Premium
Correcting for measurement error in binary and continuous variables using replicates
Author(s) -
White Ian,
Frost Chris,
Tokunaga Shoji
Publication year - 2001
Publication title -
statistics in medicine
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.996
H-Index - 183
eISSN - 1097-0258
pISSN - 0277-6715
DOI - 10.1002/sim.908
Subject(s) - statistics , binary number , confounding , observational error , mathematics , value (mathematics) , regression , binary data , regression analysis , econometrics , arithmetic
Measurement error in exposures and confounders leads to bias in regression coefficients. It is possible to adjust for this bias if true values or independent replicates are observed on a subsample. We extend a method suitable for quantitative variables to the situation where both binary and quantitative variables are present. Binary variables with independent replicates introduce two extra problems: (i) the error is correlated with the true value, and (ii) the measurement error probabilities are unidentified if only two replicates are available. We show that – under plausible assumptions – adjustment for error in binary confounders does not need to address these problems. The regression coefficient for a binary exposure is overadjusted if methods for continuous variables are used. Correct adjustment is possible either if three replicates are available, or if further assumptions can be made; otherwise, bounds can be put on the correctly adjusted value, and these bounds are reasonably close together if the exposure has prevalence near 0.5. Copyright © 2001 John Wiley & Sons, Ltd.