My thesis is simple: cohesion, the interdependence of textual elements, can be used to build reliable doze tests. Furthermore, reliable doze tests built using cohesion are parallel. This paper seeks to explore those hypotheses, and point out why the issues surrounding them are important. Developing good language tests is difficult. Tests must be both reliable and valid for their interpretations and uses. To be reliable, the potential sources of error in the test itself must be minimized. Reliability is affected by not only the examinee's language ability, but also by test methods, attributes of the test taker not related to the language abilities we want to measure, and random factors. Ideally, only the language abilities that we want to measure are reflected by the test scores, but in practice, this is not the case. Unfamiliarity with a test method, or dislike of a test method, affects the examinee's performance. A test on listening, for example, should not depend on a student's ability to comprehend complex written instructions. Finally, random factors such as lighting, what a test taker had for breakfast, or the temperature in the testing environment can have an effect on scores. In order for a test to be reliable, these influences must be minimized.