Modelling Visual Acuity Distributions to Inform a Simplified Cost Effectiveness Analysis

OBJECTIVES: Clinical trials in ophthalmology frequently focus on the mean change in visual acuity from baseline, however economic models submitted to HTA bodies have previously used a Markov structure and used health states to group visual acuity scores and capture the change in vision. Accurately capturing the distribution of patients rather than focusing on the mean is essential when costs and quality-of-life express a non-linear relationship with vision. Previous research in migraine has utilised parametric models to estimate the distribution around the mean to allow simplified modelling of the mean and reduce the number of health states, incorporating summary data and facilitate extrapolation. The objective of this study was to fit parametric models to visual acuity to inform the development of simplified cost effectiveness models.

METHODS: Publicly available 5-year patient-level data from the multicentre randomized Protocol T clinical trial for diabetic macular oedema (DMO) patients was used to fit negative binomial, beta binomial, and Poisson distributions to EDTRS scores. The analysis fitted separate distributions at each study visit and for each eye. As the aim of the research was to fit a robust distribution to the data, treatment arm was not factored into the analysis however assessment of a multimodal distribution were considered. Distributions were considered through goodness-of-fit statistics and visual inspection.

RESULTS: The Poisson distribution presented the best fitting curve for the baseline data, however over time this changes to a negative binomial distribution. Result sensitivity was tested through data transformations, visual acuity score groupings, logMAR scores, and separate correlation between eyes, however results remained broadly consistent across all scenario analyses.

CONCLUSIONS: Visual acuity data from DMO patients presents a challenge to fitting robust distributions due to the heavily skewed distribution. This study shows that this approach still offers a valuable modelling option when no patient-level data is available.