ABSTRACT
Data collected from surveys often have missing values due to nonresponse. For example, some sampled individualsmay not answer questions about financial matters or their personal life. In a political election survey, some respondents may not reveal their preferred political party or candidate. Censuses, like surveys, also face challenges from missing data. In an earlier era, statisticians struggled with very limited tools to address potential errors in estimation due to missing data. For example, in his seminal textbook on sample-survey techniques, Cochran (1977, Ch. 13) showed that there is potential for substantial bias in estimates of quantities of interest due to survey nonresponse but offered little in the way of guidance unless the investigator was prepared to pursue call-backs to try to convert initial nonrespondents into respondents. For binary outcomes, Cochran considered intervals incorporating bounds reflecting the possibilities that all of the missing items either took on the value 0 or took on the value 1, but he acknowledged that such intervals might be so wide as to be of little practical use.
