Statistical methods for modeling repeated measures of maternal environmental exposure biomarkers during pregnancy in association with preterm birth

Abstract

Background: It is of critical importance to evaluate the role of environmental chemical exposures in premature birth. While a number of studies investigate this relationship, most utilize single exposure measurements during pregnancy in association with the outcome. The studies with repeated measures of exposure during pregnancy employ primarily cross-sectional analyses that may not be fully leveraging the power and additional information that the data provide.

Methods: We examine 9 statistical methods that may be utilized to estimate the relationship between a longitudinal exposure and a binary, non-time-varying outcome. To exemplify these methods we utilized data from a nested case-control study examining repeated measures of urinary phthalate metabolites during pregnancy in association with preterm birth.

Results: The methods summarized may be useful for: 1) Examining sensitive windows of exposure in association with an outcome; 2) Summarizing repeated measures to estimate the relationship between average exposure and an outcome; 3) Identifying acute exposures that may be relevant to the outcome; and 4) Understanding the contribution of temporal patterns in exposure levels to the outcome of interest. In the study of phthalates, changes in urinary metabolites over pregnancy did not appear to contribute significantly to preterm birth, making summary of average exposure across gestation optimal given the current design.

Conclusions: The methods exemplified may be of great use in future epidemiologic research projects intended to: 1) Elucidate the complex relationships between environmental chemical exposures and preterm birth; 2) Investigate biological mechanisms in prematurity using repeated measures of maternal factors throughout pregnancy; and 3) More generally, address the relationship between a longitudinal predictor and a binary, non-time-varying outcome.

Figures

Figure 1

Scatterplot of fitted intercepts and slopes from the mixed effects model with MEHP regressed on gestational age.

Figure 2

Fitted smooth curves between phthalate levels and gestational age.

Figure 3

Estimated mean of clusters* suggested by the Gaussian mixture model, stratified by study visit.

Figure 4

Mean trajectories of clusters* based on functional curves of phthalates vs. gestational age. *Clusters are constructed based on functional k-means clustering (Functional clustering model) of the smooth curve of residuals (from a regression model of MEHP or MBP on relevant covariates) against gestational age. N=443 for each phthalate model.