The statistical modeling of the spatial theory of voting (including non-policy-biases) leads to so far undiscussed problems due to survey related methodological problems: Eliciting answers to a set of issue items implies high cognitive and motivational requirements on the part of the respondents. We argue that Nonresponse is particularly more pronounced the higher the number of supply parties/candidates and the higher the number of “hot” issues. For example , given a fixed choice set of five J parties and the evaluation of K issues, the respondent has to provide (J*K)+K positions on the usual 11- or 7- point scales in addition to the other part of the questionnaire. The resulting sample reduction due to Item-Nonresponse in a conditional logit/probit setting is dramatically high when using listwise deletion, and somewhat smaller when applying alternative-wise deletion. Dropout leads to higher standard errors, and more generally it is questionable whether the sample still is unbiased.
Therefore, the debate on the relevance of issue voting, with opponents usually referring to the low predictive power of the spatial component should be reconsidered anew: Item-Nonresponse blurs the empirical evaluation of a model: Due to the high, probably systematic dropout it is impossible to distinguish whether the hypotheses of the spatial model are wrong or whether they are correct. For the empirical assessment of this risk we use German Election Studies in order to identify respondents with a high risk of Nonresponse. Using advanced methods of multiple imputation we intend to detect and correct the bias in spatial voting models.
