2000 Volume 27 Issue 2 Pages 181-200
The problem of false negatives, people who really have attitudes but refrain from expressing them, could seriously bias the analysis, but has largely been neglected. Using a survey data including a number of “DON'T KNOW” responses, this paper examined whether “DON'T KNOW” respondents had underlying attitudes. We treated these nonresponses as nonignorably missing, in the sense that “DON'T KNOW” responses are related to the answer of the question in some partially unknown way. We proposed a method to estimate parameters in a logit model when the covariates are nonignorably missing. The method simultaneously employed two generalized linear models: the proportional odds model for the response variable “Party-Support”, and the multinomial logit model for the nonresponse. We found that “Don't know” responses to the Cabinet support question depended on whether the respondents supported the Cabinet, indicating the existence of false negatives. We also found that determining which party to support was based on voters' ideology, city size and stance toward the Cabinet, even with the false negatives.