In-Sun Nam and Leon Aarons
University of Manchester
For many drugs that are excreted renally, dosage regimens are often adjusted for renal function. Serum creatinine concentration is the most widely used measure of glomerular filtration rate (GFR) which itself is the most commonly used index of renal function. There are more accurate measures of GFR, such as creatinine clearance (CL) and especially, iothalamate CL, inulin CL, or iohexol CL, but these involve considerable practical problems. Several formulae relating plasma creatinine level to creatinine clearance exist, of which the most widely used formula is due to Cockcroft-Gault [1]. Since serum creatinine and creatinine CL were potentially important covariates related to the CL in a particular drug modelling exercise, alternative ways of imputing missing values other than the carry-forward method were investigated, as the latter was likely to introduce bias into the estimates of interest and their standard errors, and therefore cause an impact on various hypothesis tests [2]. Moreover, the proportion of missing data was too great in the dataset of interest for whole cases to be deleted.
The two major classes of modern missing data procedures are multiple imputation and maximum likelihood, and they are likely to yield almost identical results if the two are utilised in comparable ways [3], because they are derived from similar theoretical foundations. Firstly the two methods are generally fully parametric, utilising joint probability models for the observed and missing data, and secondly missing values are viewed as a cause of random variation to be averaged over. Of the two methods, multiple imputation has been widely used in the behavioural, biomedical, and social sciences, due to increased access to new computational methods and tools [4].
Considering the half-life of creatinine and reported autocorrelation between successive levels, especially due to its dependence of meat consumption, a continuous first order autoregressive multiple imputation model which was a non-explosive model satisfying the stationarity constraint, was originally utilised to describe the dependency between any two time-adjacent serum creatinine values. It was suggested that the slope of the reciprocal of serum creatinine versus time did not permit an accurate assessment of the progression rate of renal disease [5]. Nonetheless, due to practical difficulties, a simultaneous PK analysis of the specific example dataset was performed with serum creatinine imputations using a simpler model structure with weighted means of observed serum creatinine levels. Throughout, a Bayesian approach was taken with implementation via Markov chain Monte Carlo methods. The results were compared with mean imputation and the carrying-forward methods. Our method closely resembles the error in variable approach taken by Bennett et al. [6].
References
[1] Cockcroft, D. W. and Gault, M. H. Prediction of creatinine clearance from serum creatinine. Nephron, 16:31-41, 1976.
[2] Little, R. J. A. and Rubin, D. B. Statistical Analysis with Missing Data. John Wiley, New York, 1987.
[3] Collins, L. M., Schafer, J. L. and Kam, C. A comparison of inclusive and restrictive strategies in modern missing data procedures. Psychol. Meth., 6:330-351, 2001.
[4] Schafer, J. L. Multiple imputation: A primer. Stat. Meth. Med. Res., 8:3-15, 1999.
[5] Levey, A. S., Perrone, R. D. and Madias, N. E. Serum creatinine and renal function. Ann. Rev. Med., 39:465-490, 1988.
[6] Bennett, J. and Wakefield, J. Errors-in-variables in joint population pharmacokinetic/pharmacodynamic modeling. Biometrics, 57:803-812, 2001.
Reference: PAGE 12 (2003) Abstr 402 [www.page-meeting.org/?abstract=402]
Poster: poster