Abstract: Many statistical methods that use low-level election vote count data to detect election frauds have the limitation that they have a hard time distinguishing distortions in vote counts that stem from voters’ strategic behavior from distortions that originate with election frauds. Identifying latent components that underlie election forensics statistics and other contextual variables can help show the extent to which the statistics measure fraudulent as opposed to strategic behavior. We use an active-learning procedure with a support vector machine to classify complaints about German federal elections during 1949–2009 to show the diversity of the complaints, which we use as contextual data. We also use variables that measure strategic voting in those elections. For the elections of 2005 and 2009 we use latent variable methods to assess whether the parameters of a positive empirical model of frauds connect through latent variable structure to either the complaints or the strategic variables. Geographic ambiguity about the locations at which some complaints occur motivates embedding a geographic mixture structure in the latent variable model. The “fraud” parameters connect to both complaints and strategic behavior.