ISOQOL Afternoon Workshops
October 11, 20069. Dealing With Incomplete Longitudinal Data
Desmond Curran, Omega Research Limited, Dublin, Ireland, Herbert Thijs, Center for Statistics, Hasselt University, Diepenbeek, Belgium
In clinical trials, patients are often followed for a specific period of time yielding longitudinal Quality of Life (QoL) data. In these studies the researcher is almost always confronted with the problem of incomplete data and for this reason this workshop is devoted to dealing with missing data. More precisely, we will examine the most widely used methods of longitudinal data analysis and how they are affected by different types of missing Health related QoL (HRQOL) data. The main objective of this workshop will be to present the definitions of Missing Completely at Random (MCAR), Missing at Random (MAR) and Missing Not at Random (NMAR) and to understand how different types of missing data and related assumptions, can influence the results and conclusions obtained from the statistical methods used. With respect to methodology we will start by providing an overview of simple methods and highlight their associated limitations. Basically we will conclude that these methods can no longer be used on their own and consequently we will focus on several methods and approaches which can be used to deal with missing QOL data such as Multiple Imputation using different strategies, Selection models (Local and Global Influence) and Pattern mixture models applied where possible to both continuous and discrete data. The main goal finally should be to convince researchers to treat missing data problems by means of a sensitivity analysis. Participants to the workshop should have some experience with the analysis of longitudinal data using mixed effects models. After completing this workshop, participants should be able to: 1. Understand the limitations of simple imputation techniques. 2. Identify settings where multiple imputation is appropriate. 3. Identify which imputation techniques are suitable for missing data in longitudinal studies. 4. Perform analyses of multiply imputed data.
Level: Advanced
10. Introduction to the Construction of Computer Adaptive Tests on the Basis of Item Response Theory
Otto B. Walter, Psychology, University of Munster, Munster, Germany
So far there is no gold standard of how to construct computer-adaptive diagnostic tests on the basis of item response theory (IRT). This is particularly true in applications of this approach to quality of life measurement. There is rising interest in these methods as computer-adaptive tests open up the opportunity of precise measurement while reducing the item burden placed on patients. However, the variety of models to choose from as well as the mathematical intricacies of this approach can be daunting. A main purpose of the workshop is to provide insight into the construction and mathematical background of computer-adaptive tests by reviewing and comparing the methods that have been applied to the construction of IRT based instruments. Commonly employed IRT models are discussed, with an emphasis on polytomous models. The workshop outline includes topics such as data collection, data preparation, testing for unidimensionality, parameter estimation, evaluation of model fit, item information, and latent trait computation. Illustrative analyses of data sets will be provided throughout the workshop. The workshop will be divided in lecture, exercises and discussion.
Level: Advanced
11. Advanced Psychometric Methods: Application In Pro Instrument Devleopment And Evaluation
Dennis Revicki, Donald Stull, Center for Health Outcomes Research, United Biosource Corp., Bethesda, MD
The development and psychometric evaluation of PRO instruments requires the application of a number of different techniques, including exploratory and confirmatory factor analysis (FA), item response theory analysis, and structural equation modeling (SEM). We will provide a brief overview of psychometric analyses and will then focus on the application of (1) exploratory and confirmatory factor analysis for understanding of new measures and (2) use of SEM for testing construct validity. Exploratory and confirmatory FA can be used to examine the relationships among items with a PRO measure or among different domains or multiple PRO measures. These techniques are useful for understanding the internal structure of PRO instruments and for understanding construct validity. The workshop will describe the main methods of FA and illustrate these methods with examples from the instrument development literature. SEM is a powerful analytic technique that combines FA and path analysis in a simultaneous, confirmatory approach. Using SEM, the researcher can specify and evaluate hypothesized relationships between observed and latent (unobserved) constructs as well as relationships among the latent variables. SEM can also estimate the reliability and validity of measurement models while explicitly modeling measurement error. A researcher specifies a measurement model and a structural model which specifies relationships among the latent variables to examine construct and criterion-related validity. If the observed covariances are consistent with the model-implied covariances, the researcher has evidence supporting the construct validity of the PRO measure. This workshop will demonstrate the main methods, testing assumptions and criteria, and provide examples to illustrate the methods of SEM.
Level: Advanced
12. Evaluating Change In Patient-Reported Outcomes
Kathleen W. Wyrwich, Research Methodology and Health Services Research, Saint Louis University, St. Louis, MO
Although numerous measures have been developed for the evaluation of health-related quality of life (HRQoL), strategies for identifying meaningful intra-individual and group change in these measures have not kept pace with instrument development. As a result, clinical trial researchers, quality assurance assessment teams, practicing clinicians, and patients are without established standards to evaluate change in HRQoL measures. This course will review, critique and compare the methods that have been applied to establish intra-individual and group HRQoL change standards, which include anchor- and distribution-based techniques. Practical approaches to improving and advancing HRQoL change evaluations that enhance the interpretation of intra-individual, as well as a review of controversies that have developed will be provided. In addition, the course will explore future qualitative and quantitative challenges in this area of HRQoL research. The workshop outline includes: 1. Who are the stakeholders in HRQoL change evaluations?; 2. Review and critique of evaluation methods to date (Anchor-based and Distribution-based); 3. Relationships between evaluation methods; 4. Controversies associated with these methods; 5. Practical approaches; 6. Challenges ahead; and 7. Additional questions and discussion. The workshop will be divided between 70% lecture, 10% class exercises, and 20% discussion and answer periods. Participants are strongly encourages to bring a hand calculator.
Level: Advanced
13. Assessing QoL in Palliative Care
Stein Kaasa, Cancer Reserach and Molecular Medicine, NTNU, Trondheim, Norway, SB Detmar, Prevention and Health, TNO, Leiden, Holland, Jon Haarvard Loge
The principal goal of palliative treatment is to prevent and relieve symptoms and to rehabilitate patients if possible. Patients in a palliative stage of their disease have several symptoms of physical, psychological and social character simultaneously and often also existential challenges. Palliative care research is needed to test effectiveness of various types of interventions, and also to increase the basic (fundamental) understanding of physical, psychological, social and spiritually problems. The primary outcome in palliative care research is almost always of subjective nature, i.e. quality of life related issues and valid outcomes are needed. There are several challenges related to the use of subjective outcomes, such as: * Which set of outcomes should be used in the specific study? * The lack of international consensus on outcomes * Missing data in prospective studies * The handling of several outcomes in small studies * The lack of clinical intuitive understanding of health related quality of life outcomes These issues will be discussed in the workshop together with: * Study design, how to plan studies and how to conduct studies * How to choose outcomes in studies * How to implement new data into clinical practice * How to perform QL assessments in palliative treatment
Level: Basic
14. Quantifying Health States by Thurstone Scaling
Paul F.M. Krabbe, Medical Technology Assessment, Radboud University Nijmegen Medical Centre, Nijmegen, The Netherlands
The goal of all health care services activities and programs is to improve or sustain the health of people. Thus, it is not surprising that, over the years, there has been considerable interest and activity in developing methodologies to quantitatively measure the overall health status of patients and populations. So far, the present methodology has been dominated by theories and measurement techniques from (health) economists. Apart from the fact that measurement techniques based on economic theory are not very practical to conduct, numerous empirical studies have shown that elicited values (utilities) are affected by several biases and axiomatic violations. Scaling models developed by psychometricians and others are a combination of simple measurement tasks and specific data analysis. The attractiveness of the use of scaling models (e.g. Thurstone scaling, multi-dimensional scaling, Rasch model) in the case of quantifying health states (e.g. values, utilities, strength of preferences) is based on uncomplicated and cognitively simple judgment tasks (ranking, choices) that guarantee response data of good quality. These data provide enough information, after additional analytical computations, to arrive at single metric measures for health states at a group level. Although scaling models have been applied with considerable success in other research areas, they have hardly been explored nor applied in the field of medicine. In this workshop the first developed, and still most important scaling model, namely Thurstone scaling, will be explained. Outline of the workshop: I) Introduction of the conceptual ideas behind scaling models and their relationships to measurement theory. II) Explenation of Thurstone scaling and a step by step demonstration of this scaling model. III) Hands on experience with Thurstone scaling in a class exercise (attendants' responses will be used). IV) Early results from our own empirical studies will be presented.
Level: Basic
15. Prognostic Assessment of Baseline HRQOL in Oncology
Corneel Coens, Data Center, EORTC, Brussels, Belgium
Prognostic factor analyses are used in oncology to identify variables that are independent predictors of outcome. Since the advent of methods for measuring health-related quality of life, several studies have been published in which QoL variables have been identified as important prognostic factors in addition to clinical factors. This finding has considerable importance, particularly in advanced disease where treatment is generally palliative and the aim is to optimize QoL.
However, due to the specific nature of QoL data, classical analysis techniques are not always appropriate and might lead to parameter estimates of incorrect magnitude, incorrect sign etc. This workshop aims to give an overview of the issues related to the assessment of the prognostic value of baseline QoL in oncology trials and to propose some practical recommendations to circumvent these problems. Examples will be drawn from application of the discussed techniques on an existing data set.
Workshop outline:
- Introduction
- Definitions: prognostic/predictive/risk groups
- Prognostic models
- Classic methodology
- QoL specific issues
- Multiplicity
- Multi-collinearity
- Missing data
- Analysis plan
- Variable selection
- Building the model
- Checking the model
- Validation
- Internal validation
- External validation
- Bootstrap: variable selection and model stability