What to do with Don't Know responses in survey data is dependent upon a prior understanding of their causes. A proximity analysis of the Don't Know responses alone yields the same scale as does a scalogram analysis of the definite responses alone indicating response uncertainty as one source of the response. The quality of the scales is improved by eliminating ambiguous items, indicating stimulus ambiguity as another source. The data used are responses to items on attitude toward abortion obtained in an island-wide survey of Taiwan. Clyde H. Coombs is Professor of Psychology, and Lolagene C. Coombs is Research Associate, Population Studies Center, at the University of Michigan. The research upon which this article was based was supported by National Science Foundation Grant GB-41877, and National Institute of Child Health and Human Development Grant HD03002-06 to The University of Michigan. The authors wish to thank James B. Rogers who carried out most of the analyses reported here. POQ 40 (1976) 497-514 This content downloaded from 157.55.39.211 on Tue, 27 Sep 2016 05:08:15 UTC All use subject to http://about.jstor.org/terms 498 COOMBS AND COOMBS I don't want to get involved. The item may be a sensitive question and the respondent does not want to reveal how he feels, the item may not be understood, or the item may be multidimensional and induce conflict between incompatible values, etc. The frequency of DK responses due to such individual personality and attitudinal factors may reflect broader social differences; for example Sicinski (1970) has reported cross-national differences in the tendency to use the DK response revealing cultural differences in the readiness to admit lack of knowledge or opinions about the subject of the question. Such DK responses pose a particular problem if the study involves asking a series of questions with the intention of developing a scale. The problem then is not to classify reasons for the uncertain responses but to decide what they mean in terms of the scale to be developed and what to do about them in analyzing the data. A great variety of reasons could be suggested to account for DK responses in survey interviews when the response is not being used solely as an intermediate response category, and probably at one time or another each of them has been a true explanation. But most of these are indistinguishable in the data and we will treat them together under the rubric of question or item ambiguity. This source of DK's will be distinguished by the fact that the DK response is not scale dependent, in contrast to DK responses which are evoked by difficulty in discrimination and hence are scale dependent. Of course the deeper reason for item ambiguity is a matter of substantial interest-but whether it is poor editing or a sensitive question is more a matter of clinical interpretation at this stage of the art. In any case the detection and elimination of ambiguous items improves a scale's reliability and interpretability. This is not necessarily accomplished by the elimination of items with a large number of DK responses, however, because many of them may be scale dependent, and DK responses which are scale dependent are, as we shall show, as useful as any other allowable response category. This paper examines how the two kinds of DK responses, those that are scale dependent and those that are due to item ambiguity, may be discriminated to a significant degree. Such a distinction bears on the problem of what to do about the don't knows that plague survey interviews. We will do this by showing that DK responses alone can be used to construct a scale and that the scale so obtained is identical to the one obtained using Guttman scalogram analysis. Further, we find that the quality of the scale is directly related to the degree that the DK responses are scale dependent and not induced by item ambiguity. This application utilizes a scale of attitude towards abortion, but the method of analysis is applicable to other questions from which attitudinal scales are to be developed. The analysis begins with a summary of a conventional scalogram analThis content downloaded from 157.55.39.211 on Tue, 27 Sep 2016 05:08:15 UTC All use subject to http://about.jstor.org/terms