Nonresponse Bias

Nonresponse occurs when all the sampling units selected for a sample in a survey are not interviewed. A sampling unit could be an individual, household, business, or other entity being interviewed. The term nonresponse bias refers to the potential bias that can occur in surveys due to nonresponse. Particularly since the start of the 21st century, surveys based on probability samples have been experiencing declining response rates, and therefore nonresponse bias has become a growing concern for surveys in all fields. Nonresponse bias can impact education surveys as well as educational measurement and research based on surveys. If survey participants are systematically different from nonparticipants on measures related to the study, then the accuracy of the estimates, analysis, and inferences from the survey results will be affected. This entry describes the definition, identification, and measurement of nonresponse bias and describes the techniques used to adjust for it. It concludes with a list of resources for further reading on nonresponse bias.

[Page 1152]A famous example of nonresponse bias is from the 1936 presidential election in which Democrat Franklin D. Roosevelt and Republican Alfred Landon were the two candidates. The Literary Digest voter survey predicted that Landon would beat Roosevelt. The prediction was based on only 2.4 million responses from a total of 10 million mail-in questionnaires (a 24% response rate). Poll results indicated that Landon would win a majority and Roosevelt was expected to get only 43% of the vote. Actually, Roosevelt won the election with 62% of the vote. Nonresponse bias was one of the reasons for this error because respondents tended to be Landon supporters and nonrespondents tended to support Roosevelt. An additional reason for the error was sampling bias due to the undercoverage of low-income voters who tended to be Democrats.