As a final test, we performed two out-of-sample predictions using our estimates from the 1995 2007 period.
Second, to further test the generalizability of our findings, we performed an out-of-sample prediction using GDP data on the US and Japan.
However, the sufficiency of the weekly sampling for real time predictions using ANN models needs to be further evaluated against longer weekly data sets as they become available.
Given low occurrence data samples, we tested model predictions using the jackknife manipulation proposed by Pearson et al. [71], the only robust test for evaluating models based on small samples: N−1 points are used to develop N jackknifed models.
We further determined the importance of the sampling points for the lineage predictions using jackknife methodology, in order to investigate whether sampling points of a given lineage may be more important than in other lineages and thus whether differing sample sizes among lineages may have biased our predictions and conclusions.
We tested the potential importance of epidemiological parameters on model predictions using Latin Hypercube Sampling [54].
For the primary analysis of the impact of advice to remain active compared to advice to rest, the method of recycled predictions (using 1000 bootstrap samples) will be used to estimate the absolute difference in the probability of being fully recovered between the two groups, with a 95% confidence interval.
Likewise, for the secondary research question, the impact of early compared to delayed physiotherapy, the method of recycled predictions (using 1000 bootstrap samples) will be used to estimate the absolute difference in the probability of being fully recovered between the early physiotherapy group and the two delayed physiotherapy groups combined, with a 95% confidence interval.
The best predictions over all samples and methods were made by miRDeep2 on sample 2. Thus, in addition to looking at the top predictions using the combinatorial criteria described above, we also looked at miRDeep2 predictions for each sample independently.
Unsupervised clustering and sample prediction analyses using genes reported as PR-A and PR-B-specific targets further showed that BRCA-mutated samples did not exhibit an altered PR response.
