Maximum likelihood perimetric progression analysis: Using raw (trial-by-trial) response data to estimate progression more robustly

Abstract

Purpose: To describe and demonstrate a more efficient (Maximum Likelihood) method for quantifying visual field progression. Design: Monte Carlo simulation. Methods: Trial-by-trial response data were simulated using a stochastic psychometric model (a simulated observer ). Simulated Differential Light Sensitivity (DLS) decreased between tests to mimic long-term visual field progression. Progression slopes were fitted, either by fitting a regression slope to independent DLS estimates from each test (conventional method), or by fitting all the raw data combined in a single model (proposed maximum likelihood method). Results: The proposed ML method seldom performed worse than a conventional, regression-based approach, and often performed better. For an idealized observer with a lapse (false negative) rate of 0 and a guess (false positive) rate of 0, both methods were equally precise. However, as lapse rate increased, the ML method exhibited less random measurement error. For small numbers of trials this increase in precision translated to a negative progression slope being detected with 95% confidence at least one year/assessment sooner. The only time the ML method was observed to perform worse was when very few trials (N = 4) were combined with very high lapse rates ({lambda} = 0.3): an unlikely but not inconceivable scenario. Conclusions: Combining raw, trial-by-trial response data in a single ML model can provide a more robust estimate of visual field progression than conventional methods (e.g., linear regression), at no additional cost to the patient or clinician (i.e., no additional trials).