Statistical Concerns in Analysis of Adiponectin and Cortisol Data
Allan Walkey, Boston University School of Medicine
3 December 2009
We read with interest the report by Venkatesh et al investigating changes in adiponectin during critical illness. Results show a significant relationship between adiponectin concentration and cortisol levels in critically ill subjects (eg., an R2 of 0.32 with p=0.01 at day 3). Upon further review of the data we observed that one extreme outlier exists in the cortisol day 3 data set (level 2620 nmol/L, with all other values below 600 nmol/L). As the authors point out, this outlier skews the cortisol data to a non-parametric distribution. Thus, in this case, the use of a simple linear regression to analyze the relationship between adiponectin and cortisol violates the normality assumptions needed for valid linear regression analyses. More appropriate analyses of these data would include reporting the Spearman correlation coefficient (as mentioned in Methods section) or log-transformation of the data for linear regression analysis. Performing these analyses, we found the Spearman correlation coefficient for the relationship between adiponectin and cortisol to be 0.51 with p=0.17; alternatively, linear regression with log-transformed values led to an R2=0.06 with p=0.33. Results demonstrate that when using non-parametric statistical methods, the relationship between adiponectin and cortisol at day 3 is no longer statistically significant. Despite this, we agree with the authors conclusion that “the relation between adiponectin and the inflammatory response, organ dysfunction and outcome in critical illness” is a subject worthy of future investigations.
Critical Care
Statistical Concerns in Analysis of Adiponectin and Cortisol Data
3 December 2009
We read with interest the report by Venkatesh et al investigating changes in adiponectin during critical illness. Results show a significant relationship between adiponectin concentration and cortisol levels in critically ill subjects (eg., an R2 of 0.32 with p=0.01 at day 3). Upon further review of the data we observed that one extreme outlier exists in the cortisol day 3 data set (level 2620 nmol/L, with all other values below 600 nmol/L). As the authors point out, this outlier skews the cortisol data to a non-parametric distribution. Thus, in this case, the use of a simple linear regression to analyze the relationship between adiponectin and cortisol violates the normality assumptions needed for valid linear regression analyses. More appropriate analyses of these data would include reporting the Spearman correlation coefficient (as mentioned in Methods section) or log-transformation of the data for linear regression analysis. Performing these analyses, we found the Spearman correlation coefficient for the relationship between adiponectin and cortisol to be 0.51 with p=0.17; alternatively, linear regression with log-transformed values led to an R2=0.06 with p=0.33. Results demonstrate that when using non-parametric statistical methods, the relationship between adiponectin and cortisol at day 3 is no longer statistically significant. Despite this, we agree with the authors conclusion that “the relation between adiponectin and the inflammatory response, organ dysfunction and outcome in critical illness” is a subject worthy of future investigations.
Sincerely,
Allan Walkey, MD
Ross Summer, MD
Competing interests
We declare no competing interests.