Empirical likelihood inference for longitudinal data with covariate measurement errors: An application to the LEAN study

TL;DR

This paper introduces a new empirical likelihood method for analyzing longitudinal data with covariate measurement errors, accounting for different error distributions, and demonstrates its efficiency and effectiveness through simulations and application to the LEAN study.

Contribution

A novel empirical likelihood approach for longitudinal data with replicate measurement errors that accounts for distributional differences and improves estimation efficiency.

Findings

The method effectively eliminates measurement error effects.

It provides more efficient estimates than previous methods.

It reveals a significant intervention effect on BMI in the LEAN study.

Abstract

Measurement errors usually arise during the longitudinal data collection process. Ignoring the effects of measurement errors will lead to invalid estimates. The Lifestyle Education for Activity and Nutrition (LEAN) study was designed to assess the effectiveness of intervention for enhancing weight loss over nine months. The covariates systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured at baseline, month 4, and month 9. At each assessment time, there were two replicate measurements for SBP and DBP. The replicate measurement errors of SBP follow different distributions, as does DBP. To account for the distributional difference of replicate measurement errors, a new method for analyzing longitudinal data with replicate covariate measurement errors is developed based on the empirical likelihood method. The asymptotic properties of the proposed estimator are established under some regularity conditions. The confidence region for the parameters of interest can be constructed based on the chi-squared approximation without estimating the covariance matrix. Additionally, the proposed empirical likelihood estimator is asymptotically more efficient than the estimator of Lin et al. (2018). Extensive simulations demonstrate that the proposed method can eliminate the effects of measurement errors in the covariate and has a high estimation efficiency. The proposed method indicates the significant effect of the intervention on BMI in the LEAN study.