Li L, Huang J, Sun S, Shen J, Unverzagt FW, Gao S, Hendrie HH, Hall K, Hui SL. Selecting pre-screening
items for early intervention trials of dementia-a case study. Stat Med. 2004 Jan 30;23(2):271-83.
Our goal was to review and extend statistical methods for discriminating between normal subjects and those
with dementia or cognitive impairment. We compared six different methods to one constructed by expert opinion,
in their brevity and predictive power. The methods include logistic regression and neural networks, with standard
and least absolute shrinkage and selection operator (LASSO) variable selection, as well as decision trees with and
without boosting. These methods were applied to the baseline data of a subgroup of subjects in a dementia study,
using their screening interview items to predict their clinical diagnosis of normal or non-normal (cognitively
impaired or demented). The derived models were then validated on a different subgroup of subjects in the same
study who had the screening and clinical diagnosis two to five years later. Performance of different models was
compared based on their sensitivity and specificity in the validation sample. Generally, the six statistical methods
performed slightly to moderately better than the expert-opinion model. Neural networks generally performed better
than the logistic and decision tree models. LASSO improved the performance of logistic and neural network models, but
it eliminated few input variables in the neural network. The single decision tree performed at least as well as the
standard logistic model, and with fewer items, making it an attractive pre-screening tool. Using the boosting option
for decision trees did not substantially improve the performance. We recommend that for each situation, different methods
of classification should be attempted to obtain optimal results for a given purpose. Copyright 2004 John Wiley & Sons, Ltd.
Email us comments or suggestions