|
Activity Number:
|
56
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Sunday, August 8, 2004 : 4:00 PM to 5:50 PM
|
|
Sponsor:
|
Section on Statistics in Epidemiology
|
| Abstract - #301914 |
|
Title:
|
Combining Variables by the Use of Principal Components Can Lead to an Index with Low Reliability--Two Examples
|
|
Author(s):
|
John S. Grove*+ and Clementina Ceria and J. David Curb
|
|
Companies:
|
John A. Burns School of Medicine and University of Hawaii, Manoa and Pacific Health Research Institute
|
|
Address:
|
1960 East West Rd., Honolulu, HI, 96822,
|
|
Keywords:
|
principal components ; reliability coefficient ; intraclass correlation
|
|
Abstract:
|
Investigators often face having either several imprecise measures of the same underlying variable or a set of variables which are thought to be related. Using them simultaneously as independent variables in a regression model creates estimates which are difficult to interpret and have large standard errors due to intercorrelations. One common reaction is to create an index of the variables from the first eigenvector of principal component (PC) analysis. But since variables are weighted according to their variances, variables with high random error variance may contribute the most to the first component. Two examples are given of this. The first PC of systolic blood pressure (SBP) measured at two exams with different reliabilities resulted in a lower Cox regression coefficient for heart disease incidence on SBP than using the better of the two measurements alone. Using data from the Pacific Genetics Network study on balance measured by an electronic balance platform, the first principal component of eight variables had a reliability coefficient (RC) of only 0.4, whereas two of the input balance variables had RC's > 0.6. The best linear function for balance had a RC of 0.8.
|
