Psychometric methods for health outcomes II (04)

Chair: Sik-Yum Lee, Thursday 23rd July, 9.40 - 11.00, Dirac Room, Fisher Building. 

Wilco Emons, Department of Methodology and Statistics, Tilburg University, The Netherlands. On the combined use of parametric and non-parametric IRT models for polytomous items with applications to scale analysis in clinical and medical psychology. (228)

Ulrich Reininghaus, Tom Burns, Rosemarie McCabe, Stefan Priebe  Academic Unit for Social and Community Psychiatry, Queen Mary College, London, UK, Tim Croudace, Department of Psychiatry, University of Cambridge, UK. Identifying differential item functioning and population heterogeneity in the measurement of patient reported outcomes in psychosis. (252)

Elizabeth Savoca, Smith College, Northampton MA, USA. Accounting for Misclassification Bias in Binary Outcome Measures of Illness: The Case of Posttraumatic Stress Disorder in Male Veterans. (070)

Jon Heron, Carol Joinson, Glyn Lewis, Ricardo Araya, ALSPAC, University of Bristol, UK, Tim Croudace, Department of Psychiatry, University of Cambridge, UK. Pubertal timing and the onset of depressive symptoms in a cohort of girls from ALSPAC. (249)

ABSTRACTS 

On the combined use of parametric and non-parametric IRT models for polytomous items with applications to scale analysis in clinical and medical psychology. (228)
Wilco Emons
Item response theory (IRT) models are increasingly used in the domain of clinical and medical psychology. Applications include scale construction, scale revision, linking, item banking, and meta analyses. Most studies in these fields involve scales comprising polytomous items (e.g., Likert items). For these scales, the graded response model (GRM) is often used for IRT modeling. To draw valid conclusions, however, it must be ascertained whether the assumptions underlying the GRM hold in the data. This is of paramount importance in the context of clinical and medical psychology, because clinical scales tend to have characteristics that violate the assumptions made in estimating and applying the GRM. The current study investigates how nonparametric IRT approaches can be used as complementary diagnostic tools for evaluating the goodness-of-fit of the GRM. Model fit assessment by means of nonparametric IRT approaches will be addressed at the scale level, the item level, and person level. Results are presented for simulated data and real data on questionnaires used in the field of clinical and medical psychology.

Identifying differential item functioning and population heterogeneity in the measurement of patient reported outcomes in psychosis. (252)
Ulrich Reininghaus, Tom Burns, Rosemarie McCabe, Stefan Priebe and Tim Croudace
Patient reported outcomes (PROs) increasingly important in the evaluation of treatment effects in psychosis. However, some authors contest the use of self-report instruments in this population. While there is evidence suggesting that PROs may be influenced by a number of socio-demographic and clinical factors such as age, ethnicity, or psychiatric symptoms, little is known about differential item functioning in existing scales with regard to these factors. Given the increasing use and popularity of PROs in psychosis research, there is a need to assess whether existing scales are biased against particular subgroups of patients, e.g. with different depressive or psychotic symptom levels, from different ethnic backgrounds, or of different ages. The current research therefore aims to identify differential item functioning and population heterogeneity of existing PRO scales using multiple indicator multiple cause (MIMIC) models in MPlus, Version 5.2. MIMIC models are applied to large sets of patient data from European studies assessing PROs in psychosis on the basis of widely used scales (e.g. Lancashire Quality of Life Profile, Camberwell Assessment of Needs, Patient Satisfaction Questionnaire). The use of MIMIC models in this applied research context will be illustrated and preliminary findings presented and discussed.

Accounting for Misclassification Bias in Binary Outcome Measures of Illness: The Case of Posttraumatic Stress Disorder in Male Veterans. (070)
Elizabeth Savoca
Empirical analyses of the determinants of adult health frequently rely on large-scale epidemiological studies of the health of the general population, where illness is often measured by a dichotomous indicator of the respondent’s reported presence or absence of a diagnosis or condition. A large body of research has documented substantial response errors in survey classifications of both physical and mental impairments, even among survey instruments that are designed to simulate clinical appraisals. Classification errors in survey health indicators can have serious consequences not only for prevalence estimates but also for inferences about measures of association. This principle is well-known among biostatisticians and epidemiologists, but has yet to make a significant impression on empirical research in other disciplines. This paper uses a maximum likelihood approach to account for classification errors in an outcome variable in an analysis of the antecedents of Posttraumatic Stress Disorder in veterans. Results show that when adjusted for errors in diagnoses, the sample PTSD prevalence estimate falls significantly; that failure to correct for misclassification in PTSD dramatically understates the effects of risk factors, both war-related and not war-related; and that the downward bias remains even when the model incorporates differential classification errors.

Pubertal timing and the onset of depressive symptoms in a cohort of girls from ALSPAC. (249)
Jon Heron, Carol Joinson, Glyn Lewis, Ricardo Araya and Tim Croudace
ALSPAC, the Avon Longitudinal Study of Parents and Children, is a UK-based birth-cohort study set up in the early 90’s. Data on young people has been collected regularly since birth via questionnaires and hands-on clinic assessment. The work presented will focus on three measures of the short form (13 items) of the MFQ (Moods and Feelings Questionnaire), which was administered during the clinic assessments at ages 10.5, 13 and 14 years. Age at onset of menarche will be used as a measure of pubertal timing and hence this presentation will be restricted to the girls in the cohort. The effect of early life adversities will also be examined. Split-half scales derived from the MFQ items have been modelled using a variety of latent variable models. In initial work, a second-order growth model permits the assessment of the effect of early pubertal onset on both baseline levels of symptoms and also change across the time period. The results will be compared with models more suited for the skewed nature of the manifest measures such as semi-continuous growth models or the incorporation of a mixture component.