In this article, we present a theory of test translation whose intent is to provide the conceptual foundation for effective, systematic work in the process of test translation and test translation review. According to the theory, translation error is multidimensional; it is not simply the consequence of defective translation but an inevitable fact derived, among many other reasons, from the tension between translation error dimensions--broad categories of translation errors such as those related to semantics, register, or the construct being measured--and the fact that languages encode meaning in different ways. While it cannot be eliminated, translation error can be minimized. The extent to which the translation of a test item is acceptable or objectionable can be understood as a probabilistic space defined by the frequency and the severity of translation errors. Accordingly, the translation of an item can be objectionable because it has few severe errors, many mild errors, or many severe errors. To illustrate the theory, we discuss the methods used in an investigation that examined the quality of items translated in Mexico in the TIMSS-1995 international test comparison and present results from that investigation. The theory can contribute to ensuring proper implementation of test adaptation guidelines in international comparisons and to improving activities related to test translation and test translation review. (Contains 4 tables, 1 figure and 3 footnotes.)