A model to predict 24-h urinary creatinine using repeated measurements in an occupational cohort study

Creatinine measurements can be used to standardize urinary pesticide concentrations and to estimate "completeness" of urine collections. Published statistical models exist to predict 24-h creatinine, but many were developed assuming independence among observations. Using correlated repeated measurement data collected from an occupational cohort, the objectives were to create a predictive model for 24-h urinary creatinine and to compare the predictive capability of this model to earlier published models. Using a mixed-model methodology, the appropriate covariance structure was identified and utilized to model the measurements. A backwards elimination model building technique applied to the model building data set (110 adult male subjects and 457 creatinine values) yielded a final model that included variables for body mass index (BMI), height, diabetes, allergies, medical conditions that affect kidney function, use of creatine supplements, and anti-inflammatory medications. Using an external model validation data set (21 adult male subjects' creatinine values, n=91 observations from a total of 275) the predictive performance of the model was evaluated using the mean square prediction error (MSPR) and the Pearson's correlation coefficient (r); its performance was better (MSPR=279184, r=0.43) than any of the earlier models investigated (MSPR: range 658860-393139; r, range 0.18-0.38). In conclusion, the use of a covariance structure that allowed repeated measurements for any one individual to be correlated, improved the predictive performance. For purposes of incomplete urine sample identification in observational studies, it is necessary to collect information in addition to age, gender, and BMI, which are typically used in these settings.