Full, Base and Cross Sample Clauses

Full, Base and Cross validation Datasets To ensure that the data selected for EEM development were representative of more than one of the monitored sites, the EPA limited the dataset for use in EEM development to those records for which data values were available for all selected predictor variables. The number of 30-minute observations for which NH3 emissions and all of the selected candidate predictor variables were available was 10,783. This dataset is referred to as the “full” dataset. From the full dataset, the EPA withheld a cross-validation dataset for use in testing the EEMs and ensuring that any relationships predicted by the EEMs would apply generally, rather than being attributable to anomalies that might be present in the data. To create the cross- validation dataset, the EPA randomly selected 2,191 (approximately 20 percent) of the 10,783 observations in the full dataset. The remaining 8,592 observations comprised the “base” dataset. Ideally, 20 percent of the observations representing different combinations of conditions would be withheld to ensure that neither the cross-validation dataset nor the base dataset over or under-represents any one set of conditions. Because there are so many predictor variables for the NH3 lagoon dataset, sub-dividing the data into many different sets of conditions would result in subsets with few observations, and it would be difficult to ensure that 20 percent of any one set is actually withheld. To ensure that the cross-validation and base datasets selected initially did not disproportionately over- or under-represented one or more sets of conditions, the EPA created two additional cross-validation datasets and corresponding base datasets by randomly selecting approximately 20 percent of the observations to withhold. Although this report presents results for the first base and cross-validation datasets only, the EPA checked for notable aberrations in results for the other two cross-validation datasets as well.