Sound an Alarm

Abstract

Considerable research is conducted which fails to examine the problems associated with non-sampling bias—a bias introduced by loss of information due to non-coverage or observation error. It is extremely important for the researcher to analyze his data for possible bias as well as to assess the possible effect a discovered bias might have on the results. This paper demonstrates how a relatively small loss of information caused by a 9.3 per cent loss of data from a total sample of nearly 20,000 cases effectively re duced the power of the sample to less than 2,000 cases. The assumption that a loss of information does not result in serious bias is not justifiable. Too few researchers realize that a very large, carelessly chosen sample that contains a bias for which it is difficult to adjust is not preferable to a smaller, carefully con trolled random sample. The latter frequently costs less and may yield more valid and reliable results.