Objective: We propose a nonparametric method for handling different types of missing covariates in longitudinal studies. We show that the proposed method outperforms multiple standard approaches via extensive simulations as well as an application to a biomarker study in the Parkinson’s disease (PD) longitudinal cohort.

Background: Missing covariates are common in longitudinal studies. There are various causes of missing data; e.g., a study design may only require a subgroup of PD patients to take invasive biomarker tests for cerebrospinal fluid (CSF) amyloid β 1-42 (Aβ 1-42). The Penn PD dataset consists of 408 PD patients, more than half of which missed the measurements of CSF Aβ 1-42. Motivated by this PD research, we develop a novel statistical method that utilizes complete auxiliary information to impute the missing covariates.

Method: We have developed a new statistical method utilizing auxiliary variables that are related to the missing covariates without assuming any distributional assumptions. The proposed method can deal with continuous or discrete, time-invariant or -varying missing covariates. Besides, we allow for different types of auxiliary variables. We have compared the proposed model with standard missing data methods, including available case analysis (ACA) and multiple imputation (MI), through percentage bias and relative efficiency. We have applied the proposed method to the Penn PD study to investigate how abnormal measurement of CSF Aβ 1-42 affects the change in age-adjusted Dementia Rating Scale total score, where the apolipoprotein E information is used as the auxiliary variable.

Results: We find that the proposed method produces unbiased results under all the considered missing data scenarios. The proposed method is consistently more efficient than ACA, and recovers a great deal of efficiency loss caused by missing data. Even when the relationship between missing covariates and auxiliary information is extremely nonlinear, the proposed method still gives unbiased estimates and maintains higher efficiency gains than ACA, whereas standard MI methods are biased.

Conclusion: The method proposed in our study provides a robust and efficient approach to dealing with different types of missing covariates in longitudinal studies. It is particularly attractive when the underlying relationship between missing covariates and auxiliary information is unknown.

« Back to 2022 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/robust-methods-for-missing-covariates-in-longitudinal-studies-with-application-to-biomarker-research-in-parkinsons-disease-dementia/