Provenance and Limits

Record provenance while you have it. Source, exact URL, fetch date, coverage, license, and anything odd you noticed.

Datasets get revised and removed. The version you used may not be there next year, and the fetch date is what lets a reader understand a discrepancy rather than assume you made an error.

Every honest piece of data work states what it cannot show.

What the data does not cover, and who it leaves out.

What you dropped in cleaning, and how much.

Which numbers are estimates rather than counts.

Whether the comparison you drew is correlation rather than cause.

This is where the Sankofa in Sankofa Code has some weight. Going back to fetch what was left behind includes being straight about what the record never held in the first place, and who was left out of it.

✦ The sentence to be able to write

"This shows X, among the people this dataset covers, which excludes Y, over the period Z." If you cannot fill in all four, the finding is not ready. Not because it is wrong, but because it cannot yet be checked by anyone else.