ABSTRACT
In Section 3.6 of Chapter 3, we briefly described covariate measurement error problems in regression models and briefly reviewed several general approaches to address measurement errors. It is known that the naive method which ignores covariate measurement errors in regression models may produce biased and misleading results. To visually illustrate this, Figure 5.1 shows how the estimates of regression coefficients may be biased if measurement errors in a covariate is not addressed, based on an artificial dataset generated from the simple linear regression y = β0 + β1x + e. To formally address measurement errors, we typically need to have validation data or replication data (Carroll et al. 2006). In practice, these validation data or replication data may not be available. However, for longitudinal studies the repeated measurements within each individual may be viewed as replication data, which allows us to partially address measurement errors in time-dependent covariates.
