Analysis of visual field progression in glaucoma using penalized likelihood smoothing

Glaucoma is one of the leading causes of irreversible blindness in the world, the symptoms of which are not apparent until much damage is done to the optic nerve. Hence early detection is necessary and the perimetric visual field test is one way to detect progression early. The high variation in the perimetric results make analysis difficult. In this thesis, prediction of a future visual field is improved by borrowing information from the neighboring locations in the visual field. A penalized log likelihood which accommodates censoring, and with a quadratic penalty is used to get estimates of the slopes for all locations. The amount of smoothing required for each patient is given by the value of a smoothing constant lambda, where the prediction errors measured by a mean squared prediction error criterion is the smallest. With smoothing, a minimum of about 4% and a maximum of about 70% gain was obtained for all individuals. If a fixed common value of the smoothing constant is used for all patients considerable gain is still achieved.