Abstract
Composite endpoints combine multiple outcomes or components in order to determine the disease status of the patient or assess the efficacy of treatment in randomised clinical trials. A missing composite outcome is the result of one or more component being missing, which leads to the overall composite score being missing at the end of the study or in the intermediate assessments. Currently, if the primary composite endpoint is missing then direct imputation of the composite outcome is employed, which means that the partially available data in some of the components used to construct the composite outcome are ignored. The Disease Activity Score for 28 joints (DAS28) is a continuous measure of the current status of disease activity in Rheumatoid Arthritis (RA). Little is known on how best to impute composite outcomes. The thesis aim is to improve the understanding and handling of missing continuous composite outcome measure in RA trials.A review of the current literature showed discrepancies between the recommendation of the CONSORT statement and current practice in the handling and reporting of missing composite outcome data. Notably recommendations are in need for the authors to provide better reporting of missing data at the component level of the composite outcome and more extensive sensitivity analyses for the primary analyses. The methotrexate in psoriatic arthritis (MIPA) trial was reanalysed to demonstrate the differential missingness in the components of the composite by trial arms as well as over the follow up time. In addition, the tumour necrosis factor inhibitors against combination intensive therapy (TACIT) trial explored the extent of the unobserved data in the trial and examined the plausibility of the missing data mechanisms that give rise to missingness in the components of the composite outcome.
Monte Carlo simulations under different missing data mechanisms and levels of missing patterns were used to assess the comparative performance of imputing missing composite outcome directly or indirectly via its components. Multiple imputation with the option of linear regression or predictive mean matching (PMM) with five donors at the four follow-uptime points were used. Five substantive models were considered. Three models were at a single primary endpoint (cross-sectional), and two models, the data was analysed longitudinally.
For all the models, irrespective of missing data mechanisms and type of missing data patterns, the better performing MI approach was the indirect imputation of composite outcome via its components. This was in terms of smaller mean squared error arising from a greater precision. The two MI strategy (linear or PMM option) performed similarly for item missing pattern and dropout patterns under the three mechanisms; however, convergence issues were observed for the combination of item missing and dropout patterns under the MAR and MNAR mechanisms. Imputing the components jointly or individually were evaluated for handling missing composite outcome and produced similar results.
A separate analysis examined whether the imputation model was correctly specified by imputing the data as the missingness occurred. This was a sensitivity analyses to assess if the lost information due to the missingness could be regained under the direct or indirect imputation methods. It was found that the indirect imputation methods gain more precision than the direct method.Re-analysis of the TACIT trial result also confirmed the findings from the simulation that the indirect imputation methods produces more precise standard errors that are close to the benchmark standard errors.
These findings could be further examined in dataset with different levels of missing data patterns. The results from this thesis provide researchers with guidance for the handling and reporting of missing continuous composite outcome data in order to minimise bias and maximise precision in estimates based on composite outcomes.
| Date of Award | 1 Aug 2020 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | David Scott (Supervisor), Toby Prevost (Supervisor) & Brian Tom (Supervisor) |