Psychometric methods for health outcomes I (03)

Chair: Christophe Lalanne, Wednesday 22nd July, 11.45 - 13.05, Boys Smith Room, Fisher Building. 

Christophe Lalanne, Martin Duracinsky, Catherine Acquadro and Olivier Chassany, INSERM U669 and Department of Clinical Research, Hospital Saint-Louis, Paris, France. Descriptive and explanatory IRT modeling of a new Quality of Life questionnaire specific for HIV patients. (227)

Myriam Blanchin, Jean-Benoit Hardouin, Tanguy Le Neel, Gildas Kubis and Véronique Sébille  Faculte de Pharmacie, Universite de Nantes, France. Comparison of three methods for the analysis of longitudinal  patient reported outcomes. (009)

Celestina Barbosa-Leiker, Bruce R. Wright, G. Leonard Burns, Craig D. Parks and Paul Strand, Health and Wellness Services, Washington State University, USA. Longitudinal measurement invariance and latent growth modeling applied to the metabolic syndrome. (053)

Jean-Benoit Hardouin, T Le Néel, G Kubis, V Sébille, Faculty of Pharmaceutical Sciences, University of Nantes, France and B Fallissard, INSERM U669, Paris, France. Power with CTT and IRT based approached for the comparison of two groups of patients. (196)

ABSTRACTS 

Descriptive and explanatory IRT modeling of a new Quality of Life questionnaire specific for HIV patients. (227)
Christophe Lalanne, Martin Duracinsky, Catherine Acquadro and Olivier Chassany
PROGOL-HIV is a new Health-related Quality of life (QoL) questionnaire specific for HIV patients. Multicentric pretesting of items was carried out in 8 countries while collecting sociodemographic and biomedical data on patients. Psychometric validation proceeded from analysis of inter-items correlations through Exploratory Factor Analysis which isolates eight main dimensions explaining more than 60% of responses variance to a set of 39 Lickert-type items. Convergent validity was assessed using multi-trait scaling analysis and correlation of QoL scores with MOS-HIV questionnaire already in use. Subgroup analyses indicate that scaled QoL scores are also sensible to biomedical markers, with number of symptoms, depression and CD4 concentration having larger effect size. Rasch-based models can be expressed as generalized linear mixed models which allows to adjust items parameters estimates to country-dependent variations in average QoL level, while providing a way to test for differential item functioning (DIF). Influence of person-level covariates was evaluated on the physical health and symptoms subscale. Responses were also analyzed as polytomous (partial credit vs. rating scale model), providing additional insights on country-specific respondents' behavior. This analysis helped to validate the preliminary cultural validation study and yielded more reliable estimates of QoL levels both at the individual and country level.

Comparison of three methods for the analysis of longitudinal  patient reported outcomes. (009)
Myriam Blanchin, Jean-Benoit Hardouin, Tanguy Le Neel, Gildas Kubis and Véronique Sébille
In health sciences, it is frequent to deal with Patient Reported Outcomes (PRO) data. Two approaches are commonly used to analyse this type of data: Classical Test Theory (CTT) and Item Response Theory (IRT). Moreover, longitudinal data are often collected to allow analysing the evolution of an outcome over time. In this case, the correlation between measurements from each patient over time is important to consider. The most adequate strategy to analyse longitudinal latent variables, which can be either based on CTT or IRT, remains to be identified. In this work, our aim was to compare three methods to analyse such longitudinal latent variables through a simulation study. Rasch Mixed Models (RM) and Longitudinal Rasch Model (LRM) methods are based on the latent variable (IRT) whereas Score Mixed Models (SM) method is based on the score (classical approach). The three methods have shown comparable results in term of type I error (close to 5%). LRM and SM methods presented comparable power and unbiased time effect estimations. The RM method presented lower power than SM and LRM and biased time effect estimations when a time effect was simulated. This method seems inadequate to analyse longitudinal PRO data and should be avoided.

Longitudinal measurement invariance and latent growth modeling applied to the metabolic syndrome. (053)
Celestina Barbosa-Leiker, Bruce R. Wright, G. Leonard Burns, Craig D. Parks and Paul Strand
The primary aims of this research were to examine the development of the Metabolic Syndrome (MetS) over 6 years via second-order latent growth modeling and to examine the relationship between perceived stress and the MetS. Although the factor structure of the MetS has been demonstrated through confirmatory factor analysis, the establishment of longitudinal measurement stability of the MetS construct has not been verified. Utilizing the Spokane Heart Study, this research examined the longitudinal measurement invariance of the MetS indicators (glucose, body mass index (BMI), high-density lipoprotein, triglycerides, and diastolic blood pressure) as specified by a 1-factor model. Results indicated that longitudinal measurement invariance failed at the test of equal intercepts, implying that predicted values of the MetS indicators may differ across time when there is a constant level of the MetS. Thus, this research modeled two primary indicators of the MetS, glucose and BMI, and also examined if between-person variation could be explained by age, sex and perceived stress. Results indicated that glucose was best explained by a linear growth model, whereas BMI linearly increased and then leveled off. Age and sex predicted the intercept variance in glucose, while sex and perceived stress predicted intercept variation in BMI.

Power with CTT and IRT based approached for the comparison of two groups of patients. (196)
Jean-Benoit Hardouin, T Le Néel, G Kubis, V Sébille and B Fallissard
Perceived health measures are increasingly used in clinical trials. These measures differ from other measurements because such patient’s characteristics cannot be directly observed and measured. They are evaluated using self-assessment questionnaires and referred to as Patients-Reported Outcomes (PRO). Two main types of analytic strategies can be used for such data: classical test theory (CTT) and models of Item Response Theory (IRT). The choice of the statistical strategy for PRO data analysis seems to be driven to date by the researchers’ practice and familiarity with one approach or another. It might be hypothesised that IRT should provide a better and more powerful strategy than CTT to detect clinically meaningful effects. Nevertheless, whether IRT or CTT based analysis can be considered equivalent regarding power remains unknown. The purpose of our work was therefore to study and compare the statistical properties of CTT and IRT by a simulations study regarding power. Without any missing data, IRT and CTT seems to provide comparable power when comparing two groups of patients, whatever the number of individuals, items in the questionnaire or value of the effect size. Nevertheless, the observed power obtained with the two approaches is lower than the one computed by using classical formulas.