Abstract
Understanding the variability in drug response forms an important aspect of pharmacometrics. Various biological, statistical, clinical and mathematical concepts need to be considered to reach a unified decision point to understand and quantify sources of variability. This current work involves studies on methodological and clinical exploratory evaluation of covariates in the context of pharmacometrics. Studies have been conducted using theoretic approaches on the design of pharmacokinetic (PK) studies for latent covariates, use of a reduction in random between subject variability as a covariate selection criterion and evaluated methods to handle non-ignorable nuisance covariates. Exploratory studies were also conducted in a clinical & experimental framework for identification of suitable metrics of organ function as covariates to predict drug clearance.
Part I of this thesis includes methodological evaluation of covariates with Chapters 2, 3 and 4. Part II involves clinical exploratory evaluation of covariates with Chapters 5 and 6.
Chapter 2 involved studies on the design of pharmacokinetic studies for latent covariates. The motivating context for this work was from a single nucleotide polymorphism (SNP) believed to influence clearance. This led to exploration of the concept of latent covariates which can have uncertainty in both their distribution and frequency. Simulation studies were conducted in both linear regression and nonlinear mixed effects modelling (NLMEM) frameworks assuming both even and uneven frequencies of the covariate. The designs for latent covariates were evaluated assuming continuous, ordinal and nominal distribution of covariates. Initially, the designs were evaluated in a theoretic framework using linear regression. Then, these were evaluated in a NLMEM framework assuming direct influence of latent covariate or indirect influence of latent covariate via another observable continuous covariate on parameter of interest. It was observed that continuous models performed better than categorical models.
A covariate selection criterion was evaluated in Chapter 3. In pharmacometric analysis, a reduction in random between subject variability is used as part of standard criteria for selection of a covariate. The covariate is not selected if it failed to reduce random between subject variance (BSVR) in the model. Studies were conducted in a simulation framework to assess nested covariate models (NCM) and not nested covariate models (NNCM). Further, covariate-η interaction models were explored but were found to be marginally important. NCMs were found to be more robust to model misspecification than NNCMs which may not result in a reduction in BSVR.
Chapter 4 explores analysis methods for handling nuisance covariates. The frequency with which a covariate occurs is important when interpreting its effect size. Covariates like genotypes and concomitant medication are sometimes present at low frequencies or as rare events. Due to alpha error inflation, estimates of their effect size may be false. These are termed nuisance covariates. If ignoring the covariate influences bias in parameter estimates then these covariates were considered non-ignorable. Simulation studies were conducted to assess various methods for dealing with nuisance and non-ignorable covariates. It was found that addition of a fixed effect parameter for the non-ignorable nuisance covariate handled it effectively but that this coefficient should not be used for inference purposes. The use of Box-Cox transformation of η and case deletion were considered but found to be less effective.
Chapter 5 involved experimental work to quantify metrics of ageing. It is believed that variability would be better predicted with more reliable metrics of ageing such as biological age (BA) rather than chronological age (CA). It is proposed in this thesis that BA indices can be derived from the markers of ageing such as leucocyte telomere length (LTL) and/or its associated SNPs as modifiers on CA. A pilot study was conducted in healthy volunteers to obtain blood samples to collect DNA from young and older participants of either sex. An assay for LTL was implemented using a published qPCR assay and measured values of LTL in young and older subjects. SNP genotyping for selected SNPs was performed using Taqman® custom made probes. This chapter deals with the pilot clinical study and assay implementations. These assays were able to be implemented reliably so that they can be reasonably used for intended clinical exploratory studies. Initial findings found an association between LTL and CA that was similar to literature reports.
Chapter 6 deals with evaluation of metrics of BA for predicting kidney function. Additional blood samples were procured from an on-going clinical study. This study included pharmacokinetic data on 51Cr-EDTA as a kidney filtration marker. Using a population pharmacokinetic approach the available BA metrics were evaluated to predict 51Cr-EDTA clearance. The BA markers were compared against CA and estimated creatinine clearance as negative and positive controls respectively for predicting kidney function. It was found that BA markers did not perform better than CA and both performed worse than estimated creatinine clearance as a covariate for 51Cr-EDTA clearance.
So, to conclude, various methodological studies were performed which included design of studies for latent covariates, to assess BSVR as a covariate selection criterion and to assess methods for handling often missing and low frequency covariates as may be present in some genetic studies. In addition, these covariates were also considered and evaluated in a clinical study of ageing. Further work on covariates for estimating clearance changes due to ageing are required.