Bias in Area Under the Curve for Longitudinal Clinical Trials With Missing Patient Reported Outcome Data

Melanie L Bell,Diane L Fairclough,Madeleine T King

doi:10.1177/2158244014534858

Melanie L Bell, Diane L Fairclough + Show 1 more

Open Access

https://doi.org/10.1177/2158244014534858

Copy DOI

Abstract

A common approach to the analysis of longitudinal patient reported outcomes (PROs) is the use of summary measures such as area under the time curve (AUC). However, it is not clear how missing data affects the validity of AUC analysis. This study aimed to compare the use of AUC summary measures (in individuals) with AUC summary statistics (on groups, calculated from the estimated parameters of a mixed model) when data are complete, missing at random, and missing not at random. A simulation experiment based on a two-armed randomized trial was carried out to investigate the precision and bias of AUC in longitudinal analysis where missingness, trajectory, and missingness allocation were varied. Summary measures AUC with ad hoc approaches to missing data were compared with mixed model AUC summary statistics. AUC summary statistics were consistently superior to AUC summary measures in terms of precision and bias. The bias of AUC summary statistic approach was very small, even when data were missing not at random and when differential attrition between groups existed. AUC summary measures on individuals should not be used to analyze longitudinal PRO data in the presence of missing data.

Highlights

Patient reported outcomes (PROs), including quality of life (QoL), are becoming recognized as important elements in providing evidence for medical product labeling (Food and Drug Administration, 2009; Patrick et al, 2007)
Many applied researchers use substandard approaches; two reviews on the handling of missing data in randomized controlled trials (RCTs) showed that most RCTs have missing PRO data and have used problematic approaches (Fielding, Maclennan, Cook, & Ramsay, 2008; Wood, White, & Thompson, 2004), which can result in bias and/or inefficiency
We present the bias divided by the estimated standard error (SE) that represents how far off the t statistic for the test of difference in area under the time curve (AUC) is from the t statistic computed from a non-biased estimate

Summary

Introduction

Patient reported outcomes (PROs), including quality of life (QoL), are becoming recognized as important elements in providing evidence for medical product labeling (Food and Drug Administration, 2009; Patrick et al, 2007). Some researchers have advocated for keeping PRO analysis simple (Cox et al, 1992), it is not clear how this can be accomplished when data are missing, as PRO data often are, because they are often suspected of missing non-randomly Many applied researchers use substandard approaches; two reviews on the handling of missing data in randomized controlled trials (RCTs) showed that most RCTs have missing PRO data and have used problematic approaches (Fielding, Maclennan, Cook, & Ramsay, 2008; Wood, White, & Thompson, 2004), which can result in bias and/or inefficiency L. Bell & Fairclough, 2013; Carpenter & Kenward, 2008; D. F. Fairclough, 2010)

Objectives

Results

Conclusion