Abstract:
|
High-frequency digital biomarkers can improve the detection and prediction of human disorders and diseases. However, it is challenging to leverage biomarkers in traditional modeling techniques as the outcome of interest is often assessed less frequently than the biomarkers. In addition, biomarkers rarely are completely observed. Thus, it is necessary to summarize biomarker data to match the outcome’s frequency while simultaneously imputing missing biomarker values. Via simulation, we assess two processes involving imputation of biomarkers for longitudinal analysis of a binary outcome: (1) Impute then Summarize (IS), where we impute at the finest time granularity (daily), then summarize, and (2) Summarize then Impute (SI), where we summarize the biomarker then impute at the summary level. Our results show that accuracy of coefficient estimation depends on percent missing data, length of consecutive missing days, and the rate of trajectory change of the biomarkers. A real data analysis involving mild cognitive impairment is used to demonstrate the empirical differences.
|