Accounting for nonignorable unit nonresponse and attrition in panel studies with refreshment samples