Activity Number:
|
141
|
Type:
|
Topic Contributed
|
Date/Time:
|
Monday, August 12, 2002 : 2:00 PM to 3:50 PM
|
Sponsor:
|
Section on Health Policy Statistics*
|
Abstract - #301055 |
Title:
|
A Hierarchical Model for Detecting Treatment Interaction in Meta-Regression of Individual Patient and Summary Data
|
Author(s):
|
Paul Stark*+ and Christopher Schmid+
|
Affiliation(s):
|
Tufts New England Medical Center and New England Medical Center
|
Address:
|
750 Washington St. Box 63, Boston, Massachusetts, 02111, USA 750 Washington Street, Box 63, Boston, Massachusetts, 02111,
|
Keywords:
|
Meta-regression ; Multilevel models ; Treatment interactions ; Randomized controlled trials
|
Abstract:
|
We develop Bayesian multilevel models that incorporate both patient-level and study-level forms of variables to test the significance of aggregate and individual level effects for detecting treatment interactions in meta-analyses of binary and continuous responses. We compare the use of aggregate and individual-level variables for detecting treatment interactions with factors that vary at the patient level within the study, using 1860 patients from 11 randomized controlled trials of angiotensin converting enzyme inhibitors for treatment of non-diabetic renal disease. Meta-regression is unable to detect an interaction between treatment and baseline level of urine protein found by analysis of individual patient data for two treatment efficacy measures: change in glomerular filtration rate; or progression to end-stage renal disease or doubling of serum creatinine. We conclude that sufficient heterogeneity of treatment effects and sufficient variation among the study-level variables must be present for meta-regression to be able to detect treatment interaction. Researchers should not infer from a lack of between-study heterogeneity that the same treatment effect applies to every patient.
|