Evaluating latent class models with conditional dependence in record linkage.

Area of Science:

• Data Science
• Biostatistics
• Health Informatics

Background:

• Traditional record linkage uses latent class models (Fellegi-Sunter).
• These models assume conditional independence of field agreements, which can lead to biased estimates when violated.
• Conditional dependence impacts misclassification rates, sensitivity, and positive predictive value.

Purpose of the Study:

• To characterize the impact of violating conditional independence in record linkage.
• To evaluate methods for accounting for conditional dependence.
• To improve accuracy in record linkage for health information exchange data.

Main Methods:

• Utilized loglinear models with interaction terms identified via correlation residual plots.
• Employed Gaussian random effects models to handle conditional dependence.
• Applied proposed models to link newborn screening data from a health information exchange.

Main Results:

• Loglinear models with interaction terms showed the lowest misclassification rate in simulations.
• Gaussian random effects models outperformed the conditional independence assumption.
• Gaussian random effects models performed comparably to loglinear models in some scenarios and can handle continuous agreement measures.

Conclusions:

• Violating conditional independence assumptions in record linkage models negatively impacts accuracy.
• Loglinear and Gaussian random effects models offer improved performance over traditional methods.
• Gaussian random effects models provide a flexible approach for record linkage with conditional dependence and diverse data types.