Liang Yang1, Shuying Yang2, Claire Ambery2, Alienor Berges2, Maria Kjellsson1, Mats Karlsson1
1Department of Pharmacy, Uppsala University, 2Clinical Pharmacology Modelling and Simulation, GSK
Introduction: Dropout in clinical trial results in missing data, which impacts the trial statistical analysis and subsequent model-based meta-analysis (MBMA). Dropout are categorized as: completely random dropout (CRD) if the dropout does not depend on observed/unobserved variables of relevance; random dropout (RD) if the dropout depends on observed variables but not unobserved variables; informative dropout (ID) if the dropout depends on unobserved variables¹. The mixed-effect model repeated measure (MMRM) is a likelihood-based method to handle missing dependent variable data and shows advantages in controlling type I error and bias over last-observation-carried-forward (LOCF)². The MBMA modelling data could be aggregated-level data (AD), individual-patient-level (IPD) data or both³. The AD from literature could be further classified into two types: mean of observations ignoring dropout and least-square-mean based on MMRM4. Compared to AD, IPD allow modelling responses and covariate effects on individual level which may reduce the impact of dropout. Objectives: This study aims to analyse how different strategies to handle clinical trial dropouts impact the estimates of MBMA models. Methods: This study explored multiple types of dropout mechanisms: CRD, RD, and ID. A dataset without dropout was included for comparison. The explored models and data types were: AD model based on mean of observations, ignoring dropout (AD model), AD model based on least-square-mean calculated by MMRM with unstructured covariance metrics5 with or without covariates (MMRM-AD model), IPD model based on IPD (IPD model), and IPD model plus dropout probability model based on IPD and patient dropout time (IPD-dropout model). The datasets were simulated by an IPD-dropout model, with forced expiratory volume in one second (FEV1) as the observation and about 25% dropout at the end (24 weeks) of the trials. Different mechanisms of dropouts were applied to the dropout model: no covariate on the hazard function (CRD), disease severity as a covariate for hazard (RD) (disease severity was characterized by % predicted FEV1 GOLD Stage Category at the screening phase), and individual FEV1 prediction as a covariate for hazard (ID). For RD, disease severity was also a covariate for FEV1 baseline. The simulated IPD datasets were directly estimated by the IPD model and IPD-dropout models. The AD datasets were derived by averaging IPD. To derive MMRM-AD datasets, IPD datasets were estimated by MMRM, and the estimated least-square-mean formed the dependent variable in the datasets. The AD models were estimated using the AD and MMRM-AD datasets, with similar equations as the IPD model but without IIV. To minimize other bias, baseline estimates in AD models were scaled6. Model simulation and estimation were performed 20 times for each condition using NONMEM7.5.1 and PsN5.3.1. FOCEI+LAPLACE was used for IPD-dropout model, FOCEI was used for other models. Results: For all conditions of dropout, the IPD model and IPD-dropout model performed well with |bias| <5.1% for fixed-effect parameters and |bias| <1.3% for random-effect parameters. The IPD-dropout model did not show an advantage over the IPD model, which may be due to relatively small dropout ratio (25% at the end) and enough information was obtained from early data. Under conditions of no dropout and CRD, the AD model and MMRM-AD models showed no significant bias but higher imprecision on disease progression slope (3.0 fold at CRD) and drug effect parameters (11 and 15 fold for two drugs at CRD) compared to the IPD models. With RD, where disease severity affected dropout and FEV1 baseline, both the AD model and the MMRM-AD model performed poorly: estimates of disease progression slope reached the lower boundary, and bias were -75% and 18% for the two drug effects. Adding disease severity to the MMRM analysis reduced the biases of drug effects considerably to -3.5% and 3.6%. For ID, IPD and IPD-dropout models performed well, but the two AD models estimated disease progression slope at the lower boundary. Conclusions: The IPD models showed advantages in dealing with dropout with or without dropout model, while AD models were biased under the conditions of RD and ID but not CRD. Adding dropout-relevant covariates in MMRM analysis may improve AD model estimation.
1. Diggle, P. & Kenward, M. G. Informative Drop-Out in Longitudinal Data Analysis. J. R. Stat. Soc. Ser. C Appl. Stat. 43, 49–73 (1994). 2. Siddiqui, O., Hung, H. M. J. & O’Neill, R. MMRM vs. LOCF: a comprehensive comparison based on simulation study and 25 NDA datasets. J. Biopharm. Stat. 19, 227–246 (2009). 3. Riley, R. D., Simmonds, M. C. & Look, M. P. Evidence synthesis combining individual patient data and aggregate data: a systematic review identified current practice and possible methods. J. Clin. Epidemiol. 60, 431.e1-431.e12 (2007). 4. Llanos-Paez, C. et al. Joint longitudinal model-based meta-analysis of FEV1 and exacerbation rate in randomized COPD trials. J. Pharmacokinet. Pharmacodyn. 50, 297–314 (2023). 5. Kincaid, C. D. 198-30: Guidelines for Selecting the Covariance Structure in Mixed Model Analysis. COMSYS Inf. Technol. Serv. Inc at <https://support.sas.com/resources/papers/proceedings/proceedings/sugi30/198-30.pdf> 6. Yang, L. et al. A combined model meta-analysis of aggregated and individual FEV1 data from randomized COPD trials. at <https://www.page-meeting.org/default.asp?abstract=10848>
Reference: PAGE 33 (2025) Abstr 11758 [www.page-meeting.org/?abstract=11758]
Poster: Methodology - Other topics