Abstract
Background: Clinical studies often omit outcomes that allow the direct estimation of Quality Adjusted Life Years (QALYs) for use in cost effectiveness analyses crucial for informing policy decisions. This means analysts often have to estimate the relationship, “map”, between included outcomes and preference based ones like the EuroQoL EQ5D. In ankylosing spondylitis (AS), the relationship between BASDAI/BASFI and EQ5D has been established but the emergence and growing use of the Ankylosing Spondylitis Disease Activity Score (ASDAS), a new composite index used to assess clinical disease activity, means new mapping tools are required. Furthermore mapping has never been done for the entire axial spondyloarthritis (axSpA) spectrum of patients i.e. including not only patients with AS but also those with non-radiographic axSpA.
Objectives: To estimate a robust mapping between ASDAS and EQ5D (3 level version) and to test its performance out of sample (external validation) in patients with axSpA.
Methods: Data from an electronic, prospective, nationwide Rheumatic Disease Portuguese Register (Reuma.pt) provided data pertaining to 1140 patients (5483 observations) with a confirmed diagnosis of axSpA was used to develop a model to predict EQ5D from the ASDAS score. We compared a range of different statistical models developed to deal with the complex distributional features of health utility data. A range of criteria examining model fit across the spectrum of disease severity were used to select preferred models. A smaller dataset for out of sample validation from the SPondyloArthritis Caught Early (SPACE) cohort was used, providing data from 317 patients (1225 observations) at five European centres.
Results: Characteristics of patients from the Reuma.pt and SPACE are presented in the table. There is a non-linear relationship between ASDAS and EQ5D. We found that a four component mixture model based on a bespoke distribution, with one component constrained to reflect the mass of observations at full health, was the best fitting of the ASDAS models estimated (figure). ASDAS squared and age squared featured as within component variables. The model demonstrated close fit to the observed data where ASDAS was less than 4 but diverged from the mean of the data where ASDAS was higher. There is a very limited data at this more severe level of disease activity. In the out of sample testing, the model continued to perform well overall and exhibited the same divergence from the observed data only where data was sparse.
Conclusion: There is a clear relationship between ASDAS and EQ5D that we were able to model reliably using bespoke mixture model based methods. There is more uncertainty regarding model fit at very high levels of disease activity owing to the relative paucity of data from patients in such disease activity state. Future analyses may wish to focus on these severely affected patients in order to improve the robustness of model estimates.
Objectives: To estimate a robust mapping between ASDAS and EQ5D (3 level version) and to test its performance out of sample (external validation) in patients with axSpA.
Methods: Data from an electronic, prospective, nationwide Rheumatic Disease Portuguese Register (Reuma.pt) provided data pertaining to 1140 patients (5483 observations) with a confirmed diagnosis of axSpA was used to develop a model to predict EQ5D from the ASDAS score. We compared a range of different statistical models developed to deal with the complex distributional features of health utility data. A range of criteria examining model fit across the spectrum of disease severity were used to select preferred models. A smaller dataset for out of sample validation from the SPondyloArthritis Caught Early (SPACE) cohort was used, providing data from 317 patients (1225 observations) at five European centres.
Results: Characteristics of patients from the Reuma.pt and SPACE are presented in the table. There is a non-linear relationship between ASDAS and EQ5D. We found that a four component mixture model based on a bespoke distribution, with one component constrained to reflect the mass of observations at full health, was the best fitting of the ASDAS models estimated (figure). ASDAS squared and age squared featured as within component variables. The model demonstrated close fit to the observed data where ASDAS was less than 4 but diverged from the mean of the data where ASDAS was higher. There is a very limited data at this more severe level of disease activity. In the out of sample testing, the model continued to perform well overall and exhibited the same divergence from the observed data only where data was sparse.
Conclusion: There is a clear relationship between ASDAS and EQ5D that we were able to model reliably using bespoke mixture model based methods. There is more uncertainty regarding model fit at very high levels of disease activity owing to the relative paucity of data from patients in such disease activity state. Future analyses may wish to focus on these severely affected patients in order to improve the robustness of model estimates.
| Original language | English |
|---|---|
| Article number | OP0078 |
| Pages (from-to) | 52-53 |
| Number of pages | 2 |
| Journal | Annals of the Rheumatic Diseases |
| Volume | 79 |
| Issue number | Suppl 1 |
| Early online date | 2 Jun 2020 |
| DOIs | |
| Publication status | Published - 2 Jun 2020 |