|Home | About | Journals | Submit | Contact Us | Français|
We read with concern the paper by Ong, Pitts, and Younossi,(1) having ourselves conducted a similar analysis of elevated serum alanine aminotransferase (ALT) and mortality using the same dataset, i.e. the National Health and Nutrition Examination Survey (NHANES) III Linked Mortality File.(2) In contrast to their analysis, we found no increase in overall mortality with elevated ALT. Ong and colleagues reported an adjusted hazard ratio (HR) for overall mortality of 1.038 with a 95% confidence interval (CI) of 1.036–1.041. In contrast, our adjusted HR estimate for all-cause mortality was 1.2 with a 95% CI of 0.88–1.6.
While there were differences between the two papers in the selection of the analytical samples and in the definitions of elevated ALT versus what they deemed as “NAFLD”, we believe the most likely explanation for these contrasting findings was incorrect statistical methods used by Ong and colleagues for the calculation of variance from the complex NHANES III sample. NHANES has always used a highly stratified sample in order to derive health measures for the U.S. population. As such, analyses require consideration of the complex sample design in order to determine the precision and statistical significance of any statistic. As advised by the NHANES III sponsor,(3) when using the sample weights the strata and primary sampling unit pairings from the sample design should also be used in estimating variances and testing for statistical significance. To do this it is necessary to use computer software packages such as SUDAAN or Stata. Ong and colleagues reported using SAS software, which does not have a procedure to properly estimate variances from complex samples when using Cox proportional hazards regression analysis. Consequently, standard errors are underestimated and confidence intervals are inappropriately narrow. For example, if we had incorrectly analyzed our data using SAS instead of SUDAAN, we would have obtained a hazard ratio estimate of 1.192 with a 95% CI of 1.189–1.194. While our hazard ratio estimate would remain unchanged (1.2 when rounded to one decimal place), the confidence interval becomes inappropriately small, leading to the incorrect conclusion of statistical significance. This problem can be easily seen by rounding to 2 decimal places resulting in a hazard ratio of 1.19 with 95% confidence interval of 1.19 to 1.19. Such precision exists in no medical study. We suggest that if Ong and colleagues were to repeat their analyses using the recommended statistical methods for variance estimation, they would no longer find increased overall mortality in “NAFLD”.
This work was supported by a contract from the National Institute of Diabetes and Digestive and Kidney Diseases (HHSN267200700001G).