A Bayesian latent class model for predicting gestational age in health administrative data

Area of Science:

• Epidemiology
• Biostatistics
• Pharmacovigilance

Background:

• Health administrative data often lack gestational age at birth (GAB) information, limiting their use in drug safety research.
• Existing algorithms to estimate GAB from claims data may introduce bias due to inadequate modeling of GAB's distributional shape.

Purpose of the Study:

• To develop and validate a Bayesian latent class model for predicting gestational age at birth (GAB).
• To improve the accuracy of GAB estimation in large health administrative datasets for epidemiological studies.

Main Methods:

• A Bayesian latent class model using a mixture of Gaussian distributions with covariates within each class was developed.
• Markov Chain Monte Carlo methods were employed for posterior computation.
• The model was applied to a dataset of 10,043 Rhode Island Medicaid mother-child pairs.

Main Results:

• The three-class and six-class mixture models demonstrated maximal prediction accuracy for GAB.
• Medicaid women were successfully partitioned into three distinct classes based on birth timing: extreme preterm/preterm, preterm/early term, and late term.
• Obstetrical complications significantly influenced class membership.

Conclusions:

• The proposed Bayesian latent class model offers superior predictive accuracy compared to traditional linear models for GAB estimation.
• This approach effectively handles skewed response distributions and population heterogeneity, enhancing epidemiological research on drug safety in pregnancy.