Purchase this article with an account.
Deborah Goren, William H Swanson, Shaban Demirel, Stuart Keith Gardiner; Censoring sensitivities results in a non-linear model of sensitivity vs. variability for standard automated perimetry. Invest. Ophthalmol. Vis. Sci. 2014;55(13):5631.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Recently, perimetric sensitivities below ~15-19dB have been shown to be effectively uninformative of true functional status, as they are likely beyond the effective dynamic range of perimetry (Goren, 2013, ARVO). This study examines whether excluding these values alters the nature of the relation between sensitivity and variability.
Thirty-four individuals with moderate to advanced glaucoma were tested at 4 individually chosen locations (35 trials x 7 contrast levels) using size III SAP stimuli on an Octopus perimeter. Frequency of seeing curves were generated to which cumulative Gaussian curves were fit, assuming a maximum potential response probability of 95%, and used to estimate sensitivity and variability for each location tested. The sensitivity-variability relation was modeled using a linear model (Var=m*Sens+b), and a non-linear model (Var=eA-B*Sens) similar to that of Henson et al (IOVS 2000); once including all sensitivities and once excluding values less than 19dB. Akaike’s Information Criterion (AIC) was calculated for each model, first for all available data, and then for 1000 bootstrapped resamplings. AICs were compared using the paired Wilcoxon test.
When including all sensitivities, the linear model produced lower (better) AIC in the entire dataset (AIClin: 668.65; AICnon-lin: 685.96; Fig A), and in 100% of bootstrapped samples (p<0.001). When sensitivities less than 19dB were excluded, AIC for the non-linear model were lower than for the linear model overall (AIClin:212.30; AICnon-lin:211.07; Fig B) and in 72% of bootstrapped samples (p<0.001). Model coefficients from Henson et al (A=3.27; B=0.081) were within the 95% confidence limits of the nonlinear model coefficient estimates when sensitivities <19dB were excluded (A:2.72-4.56; B:0.06-0.14), but this was not true when all sensitivities were included (A:2.29-2.44; B:0.03-0.04).
Inclusion of all sensitivity values suggested a linear model relating sensitivity to variability. However when uninformative sensitivities (<19dB) were excluded, non-linear models fit the sensitivity-variability relation best, consistent with the Henson model. These findings suggest that inclusion of uninformative sensitivities may lead to erroneous conclusions regarding sensitivity-variability relations and that more appropriate conclusions are reached when sensitivities below 19dB are censored.
This PDF is available to Subscribers Only