Invited Commentary: Estimation and Bounds Under Data Fusion.

Area of Science:

• Epidemiology
• Biostatistics

Background:

• Data fusion involves combining datasets, but imputation of missing variables can introduce bias.
• Ogburn et al. highlighted bias risks when using regression imputation for completely missing variables.

Purpose of the Study:

• To investigate methods for eliminating or mitigating bias in epidemiologic data fusion when a covariate is missing in the primary dataset.
• To explore solutions for bias introduced by regression imputation of completely missing covariates.

Main Methods:

• Focuses on a linear outcome regression model with a missing covariate.
• Proposes a nonlinear imputation model for the missing covariate as a bias-elimination strategy.
• Describes two alternative approaches: using a validation dataset and deriving informative bounds.

Main Results:

• Bias in linear outcome regression can be eliminated if the imputation model for the missing covariate is nonlinear in shared variables.
• Two alternative methods can partially resolve the issue: outcome model fitting with validation data or bounds for coefficients without validation data.

Conclusions:

• Nonlinear imputation models offer a direct solution to bias in specific data fusion scenarios.
• Alternative methods provide partial resolution for bias when nonlinear imputation is not feasible, enhancing the reliability of epidemiologic data fusion.