areas.
Study Limitations
This study presented several significant methodological and analytical challenges that may limit the accuracy and applicability of results. Firstly, the datasets used for each anthropogenic variable serve only as proxies for the actual effects of these factors on mammalian extinction risk. Although each variable (human population, agricultural conversion, invasive species) has been well-documented as a driver of changes to wildlife assemblages and abundance, the specific effects are not captured in this study. On top of this, there are limited robust, open-source global data available that visualize each of these factors in a spatial context, making more precise analyses a challenge to execute. The data used in this study also varied in their temporal scales, so results may have become skewed by mismatched data collection timelines.
Additional uncertainty was introduced by the factor weights chosen for this study. Although the models varied in the weights assigned to each variable, more iterations of testing would have allowed for more precise conclusions to be drawn.
As discussed above, one of the greatest limitations to creating an accurate model using the MCE framework was the inclusion of only three anthropogenic variables. As there were several areas of high extinction risk not accounted for by any of the models, additional factors may have increased accuracy.
Finally, many prior studies emphasized the varied impacts of different anthropogenic influences (including the variables in this analysis) in different parts of the world. Performing more spatially-constrained analyses may provide more accurate results than were yielded by the global approach taken in this study.
Recommendations for Future Work
Future research in this area is crucial to build more accurate predictive models of anthropogenic effects on mammalian extinction risk. Next steps may include taking measures to mitigate the sources of uncertainty outlined above, such as:
Incorporating data that more closely represents the actual effects of each anthropogenic variable on extinction risk (reducing the methodological “distance” between proxy and reality)
Running additional models with more factor weight variation and additional membership models
Including additional anthropogenic disturbance variables to account for more human impacts on mammal populations
Conducting finer resolution analyses at smaller spatial scales to capture regional differences in model predictive capacity
Including additional validation datasets to compare model accuracy
Lastly, as this study was undertaken to understand ways in which we might build relatively simple models to predict extinction risk, future explorations should apply these built models to inform conservation strategies, from broadly mitigating anthropogenic impacts to identifying priority regions for wildlife protection.
Conclusions
Ultimately, this project succeeded in creating predictive models that varied in their ability to accurately identify areas of mammalian extinction risk. Although the parameters of my hypothesis were not supported by my results (no model reached 50% model/validation overlap), the underlying framework for using multi-criteria evaluation and suitability analyses to compile multiple anthropogenic variables into a single model for predictive purposes remains robust. Future efforts should continue to refine the models presented here, as well as incorporate additional data sources to build more comprehensive, accurate, and functional representations.