Comparison of statistical methods of handling missing binary outcome data in randomized controlled trials of efficacy studies

Mukaka, Mavuto Comparison of statistical methods of handling missing binary outcome data in randomized controlled trials of efficacy studies. Doctor of Philosophy thesis, University of Liverpool.

	PDF Mavuto_Mukaka_PhD_Thesis_Final_27_11_13.pdf - Unspecified Access to this file is embargoed until Unspecified. Available under License Creative Commons Attribution No Derivatives. Download (1MB)
	PDF (MukakaMav_July2013_14593.pdf) MukakaMav_July2013_14593.pdf - Unspecified Available under License Creative Commons Attribution No Derivatives. Download (1MB)

Abstract

The presence of some missing outcomes in randomized studies often complicates the estimation of measures of effect, even in well designed randomized controlled trials. The process may be complicated further when the efficacy rates are close to 0% or 100% as the standard binomial model is susceptible to model non-convergence. The main objective of this study was to compare the performance of multiple imputation (MI) and Complete Case analysis for dealing with missing binary outcomes when modeling a risk difference. Firstly, however, the binomial regression COPY method and the Cheung’s modified Ordinary Least Squares (OLS) method were examined using simulation processes for their appropriateness in risk difference modeling. It was found that the number of copies (for the COPY method) required to minimize non-convergence coincided with the number of copies that gave the most biased estimates of the true efficacy difference while increasing the number of copies made the problems of non-convergence and bias worse; using Cheung’s method, however, there was 100% convergence with unbiased estimates of effect size. Simulation methods were used to compare the performance of complete case (CC) analysis and several multiple imputation (MI) models for handling missing outcome data over a wide range of efficacy environments and missing value assumptions. When outcomes were missing at random (MAR) or completely at random (MCAR), MI analyses that included treatment group membership in the imputation calculations yielded unbiased estimates of efficacy differences. The CC method was found to be as good, and often better, than MI methods when outcomes were MAR or MCAR, with coverage close to 95% in many situations – but neither CC nor MI produced unbiased estimates of effect difference when outcomes were missing not at random (MNAR). It was concluded that CC and MI methods are equally good in terms of producing unbiased estimates of effect difference in most missing outcome situations, but applying the intention to treat principle (ITT) which requires all randomized patients to be included in the primary analysis of a RCT, MI should be adopted as the analysis method of first choice, accompanied by a secondary CC analysis for sensitivity purposes (i.e. to investigate the extent of any likely bias).

Item Type:	Thesis (Doctor of Philosophy)
Additional Information:	Date: 2013-07 (completed)
Uncontrolled Keywords:	Missing data, Randomized Controlled Trials, Multiple Imputation, Simulation
Divisions:	Faculty of Health and Life Sciences
Depositing User:	Symplectic Admin
Date Deposited:	07 Feb 2014 12:36
Last Modified:	17 Dec 2022 01:28
DOI:	10.17638/00014593
Supervisors:	Faragher, Brian White, Sarah
URI:	https://livrepository.liverpool.ac.uk/id/eprint/14593