Detecting potential outliers in longitudinal data with time-dependent covariates

Area of Science:

• Biostatistics
• Epidemiology
• Data Science

Background:

• Outliers can significantly distort regression model parameters and effect estimates in longitudinal data.
• Identifying visit-level outliers is critical, especially with skewed distributions of continuous time-dependent covariates.
• Accurate outlier detection enhances the reliability of associations between exposure variables and outcomes.

Purpose of the Study:

• To identify potential outliers at follow-up visits using the interquartile range (IQR) statistic.
• To assess the influence of these identified outliers on estimated Cox regression parameters.
• To evaluate the impact of extreme vitamin B12 intake on the risk of developing Islet Autoimmunity (IA).

Main Methods:

• The study utilized the TEDDY dietary longitudinal dataset, focusing on time-to-event data.
• An IQR algorithm was applied to detect outliers in continuous time-varying vitamin B12 intake at each visit.
• Extended time-dependent Cox models with robust sandwich estimators and partial residual diagnostics were used to assess outlier impact.

Main Results:

• Extreme vitamin B12 observations, particularly in Islet Autoimmunity (IA) cases, exerted a greater influence on the Cox regression model.
• Detected outliers demonstrably altered hazard ratios, standard errors, and the overall strength of association with IA risk.
• The direction of estimated effects was notably changed by the presence of these extreme values.

Conclusions:

• The IQR algorithm serves as an effective data quality control tool for identifying potential visit-level outliers during exploratory data analysis.
• Further investigation of outliers detected by the IQR method is recommended to ensure robust statistical modeling.
• This approach is valuable for handling longitudinal data with skewed, time-dependent covariates in epidemiological studies.