[Skip to Navigation]
Citations 0
Editor's Correspondence
October 13, 2003

Understanding Statistical Information: The Problem of Numerical Interpretation

Arch Intern Med. 2003;163(18):2248. doi:10.1001/archinte.163.18.2248-a

I read with great interest the article by Swedko and colleagues.1 Their article clearly demonstrated the poor ability of serum creatinine to predict renal failure in elderly patients. There remain, however, several aspects of presentation of medical information, which are gradually being recognized.

It is not surprising to understand why most primary physicians do not refer patients to nephrologists when the patient's serum creatinine level is less than 1.7 mg/dL (150 µmol/L). Interpretation of statistical information when presented in traditional format (sensitivity, specificity, and positive predictive value) remains confusing to most physicians. In a study of 48 physicians from Munich, Germany, when the results were presented as probabilities (prevalence), the physicians could correctly estimate the positive predictive values in only 10% of cases. However, when the same information was presented as natural frequency format, the percentage of accuracy rose to 46%.2 For example, using the information presented in Table 2 in the article by Swedko et al,1 the positive predictive value of a creatinine level greater than 1.7 mg/dL in predicting renal failure would be 97%, and the value of a serum creatinine level less than 1.7 mg/dL in predicting renal failure would be 26% (1 − negative predictive value).

Add or change institution