There are limitations to our analysis that could be improved in future studies. The current lack of a gold standard for CNVs across the entire size spectrum makes it difficult to accurately assess the false-discovery rate. We consider the `gold standard' data sets used in the current analysis to be the best available to date, but the analysis should be updated once higher quality sequence-based data for both deletions and duplications exist. Another limitation of the data analysis is that we have not tested every algorithm across a large range of parameters. Our settings are based on substantial experience analyzing the same type of data from the same laboratories and on previous studies where optimal parameters have been established. We note that Birdsuite (and possibly other algorithms) have been trained on HapMap samples, raising the possibility of biased results. However, we see no clear evidence of this in comparison of the HapMap and non-HapMap samples in our study. Not all experiments passed quality control thresholds, but this was mainly the case for the lower resolution platforms (Affymetrix 250K, Illumina 650Y).
