Agreement Between Two Quantitative Measurement Methods When the Underlying Latent Trait Is Not Constant

Most statistical methods that have been developed to assess the agreement between two quantitative measurement methods have (implicitly) relied on the assumption of a constant "individual" latent trait. This might be inappropriate when the "individual" is not an object but a person. Therefore, the goal of this study was to extend the standard measurement error model to cope with this limit. Four different settings were investigated: first, where the true individual latent trait was constant; second, where it was variable but without exhibiting a time trend; third, where it followed a linear time trend; and fourth, where it exhibited an approximate linear time trend. Two competing methods to estimate the parameters of the general measurement error model were assessed: the GLS estimator of Sprent and the two-stage method of Taffé. It was found that the latter generally performed better than the former to estimate the bias. In addition, it can be used when there is only a single measurement per individual by one of the two measurement methods, which is not the case with the former method.