The analysis of a patient-reported outcome measure in on-demand medication data

Abstract

This thesis deals with advanced statistical methods for the analysis of a patient-reported outcome measure in on-demand medication data. On-demand medication is only taken when necessary to relieve symptoms when they occur, such as pain. On-demand medication data often consist of observations of discrete events of when the medication was taken, including data on symptom severity for each event per patient. The data central to this thesis come from a clinical trial that investigated the effect of an on-demand medical treatment. During this trial, patients experienced events and each time they had an event, they filled out a questionnaire (patient-reported outcome) consisting of five items. These observed items intend to measure a latent construct. The resulting data of this clinical trial are longitudinal, multivariate correlated, multilevel data where events (level 1) are nested within patients (level 2) and where the correlated items of the patient-reported outcome intend to measure a latent outcome variable. The motivation of this thesis was based on the way such data are often analyzed, namely by aggregating all data at the level of the patient. This means that for each event, the sum score of the five items is calculated to approximate the latent variable score and this sum score is then averaged over the events, per patient. This procedure creates patient-level scores and these can be analyzed by conventional statistical techniques. Although easily understood by clinicians, this procedure has several statistical drawbacks that are discussed in this thesis, and alternative, advanced statistical methods are proposed that can overcome these drawbacks. In three chapters, multilevel models and multilevel structural equation models are presented and it is shown how to appropriately analyze the data using these multilevel techniques. These models acknowledge the hierarchical data structure and are able to estimate latent variables. In this thesis, it is advocated to use these advanced statistical methods over conventional statistical models as it is shown that multilevel structural equation models produce unbiased latent variable estimates, are more flexible and offer additional modeling options, compared to conventional procedures. When dealing with similar data, the presented models should be preferred. This thesis provides clinicians with guidelines in how to apply multilevel structural equation models. This final chapter of this thesis presents an overview of the literature describing methods for assessing and/or controlling placebo effects. The motivation for this chapter was based on the occurrence of placebo effects in the clinical trial that is central to this thesis. Questions that are answered is how to disentangle placebo effects from the true treatment effect, how to identify placebo responders and how to minimize placebo effects when designing a clinical trial.