Logo: to the web site of Uppsala University

Introduction: Ataxia comprises a group of neurological disorders, among which spinocerebellar ataxia type 3 (SCA3) is one of the most common autosomal dominant forms. The Scale for the Assessment and Rating of Ataxia (SARA) is a widely used tool for evaluating the ataxia severity, and has been adopted into several modified versions, such as Modified Scale for the Assessment and Rating of Ataxia (m-SARA) and Functional Scale for the Assessment and Rating of Ataxia (f-SARA). Item Response Theory (IRT) has been applied in ataxia assessment based on the SARA and has demonstrated good performance in capturing item-level measurement characteristics. This thesis applies the IRT modeling to the SCA3 population to assess the disease severity using different versions of SARA and explores applicability of the published SARA IRT models in SCA3 population.

Methods: The dataset from the SCA3 population, with 352 patients with 1168 visits, was analyzed using IRT modeling. A graded response model was applied to assess the disease severity and was compared with a published IRT model developed in autosomal recessive cerebellar ataxia (ARCA) population. Model performances were evaluated using item characteristics curves, Fisher information and individual latent variable estimates. The longitudinal IRT models were developed to capture the disease progression and further assess the published model’s transferability. The IRT model was applied on different versions of SARA, and comparisons were made based on test information, individual latent variable estimates and IRT-informed functions.  

Results: The IRT model developed on the SCA3 population dataset demonstrated good performance. The re-estimated IRT model showed improved model fit compared to the published SARA IRT model. However, individual latent variable estimates from cross-sectional and longitudinal IRT models showed similar results from the two models. IRT models were developed for six versions of SARA, and test information analysis indicated that SARA provided the most amount of information, followed by the m-SARA and then f-SARA. IRT-informed functions supported these findings in terms of measurement sensitivity.

Conclusions: While the re-estimated IRT model showed improved fit and reduced uncertainty, the published SARA IRT model achieved comparable results in individual latent variable estimation and longitudinal model estimates. Although variations in measurement performance were observed among the different versions of SARA, SARA showed the best performance.