Background: Most physicians do not understand the outcomes of medical tests (Gigerenzer, 2002; Hoffrage et al., 2000). In cancer screening, this undesirable state of affairs can cause overdiagnosis, overtreatment, and unnecessary anxiety. Unless physicians learn to make sense of health statistics, informed consent and shared decision making are impossible. Objectives: To test whether physicians can derive the positive predictive value (PPV) of cancer screening tests from the relevant statistical information, and to test what representation of information can help physicians to improve understanding. Methods: In a series of experiments, German doctors were asked to estimate the PPV of screening tests, including breast cancer and colorectal cancer screening. The information was provided either in conditional probabilities (sensitivity, false positive rate, prevalence) or in natural frequencies. Results: When information was given in conditional probabilities, as is typically the case in medical textbooks and journal articles, few physicians found a correct estimate for the various screening tests. The estimates for the PPVs varied widely, typically ranging between 1% and 90%, independent of whether the test was in the physician’s specialty or not. For instance, while the PPV for mammography screening is about 10%, the most frequent answer of 160 gynecologists was 90%. When the information was presented in natural frequencies, the variability strongly decreased and the modal answer was precisely at the correct PPV. The same result was obtained after I trained some 1000 gynecologists in translating conditional probabilities into natural frequencies. Conclusions: The majority of physicians I have studied do not understand how to derive a PPV from the relevant information if the latter is in conditional probabilities. However, there is a simple solution: a proper representation of the information in natural frequencies can turn physicians’ innumeracy into insight. There are transparent forms of information such as natural frequencies and absolute risks, and less transparent forms such as conditional probabilities, relative risks, and 5-year survival rates, which tend to cloud physicians’ thinking. If we want physicians to become statistically literate, transparent representations need to be taught in medical school and should be made mandatory in journal articles and information brochures.