Conditional predictions in joint species distribution models for mosquitoes: Does presence-absence data about additional species enhance predictions of the target species presence?
Jakob Langenbacher (05/2020-11/2020)
Betreuer: Stephanie Thomas
Species Distribution Models (SDMs) only model one species. JSDMs provide two prediction types on the species-level: (i) unconditional and (ii) conditional predictions. Both types are compared to predictions from SDMs. In addition to these methodological investigation, the ecology of the mosquitoes Culex perexiguus and Anopheles atroparvus are examined. In specific, this work compares how the environmental responses dier between the two species and whether the residual correlation between the species is negative or positive. Presence-Absence data over one season in a Mediterranean wetland in Southern Spain is analyzed. The core model for all the models is a Bayesian probit model. The SDMs use the univariate and the JSDM the multivariate version. Since the JSDM in this analysis is only different from the SDMs with respect to the multivariate nature, reasons for different results can be attributed to the multivariate nature. Model outputs based on the different erent prediction types are compared in terms of environmental parameter estimates, response curves and measures of variable importance. The predictive performance is evaluated on a test set with the discrimination metric area under the curve (AUC). The results show that unconditional predictions of JSDMs do not differ from predictions of SDMs. Conditional predictions of JSDMs, on the other hand, differ from predictions of SDMs. They, also, have a better predictive performance. Regarding the ecology, both mosquito species respond similarly to the environment and their residual correlation is positive. In conclusion, when investigating the predictive performance of JSDMs, researchers should focus on conditional predictions. Moreover, data about other mosquitoes can give valuable information about the target mosquito which can be used for improved predictions.