An item bank is a set of test items calibrated on a single scale, independent of the sample used for calibration purposes. Key concepts which arise from this definition are those of 'dimensionality', 'ability', and 'difficulty', all of which must be investigated both at the pre- and post-test construction stages. A discussion of the dimensionality of reading in English as a foreign language (EFL) suggests that there is in fact a single dimension to a construct which is often analysed in terms of separately identifiable sub-skills. Testing 'ability' is best approached within a criterion-referenced framework, where rigorous attempts to ensure content validity are necessary at the test writing stage. Using criterion-referenced measurement as a philosophy of test construction helps in the analysis of what is and is not 'difficult' about EFL reading test items. Further investigation of difficulty highlights the systematic procedure. Specific technical questions relating to item banks are discussed, in particular the use of the Rasch model is justified for the development of an EFL reading item bank. The principles discussed are applied to a practical testing situation, and the results analysed in terms of the key concepts. It is concluded that EFL reading is unidimensional and that the Rasch model is appropriate for use with data of this kind. An item bank for EFL reading in a particular context (Malaysian undergraduate testing) is derived from two separate anchor tests, which are shown to achieve similar results. Thus an item bank can be constructed, though it is debatable whether the methods of test analysis employed have any great advantage over traditional methods in situations where the population is relatively homogeneous or where test items are piloted on large representative samples of the population.