Multivariate analysis of drug safety in clinical trials with application to antimalarials in pregnancy in Malawi

Abstract

Drug safety assessment is important at every stage of drug development especially in large late-phase clinical trials that have the power to detect relatively rare or common events. Clinical trials evaluating antimalarial trials in pregnancy provide a typical setting where safety data receives non-rigorous analysis compared to efficacy data. In particular, improvements in antimalarial drug safety assessment in pregnancy have focussed on study design issues e.g. definition of safety outcomes and data collection. Safety data is usually collected longitudinally and consist of multiple correlated measured endpoints with potential time-varying observations across the study population, requiring advanced statistical methods for analysis. Analytical challenges with such complex data structure include accounting for heterogeneity, dependency between multiple safety endpoints and differences in follow-up-time, e.g. due to non-adherence or study dropout. This thesis investigates optimal statistical methods for analysis of multivariate safety data in clinical trials with application to trials on malaria chemoprevention and treatment during pregnancy. Firstly, using a systematic review, the thesis ascertained the current practice and methodological issues in statistical analysis and reporting of drug safety for malaria chemoprevention in pregnancy trials. The review showed that although most trials reported that they collected multiple longitudinal safety endpoints, statistical analyses were predominantly descriptive and nonparametric, and ignored multiple occurrence. Furthermore, impact of AEs on treatment non-adherence was rarely quantified.

Motivated by the systematic review findings, statistical methods for analysis of recurrent AEs were investigated. In clinical trials, recurrent AEs are a key multivariate outcomes that may require complex analytical methods due to correlation that arises from the individual patient unobserved heterogeneity and the dependence between events. Shared frailty with proportional hazards models can be used to address the correlation issue., However, the presence of time-dependent treatment effects can lead to poor unobserved heterogeneity estimates. Flexible univariate parametric survival models are now becoming popular in the presence of time-dependent treatment effects. Using a simulation study, flexible parametric lognormal shared frailty models were assessed whether they can improve frailty variance and hazard ratio estimates in investigating treatment effect on AE recurrence; the models were compared to the inverse Gaussian shared parametric model with proportional hazards assumption. The inverse Gaussian shared frailty model yielded higher coverage probability for the 95% confidence interval, lower mean square error and lower bias of the frailty variance estimates compared to the flexible parametric shared frailty models with non-proportional hazards and restricted cubic splines. The flexible parametric lognormal shared frailty models had unbiased log hazard ratio estimates at the expense of precision loss and high mean square error compared to the conventional inverse Gaussian shared frailty model.

In addition to clinical AEs, longitudinal continuous biochemical laboratory safety data and concomitant medication are important multiple longitudinal outcomes for interpreting drug safety in context. For example, drug safety estimates can be confounded by the exposure to the concomitant medication over the follow-up time. Joint modelling of multivariate longitudinal laboratory safety data, concomitant medication and clinical AEs efficiently harnesses the safety data in order to better understand drug safety profile. Using joint modelling, the thesis investigated whether clinical AEs vary by treatment and how laboratory outcomes (alanine amino-transferase, total bilirubin) and concomitant medication are associated with clinical AEs over time following artemisinin-based antimalarial therapy. Using the novel joint model, AEs were associated with the concomitant medication (IRR=5.75; 95% CI: 4.70, 7.03; p<0.001) but there was no evidence of association of AEs with the total bilirubin (IRR= 0.97; 95% CI: 0.60, 1.56; p=0.906) and alanine aminotransferase (IRR=1.12; 95% CI: 0.91, 1.38; p=0.269). Generally, the Poisson model yielded lower effect estimates of treatment on AE incidence compared to the joint model.

Finally, this study investigated whether nested case control (NCC) study with incidence density sampling is more efficient than NCC with path set sampling under conditional logistic or weighted cox models in assessing the effect of AEs on treatment nonadherence and participation in preventive antimalarial drug during pregnancy trial. Similar estimates were obtained under incidence density sampling and path set sampling whether weighted cox or conditional logistic models were used. The results indicated no evidence that the clinical AEs impacted treatment non-adherence or study noncompletion.

Overall, the thesis work developed and applied novel flexible parametric models to improve contextualized profiling of the antimalarial drug safety when there are multiple or repeatedly-measured and time-to-event safety outcomes. NCC was also adapted to aid assessment of impact of AEs on clinical trials and inform future trial designs.