[Skip to Content]
Access to paid content on this site is currently suspended due to excessive activity being detected from your IP address 54.211.168.204. Please contact the publisher to request reinstatement.
[Skip to Content Landing]
Citations 0
Correspondence
September 2007

Medication Use as a Confounding Factor in the Use of the Cerebrospinal Fluid tau/β-Amyloid42 Ratio—Reply

Arch Neurol. 2007;64(9):1357-1359. doi:10.1001/archneur.64.9.1357-b

In reply

We thank Dr Vanacore and colleagues and Dr Wong for their interest in our study and appreciate the opportunity to respond to their concerns. Concerning Dr Vanacore's first point, we obtained similar sensitivity, specificity, and positive and negative predictive values reported by these investigators using equations developed for the traditional diagnostic test setting, where the variable used for prediction is measured at the same point in time as the outcome. However, these statistics are less appropriate when future onset of disease is assessed in association with potential predictors measured only at baseline over varying follow-up times among subjects, as in our study. These types of longitudinal studies require that the prediction of future disease onset takes into account not only whether participants will eventually convert to a CDR greater than 0, but also the variable length of the observation period and the fact that participants could drop out of the study any time before their conversion or might never convert during the entire follow-up. Cox proportional hazards models,1 when justified by the data, have been the standard approach in analyzing these types of data and therefore were what we reported. The estimated hazards ratios and the 95% confidence intervals we reported are the standard statistics for assessing the strength of associations between potential baseline predictors and future disease onset. We used a cutoff of 0.214 only to graphically demonstrate how the baseline value of CSF ptau181/Aβ42 could be associated with future disease onset via a survival curve. The rather poor sensitivity and positive predictive values obtained with the cutoff, which are only appropriate for cross-sectional measures of predictive accuracy because they ignore the longitudinal feature of a study, are in fact mathematically expected when the total number of converters is low.

First Page Preview View Large
First page PDF preview
First page PDF preview
×