Using propensity score matching to address selection bias: An application using national data on breastfeeding and early childhood development

Propensity score (PS) matching can be employed to reduce selection bias in observational studies by balancing covariates across exposed and unexposed groups and ensuring common support.  Exclusive breastfeeding for six months is known to improve child health outcomes, but less evidence is available for developmental outcomes.  Studies of the effect of breastfeeding on child development may benefit from using PS-matching to reduce bias due to the selection processes affecting the decision to breastfeed.  This study demonstrates an application of PS-matching with nationally representative data to test whether extended breastfeeding duration improves early childhood development.  The 2012 National Survey of Children’s Health was used for children 6 months to 5 years old (n=20,032); exclusion criteria included conditions impacting development, such as autism.  Exposure was defined as extended breastfeeding (≥6 months vs <6 months or not at all).  Outcomes included risk of developmental delay (low, moderate, high), measured using the Parents’ Evaluation of Developmental Status screener.  Covariates included sixteen child, mother and household-level factors.  Propensity scores were estimated as the predicted probabilities from a logistic regression model of extended breastfeeding on all covariates.  Nearest neighbor 1:1 matching on PS was performed using a 0.02 caliper width (n=8,807 matched pairs). The counterfactual was estimated using matched analysis methods and results were compared with results from traditional multivariable regression models.  The weighted prevalence of extended breastfeeding was 47%, with 9% and 15% of children at high and moderate risk of delay, respectively.  Children exposed to extended breastfeeding were significantly less likely to experience high vs low/moderate delay (PS-matched RR=0.78, 95% CI=0.68, 0.88).  Results from the PS-matched sample were similar to, but more precise than multivariable model results.  Generalizability was sacrificed by matching unexposed to exposed on measured covariates; the matched sample was more likely to be white and Hispanic, have more highly educated parents, and have higher incomes than the general U.S. population.  Other limitations include residual confounding due to unmeasured factors and failure to match 24% of the exposed group.  Considering new analytic approaches such as PS-matching forces epidemiologists to reconsider the rigor of our methods for estimating causal relationships from observational studies.