Modeling the Association between Repeated Measures of Hemoglobin during Pregnancy and Adverse Birth Outcomes

Maternal hemoglobin (Hb) concentrations and their trajectories throughout pregnancy are important determinants of birth outcomes. However, longitudinal studies on pregnancy frequently rely on cross-sectional analyses at specific time points or utilize summary measures, overlooking valuable information contained in repeated Hb measurements.

This study aimed to illustrate various statistical approaches for modeling longitudinal Hb data and their associations with birth outcomes, highlighting the strengths and limitations of each method.

We analyzed 8 pregnancy datasets (6452 women, 13,580 Hb measurements) from the Biomarker Reflecting Inflammation and Nutritional Determinants of Anemia project using: 1) logistic regression incorporating residual Hb, 2) 2-stage mixed effect model, 3) distributed lag nonlinear model (DLNM), 4) generalized additive mixed model (GAMM), and 5) group-based trajectory modeling (GBTM). Outcomes were low birth weight (<2.5 kg), preterm birth (PTB, <37 wk), and small for gestational age (birthweight <10th percentile).

Logistic regression using residual Hb, 2-stage mixed-effects model, and DLNM did not reveal any significant associations between Hb concentrations and adverse birth outcomes. GAMM showed that women with PTB had lower Hb concentrations before 20 wk of gestation compared with those without PTB. GBTM identified 4 distinct Hb trajectory clusters, but no significant associations were found between trajectory groups and adverse birth outcomes.

These analytic approaches provide complementary insights into the relationship between maternal Hb and birth outcomes, while illustrating how inference can vary depending on the method used. DLNMs can help pinpoint critical gestational periods of vulnerability, whereas models such as GAMM and GBTM capture nonlinear trends and heterogeneous trajectories. Researchers should be aware that conclusions about Hb and birth outcomes may be highly sensitive to modeling decisions. Overall, these methods can guide researchers in selecting statistical strategies best suited to their study aims and data structure.