Voting behaviour, individual data and the ecological fallacy

We analyse a rather unique set of individual data on voting behaviour for an investiga-tion into the ecological bias. We review and clarify the conditions under which ecological inference will produce biased estimates and explain why this is going to affect different es-timation methods. Our investigation highlights that only the ecological inference methods which allow to model the effect of covariates may produce unbiased estimates. However, as the application to our data shows, even the use of covariates and more sophisticated methods of ecological inference may fail to produce satisfactory estimates. A logistic multilevel analysis is applied to assess the true effect of covariates and to estimate the importance of additional sources of random variations