Difficult-to-treat depressionAssessing and interpreting treatment effects in longitudinal clinical trials with missing data
Introduction
Treatment effects are often evaluated by comparing change over time in outcome measures; however, valid analyses of longitudinal data can be problematic, particularly if some data are missing for reasons related to the outcome measure Diggle et al 1994, Milliken and Johnson 1993. Missing data are an almost ever-present problem in clinical trials, and numerous methods for handling missingness have been proposed, examined, and implemented (Verbeke and Molenberghs 2000). In fact, there are so many methods that choosing a suitable method and interpreting its results can be difficult.
Perhaps the best place to start in determining how to analyze and interpret longitudinal data, especially in the presence of missingness, is to realize that no universally best method exists. This implies that the analysis must be tailored to the situation at hand. This, in turn, implies that the characteristics of the missing data must be understood. And thus, the objectives of this article are as follows: 1) to examine the characteristics of missing data that influence analytic choices; 2) to examine the attributes of common methods of handling missing data; and, 3) to use the data characteristics and the attributes of the various analytic methods, along with empirical evidence, to develop a robust approach for the analysis and interpretation of data from longitudinal clinical trials. Although certain statistical concepts are inherent to this discussion, we approach it from a clinical perspective and provide references to the more detailed statistical literature when appropriate. Our aim is to translate technical aspects of data analysis to an audience that is familiar with clinical research and clinical trials but is not expert in statistics. Our focus is on applications to neuropsychiatric disorders, with ideas fixed via a specific application to a clinical trial of an antidepressant; however, the concepts covered herein have a broad range of applications.
Section snippets
Missing data
In many areas of clinical research, the impact of missing data can be profound Gibbons et al 1993, Laird 1988, Lavori 1992, Little and Rubin 1987. The potential impact of missing data is best understood by considering the process (i.e., mechanisms) leading to the missingness. The following taxonomy of missing data mechanisms is now common in the statistical literature (Little and Rubin 1987)
Data are considered missing completely at random (MCAR) if the missingness does not depend on (is not
Traditional methods
A common choice in many therapeutic areas is to assess mean change from baseline to endpoint via analysis of variance (ANOVA), with missing data imputed by carrying the last observation forward (LOCF). The LOCF approach assumes that missing data are MCAR and that subjects’ responses would have been constant from the last observed value to the endpoint of the trial. These conditions seldom hold (Verbeke and Molenberghs 2000). Carrying observations forward may therefore bias estimates of
Methods
A reanalysis of data from a clinical trial is presented to illustrate the use of MMRM. The results were originally reported by Wernicke et al (1987). The study included patients with a baseline 17-item Hamilton Depression Rating Scale (HAMD17; Hamilton 1960) total score of 19 or more. Patients were randomized to placebo, fluoxetine 20 mg daily (Flx20, n = 100), fluoxetine 40 mg daily (Flx40, n = 103), fluoxetine 60 mg daily (Flx60, n = 105), and placebo (n = 48) in a 2:2:2:1 ratio.
For these
Results
Reasons for discontinuation are summarized in Table 2. As the dose of fluoxetine increased, the percentage of completers decreased, the percentage of patients who discontinued for adverse events increased, and the percentage of patients who discontinued for lack of efficacy decreased. The timing of discontinuation also varied across treatments. The Flx60 group had a higher percentage of subjects dropping out at earlier visits.
Results from analyses using the three methods are summarized in
Discussion
We have noted that no universally best approach to analysis of longitudinal data exists. In general, however, the analyses traditionally used in many longitudinal clinical trials are based on the unrealistic assumption that data are MCAR. The MAR assumption is more plausible than MCAR. Likelihood-based MAR methods can be easily implemented with commercially available software, are consistent with the intent-to-treat principle, all details can be prespecified, and these methods have been shown
Conclusion
No universally best approach to analysis of longitudinal data exists; however, likelihood-based, mixed-effects analyses developed under the MAR framework are more robust to the bias from missing data than LOCF and are valid in every scenario where LOCF is valid, and in many other scenarios as well. Therefore, likelihood-based, repeated-measures analyses are a sensible analytic choice in many clinical trial scenarios. Because the possibility of MNAR data cannot be ruled out, MNAR methods can be
Acknowledgements
This work was sponsored by Eli Lilly and Company.
The authors thank Ms. Renee Bacall for her editorial assistance and the reviewers for their detailed comments. Their contributions greatly enhanced the quality of this manuscript.
Aspects of this work were presented at the conference, “Difficult-to- Treat Depression,” held April 21–22, 2002 in San Francisco, California. The conference was sponsored by the Society of Biological Psychiatry through an unrestricted grant provided by Eli Lilly and
References (21)
- et al.
Using the general linear mixed model to analyze unbalanced repeated measures and longitudinal data
Stat Med
(1997) - et al.
Duloxetine, 60 mg once daily, for major depressive disorderA randomized double-blind placebo-controlled trial
J Clin Psychiatry
(2002) - et al.
Analysis of Longitudinal Data
(1994) - et al.
Informative dropout in longitudinal data analysis
Appl Stat
(1995) - et al.
Some conceptual and statistical issues in analysis of longitudinal psychiatric data
Arch Gen Psychiatry
(1993) - et al.
Duloxetine in the treatment of major depressive disorderA double-blind clinical trial
J Clin Psychiatry
(2002) A rating scale for depression
J Neurol Neurosurg Psychiatry
(1960)- et al.
Statistical handling of dropouts in longitudinal clinical trials
Stat Med
(1992) Missing data in longitudinal studies
Stat Med
(1988)Clinical trials in psychiatryShould protocol deviation censor patient data
Neuropsychopharmacology
(1992)
Cited by (205)
The emotional dance with depression: A longitudinal investigation of OULA® for depression in women
2020, Journal of Bodywork and Movement TherapiesMetabolomic profiles of 38 acylcarnitines in major depressive episodes before and after treatment
2023, Psychological Medicine