IRT-BASED ASSESSEMENT OF DIFFERENTIAL ITEM FUNCTIONING IN HRQL STUDIES
Christophe Lalanne, Martin Duracinsky, Andrew R. Armstrong, Olivier Chassany 
Department of Clinical Research and Development, Assistance Publique-Hôpitaux
de Paris, Paris, France

AIMS. 
We sought to provide sample-free estimates of uniform DIF on a newly developed
HRQL questionnaire, PROQOL-HIV. We considered patients characteristics
supposed to modulate HRQL, such as gender, country of residence but also
ethnicity, given the high prevalence of HIV in migrant populations.

METHODS.
We used an iterative ordinal logistic regression framework where nested models
are compared in order to test for significance of person covariates, while
persons parameters are estimated using the Graded Response Model. DIF was
tested on two constructs extracted from factor analysis: physical health and
symptoms (PHS, 14 items accounting for 17% of the total variance) and
emotional distress (ED, 10 items, 9%), on a subset of N=505 patients (71%
males, median age 43 yrs.) from 5 countries (Australia, USA, France, Brazil,
Thailand). Missing responses (<1%) were imputed using median scale score.

RESULTS. 
Items belonging to the ED dimension were more frequently flagged as exhibiting
bias compared to PHS items, especially when referring to ethnicity.
Country-related biases were equally distributed on the two scales. In both
cases, however, no DIF effects were found for gender. We reached similar
conclusions when comparing the results to those obtained when using sum or
rest scores as conditioning ability levels. Correlations between scores
computed with or without flagged items were above 0.5. Restricting analysis on
Western countries, however, yielded smaller hits. This suggests that, for
these two dimensions of PROQOL-HIV, DIF effects should have a small to
moderate influence on HRQL scores reporting, although more studies are needed
to estimate DIF effect size and provide cut-off values specific to the
questionnaire.

CONCLUSIONS.
The relevance of DIF effects need to be carefully studied through expert
reviewing, and their higher frequency on bio-psychological constructs raise
interesting questions for questionnaire development. Moreover, DIF can bring
additional information about HRQL when testing also ethnic origin and not only
country of residence in international clinical trials.