Predictive Accuracy of Polytopic Vector Analysis in Environmental Forensics: Sensitivity to Seeding Methods, Random Noise, and Sample Size

Although the sensitivity of receptor models to seeding methods, random noise, and sample size are frequently discussed in environmental forensics, a rigorous evaluation of these factors has been lacking. We generated 435,600 unique datasets with between two and seven sources of PCBs or PFAS that attempt to approximate real-world environmental datasets. Evaluation of these simulated datasets with PVA shows that EXRAWC and NNDSVD seeding methods converge more often and are typically more accurate when sources are similar to each other or are present in small proportions. The results also show that increasing sample sizes up to 25 samples can improve predictive accuracy.