Analyzing Patient Reported Outcomes in Exercise Oncolgy using Longitudinal Latent Variable Models

Specialeforsvar ved Gry Nielsen Sparre

Title: Analyzing Patient Reported Outcomes in Exercise Oncolgy using Longitudinal Latent Variable Models 

Abstract: In recent years there has been a growing interest in the exercise oncology field, where physical activity has been used as a strategy for rehabilitation in cancer patients to remedy disease and treatment related symptoms and side effects. Patients with acute leukemia experience a substantial symptom burden and are at risk of developing infections throughout the course of treatment consisting of repeated cycles of intensive chemotherapy. To date, there are no clinical practice exercise guidelines for patients with acute leukemia undergoing the first parts of chemotherapy. Patients with prostate cancer is treated with Androgen Deprivation Therapy (ADT), but adverse musculoskeletal and cardiovascular effects of ADT are widely reported. In this thesis, we will study four data sets from two studies. For the 'FC Prostate' study we use Role Emotional functioning data and Fatigue data, for the PACE-AL study we use Patient Activation Measure (PAM) and General Self-Efficacy (GSE) data. We examine the longitudinal person reported outcomes of these questionnaires concerning their health status. The patients are divided into two groups, A or B. To examine the data we use Item Response Theory models and to see if there is an effect of exercise treatment we use different kinds of Item Response Theory Models, as well as latent regression models. We will show how to estimate the IRT models in theory, but also give applied examples in R using different packages. Through the analysis of the data sets we found that there were no significant difference between the groups for the Role Emotional data, but a single patient experienced an increase in the level of health over time. The model used for the Role Emotional Functioning did not fulfill the assumption about local independence, so we cannot trust this result. For the Fatigue questionnaire we found only one patient in group A, to have an increase in the level of fatigue over time, whereas two patients had a decrease of the level of fatigue, one from each group. The Divide-by-total model and the Difference model led to the same conclusion, namely that there were no significant difference between the groups. The best model for the Patient Activation Measure data did not fulfill the local independence assumption but we solved this by removing two items from the analysis. None of the patients had a significant change in their level of health over time. The analysis showed no significant difference between the groups. For the General Self-Efficacy data we found a significant difference between the groups for the Divide-by-total models, and we found that 4 patients had a decrease in level of self-efficacy, two in each group, while 14 patients had an increase in the level of self-efficacy over time, with 13 of those in group A. Since the model did not fulfill the assumptions about local independence, we cannot trust the results. We compared the packages used for the IRT analysis and we found the TAM package to be the most reliable, stable and versatile package. For further research we suggest implementation of the two-dimensional Graded Response Model in R, in order to be able to compare the two types of IRT models. We also suggest a theoretical boundary of the Q3 statistic, since now we only have a rule of thumb, and this might not hold for smaller data sets

Vejledere:    Susanne Ditlevsen, Karl Bang Christensen, SUND
Censor:        Per Bruun Brockhoff, DTU