Because you're sampling from more recent months, you're biasing the data by including only faster files (excluding all that haven't made it this far). To compare to a sample including slower files, you should go into files further back, ideally including ones that are past the estimated...