Propensity Score Calibration in the Absence of Surrogacy

Propensity score calibration (PSC) can be used to adjust for unmeasured confounders using a cross-sectional validation study that lacks information on the disease outcome (Y), under a strong surrogacy assumption. Using directed acyclic graphs and path analysis, the authors developed a formula to pre...

Full description

Saved in:
Bibliographic Details
Published in:American journal of epidemiology 2012-06, Vol.175 (12), p.1294-1302
Main Authors: LUNT, Mark, GLYNN, Robert J, ROTHMAN, Kenneth J, AVORN, Jerry, STURMER, Til
Format: Article
Language:English
Subjects:
Bias
Biological and medical sciences
Cohort Studies
Confounding Factors, Epidemiologic
Cross-Sectional Studies
Data Interpretation, Statistical
Epidemiologic Research Design
Epidemiology
General aspects
Medical sciences
Miscellaneous
Models, Statistical
Poisson Distribution
Practice of Epidemiology
Propensity Score
Public health. Hygiene
Public health. Hygiene-occupational medicine
Validation studies
Validation Studies as Topic
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Propensity score calibration (PSC) can be used to adjust for unmeasured confounders using a cross-sectional validation study that lacks information on the disease outcome (Y), under a strong surrogacy assumption. Using directed acyclic graphs and path analysis, the authors developed a formula to predict the presence and magnitude of the bias of PSC in the simplest setting of a binary exposure (T) and 1 confounder (X) that are observed in the main study and 1 confounder (C) that is observed in the validation study only. PSC bias is predicted on the basis of parameters that can be estimated from the data and a single unidentifiable parameter, the relative risk (RR) associated with C (RR(CY)). The authors simulated 1,000 cohort studies each with a Poisson-distributed outcome Y, varying parameter values over a wide range. When using the true parameter for RR(CY), the formula predicts PSC bias almost perfectly in this simple setting (correlation with observed bias over 24 scenarios assessed: r = 0.998). The authors conclude that the bias from PSC observed in certain scenarios can be estimated from the imbalance in C between treated and untreated persons, after adjustment for X, in the validation study and assuming a range of plausible values for the unidentifiable RR(CY).
ISSN:0002-9262
1476-6256
DOI:10.1093/aje/kwr463