A comparison of confounding adjustment methods with an application to early life determinants of childhood obesity

Abstract

We implemented 6 confounding adjustment methods: 1) covariate-adjusted regression, 2) propensity score (PS) regression, 3) PS stratification, 4) PS matching with two calipers, 5) inverse-probability-weighting, and 6) doubly-robust estimation to examine the associations between the BMI z-score at 3 years and two separate dichotomous exposure measures: exclusive breastfeeding versus formula only (N = 437) and cesarean section versus vaginal delivery (N = 1236). Data were drawn from a prospective pre-birth cohort study, Project Viva. The goal is to demonstrate the necessity and usefulness, and approaches for multiple confounding adjustment methods to analyze observational data. Unadjusted (univariate) and covariate-adjusted linear regression associations of breastfeeding with BMI z-score were −0.33 (95% CI −0.53, −0.13) and −0.24 (−0.46, −0.02), respectively. The other approaches resulted in smaller N (204 to 276) because of poor overlap of covariates, but CIs were of similar width except for inverse-probability-weighting (75% wider) and PS matching with a wider caliper (76% wider). Point estimates ranged widely, however, from −0.01 to −0.38. For cesarean section, because of better covariate overlap, the covariate-adjusted regression estimate (0.20) was remarkably robust to all adjustment methods, and the widths of the 95% CIs differed less than in the breastfeeding example. Choice of covariate adjustment method can matter. Lack of overlap in covariate structure between exposed and unexposed participants in observational studies can lead to erroneous covariate-adjusted estimates and confidence intervals. We recommend inspecting covariate overlap and using multiple confounding adjustment methods. Similar results bring reassurance. Contradictory results suggest issues with either the data or the analytic method.

Immigrant Legal Status and Health: Legal Status Disparities in Chronic Conditions and Musculoskeletal Pain Among Mexican-Born Farm Workers in the United States

“…Farm workers of different legal statuses have different distributions of observed characteristics, such as age and time in the United States (see Table 1), characteristics that are also associated with health. When the covariate structure is nonoverlapping or poorly overlapping among treatment groups, using simple covariate adjustment is likely to lead to biased results (Li et al 2014;Thoemmes and Ong 2016). IPTWs account for these imbalances in the distribution of the confounders across treatment groups more effectively than regression adjustment by creating a pseudo-population that could have been sampled from a population in which the observed covariates do not affect probability of treatment.…”

Prospective Study of Insufficient Sleep and Neurobehavioral Functioning Among School-Age Children

“…We used propensity scores to define overlapping covariate values, or “common support.” We ran common-support regression after excluding 16 participants where one or the other exposure group provided few data; results were similar so we do not report them. 22 …”

Which Propensity Score Method Best Reduces Confounder Imbalance? An Example From a Retrospective Evaluation of a Childhood Obesity Intervention

“…Presence of a large overlapping area of propensity scores, or “common support,” indicates that the use of propensity scores will help to balance the intervention and control groups on key confounders (Wagner, 2012). Though there is no theoretical guidance on what exactly merits common support, it can be considered the range of propensity scores with at least five observations per propensity score in both the intervention and control groups (Li, Kleinman, & Gillman, 2014). Limiting analyses to only participants whose propensity scores fall under the common support is recommended, because characteristics of individuals outside the common support may be too different to compare without introducing significant bias in the estimate of treatment effect (Li, Morgan, & Zaslavsky, 2014; Wagner, 2012).…”