| | Statistics seminar: ``Improved HorvitzThompson Estimation of Model Parameters from Stratified Case-Control/Cohort Data" by Norman Breslow, University of Washington | |
| | |
| |
Speaker
| | Norman Breslow, University of Washington |
| | | | |
| | Date | | Nov 12, 2009 |
| | | | |
| | Time | | 4:00 pm
|
| | | | |
| | Location | | Everitt Lab 165 (SPECIAL LOCATION) |
| | | | |
| | Sponsor | | Department of Statistics |
| | | | |
| | E-Mail | | office@illinois.edu |
| | | | |
| | Phone | | 3-2167 |
| | | | |
| | Event type | | Seminar |
| | | | |
| | Views | | 484 |
| | | | |
|
|
| |
| |
| Abstract: Stratified case-control and case-cohort studies involve simple random sampling
from an infinite superpopulation at phase one and stratified sampling without replacement from a finite cohort at phase two. The asymptotic variance of the Horvitz-Thompson (HT) estimator, of parameters in a variety of (semi)-parametric models, is the sum of two components: the model-based variance of the MLE that would be calculated from complete data for the entire cohort; and the design-based variance from HT estimation of the unknown cohort total of the (efficient) influence function contributions. The second component may be reduced by adjusting the sampling weights, e.g. by calibration to known cohort totals of auxiliary variables correlated with the influence function. Asymptotic theory suggests that further reduction may be possible by calibrating to within stratum totals of these same variables.
We compare standard and adjusted HT estimates of log hazard ratios for coronary heart disease using case-cohort data from the Atherosclerosis Risk in Communities Study and software available in Lumley's R survey package.
Similar comparisons are made using simulated case-cohort samples from the National Wilms Tumor Study. The results suggest that standard case-cohort analyses, as published today in the medical literature, may entail a substantial waste of information in comparison with the adjusted analyses. This is particularly true for hazard ratios associated with covariates known for the entire cohort. There are practical limits, however, on the number of adjustment variables that may be accommodated. |
| |
| |
|
|
