Dept. of Biomedical Engineering
Oregon Health & Science University
Neuropsychological test scores provide a valuable tool for evaluating cognitive function and identifying cognitive decline. Unfortunately floor and ceiling effects, the methods for determining summary score information, and subject dropout (known as longitudinal censoring) are all drawbacks that inhibit the value of these tests for evaluating cognitive performance. By developing models which address these drawbacks, better estimates of cognitive performance can be obtained from score data. Earlier diagnosis, and possibly treatment, of cognitive decline may be possible with these improved score estimates. In this thesis we utilize censored normal (Type 1 Tobit) models for longitudinal score data subject to both ceiling and floor effects. Evaluation of ceiling-afflicted data is done on the Boston Naming Test (BNT) while evaluation of floor-afflicted data is done on the Word List Delayed Recall (WLDR) test. Simulations show that failing to account for ceiling effects results in improper estimation of change points as well as population decline estimates that are significantly different than true values. Simulation shows that in longer studies failing to account for informative dropout results in an overestimation of population mean and an underestimation of population variance. Prediction of scores at the fourth follow-up visit as well as the ability of models to classify subjects with Mild Cognitive Impairment (MCI) were evaluated for both BNT and WLDR using standard normal and Type 1 Tobit models. In the BNT, a model with quadratic decline with respect to time resulted in a classifier with an area under the Receiver Operator Characteristic (ROC) curve of 0.73 for the Tobit model, and 0.69 for standard normal, suggesting slight improvement in classification of cognitive impairment when accounting for the ceiling. Mean squared error of predicted fourth follow-up score values was 7.95 for the Tobit models and 8.05 for standard normal models. In the Word List Delayed Recall, a more widely utilized test with more data available for evaluation, a model with quadratic decline in time resulted in a classifier with an area under the ROC curve of 0.68 for Tobit models and 0.63 for standard normal models. The classifier based on the Tobit model had higher sensitivity at all ranges of specificity. Mean squared error of predicted fourth follow-up scores was 4.95 for the Tobit model and 5.35 for the standard normal model. Accounting for ceiling effects improves both classification accuracy of cognitively impaired subjects, as well as score prediction for all subjects.
School of Medicine
Perkins, Jedediah, "Censored data models for longitudinal neuropsychological score analysis" (2011). Scholar Archive. 724.