New insights into handling missing values in environmental epidemiological studies

Area of Science:

• Environmental Epidemiology
• Biostatistics
• Data Science

Background:

• Missing data are common in environmental epidemiologic surveys, posing challenges for analysis.
• Accurate handling of missing values is essential for reliable study findings.

Purpose of the Study:

• To compare different methods for managing large amounts of missing data in environmental epidemiology.
• To evaluate the performance of omission, single/multiple imputation, and Bayesian approaches using a real-world cohort and simulations.

Main Methods:

• Applied omission, single/multiple imputation (linear regression, partial least squares), and a fully Bayesian approach to the PARIS birth cohort data.
• Conducted a simulation study with 50-95% missing values, varying odds ratios (1.0-1.4) and health outcome prevalence.

Main Results:

• Omitting data reduced power and inflated standard errors; single imputation underestimated variability and increased Type I error.
• The Bayesian joint model was efficient, especially for common outcomes, showing low RMSE, reasonable Type I error, and high power.
• For less prevalent events, Bayesian methods showed increased Type I error and reduced power, but posterior distributions aided interpretation.

Conclusions:

• No single method is universally best for missing data in environmental epidemiology.
• A joint modeling approach for the missing data process and health association is more efficient when substantial data are missing.
• The fully Bayesian approach offers advantages for common outcomes, while careful consideration is needed for rare events.