There were some innefficiencies in our code. We would first download all PNADC quarters and save them to disk; then read them all again and split into panels; then identify the panels.

I moved things around to combine steps 1 and 2 into one: in the loop that loads all the data, the function now reads each quarter and already splits into panels, making it a bit faster (but probably not too much, because downloading takes the bulk of the time).