Effect of spatial correlation in detecting changes in the optic nerve head with confocal scanning laser tomography

Purpose. To assess the effect of spatial correlation of topography values on the ability to detect change. Methods. Tests for differences between the baseline and follow-up images obtained with the Heidelberg Retina Tomograph were carried out on each non-overlapping condensed pixel. The image matrix contained 64 x 64 condensed pixels (each condensed pixel = 4 x 4 original pixels). The method uses a three-way analysis of variance for each condensed pixel with factors for time (baseline, follow-up), location (within the condensed pixel) and also for image nested within time. Change is said to have occurred if the interaction of time and location or the main effect for time are significant. We simulated differing degrees of spatial correlation within a condensed pixel by varying the intraclass correlation coefficient, r. The simulations were carried out assuming no change between baseline and follow-up. We then applied the test method previously described (ARVO 1994). We also consider an alternative method based on a Satterthwaite correction which adjusts the degrees of freedom of the F- ratio from the original test. Results. Assuming alpha = 0.05, we would expect 5% false positives. Our results show that the original test performed well in the absence of spatial correlation, however, with increasing degrees of spatial correlation (as found in real images), the percentage of false positives increased significantly with 4.67%, 5.95%, 12.89% and 19.86% false positives for r = 0, 0.1, 0.4 and 0.8 respectively. However with the Satterthwaite correction the corresponding false positives were 4.21%, 3.2%, 2.04%, 4.11% respectively. Conclusions. In the presence of spatial correlation, the Satterthwaite correction reduces the number of false positives.