This suggestion runs counter to the growing emphasis on large samples built on secondary data, data which is assumed somehow to be more valid and generalizable than carefully col-lected primary data.