Despite this fact, there still exists a serious need to develop more robust predictive models.
This approach may be advantageous when the desired outcome is a screening tool (for instance, for disease, or for environmental pollution), when a single (or few) robust biomarker(s) will enable the use of simpler and more robust predictive models (for example, classical linear discriminant analysis [ 49]) and targeted detection.
In an era where tracking all of that data becomes increasingly important thanks to more robust predictive tools and considerably more processing power to make those projections, SurveyMonkey's data is likely even more valuable than it was when it raised funding in 2014.
BMA accounts for model uncertainty by averaging over the posterior distributions of multiple models, allowing for more robust predictive performance.
Thus, the idea of a biomarker profile, rather than the single use of one of these techniques, may offer more robust predictive power to determine who is going to convert to AD with acceptable reliability [ 72].
Random forest yielded more, and more robust predictive signatures, with nine out of the twelve signatures generated remaining predictive when tested against the CCLE dataset (17-AAG, AZD6244, Erlotinib, Lapatinib, Nultin-3, Paclitaxel, PD0325901, PD0332991, PLX4720).
We ran ten cross-validation replicates for each model and calculated the average AUC of the ten cross-validations (mean ± SD) to give a more robust estimate of predictive performance.
For example, Alvarado-Herrera et al. (2015) reported the lack of a more robust demonstration of the predictive validity whereas Kim et al. (2011) of the nomological validity.
An OPLS-DA model based on these core homogeneous subsets, without spectra contributing idiosyncratic responses, will have an improved predictive ability and potentially a more robust selection of biomarkers.
