A misclassification model is presented for the assessment of bias in rate ratios estimated by person-time analyses of automated medical care databases.
The model allows for misclassification of events and person-time and applies to both differential and nondifferential errors.
The focus is on medical care exposures that occur at discrete points in time (e.g., vaccinations) and on adverse events that are closely associated in time.
Bias corrections for rate ratios and binomial tests of equality of event rates during exposed and unexposed person-time are developed and illustrated.
For nondifferential under-or over-ascertainment of events, the observed rate ratio (r) is unbiased at the null hypothesis (true rate ratio R=1), negatively biased when R>1, and positively biased when R<1 (i.e., biased toward the null).
Differential under-ascertainment of unexposed events and differential over-ascertainment of exposed events positively bias r when R=1. Differential event sensitivities cause larger biases in rate ratios than differential false event rates.
False positive exposures bias observed event rate ratios more than false negative exposures.
Biases are small when event sensitivities are nondifferential and when less than 10% of database exposures and events are false.
The usefulness of the model for critical sensitivity analysis is illustrated by an example from a linked database study of childhood vaccine safety. (...)
Mots-clés Pascal : Base donnée, Informatique, Modèle statistique, Vaccination, Médicament, Toxicité, Complication, Qualité, Information, Epidémiologie, Biais
Mots-clés Pascal anglais : Database, Computer science, Statistical model, Vaccination, Drug, Toxicity, Complication, Quality, Information, Epidemiology, Bias
Notice produite par :
Inist-CNRS - Institut de l'Information Scientifique et Technique
Cote : 96-0459836
Code Inist : 002B30A01A1. Création : 10/04/1997.