Propensity score matching for causal inference and reducing the confounding effects: statistical standard and guideline of Life Cycle Committee

Abstract

Since the development of research methodology, there has always been keen interest in developing the accuracy of the research by comparing covariates. Propensity score is useful when the research covers many variables which are not intended to be included as independent variables, thus allowing the removal of certain covariates from the model. This review discusses a general aspect of propensity score matching, which begins with the mathematical principles of propensity score matching. The concept and context of propensity score matching is also explained, which includes the advantages of propensity score matching over conventional research methods and the reasons for the introduction of propensity score matching in medical research. It is our aim that readers learn how to actually obtain a propensity score. Discussion of several options for performing matching based on the propensity score is also included, and the final topic is the adequacy of the matched cohort being evaluated with standardized mean differences, and research methods including Cox regression being conducted on the matched cohort. We hope to assist readers in understanding when and how to perform propensity score matching through this paper.

Prenatal and postnatal exposure to antibiotics and risk of food allergy in the offspring: A nationwide birth cohort study in South Korea

“…To minimize the effects of confounding due to baseline covariates, we conducted a 1:1 exposure-driven propensity score matching. 23 The propensity score was estimated using an adjusted logistic regression model, with prenatal and postnatal antibiotics exposure as the independent variable and all the baseline variables presented in Table 1 as covariates. We used a "greedy nearest-neighbor matching algorithm" to match children in two groups with random selection and without participant replacement within specified caliper widths (0.001 standard deviations).…”

“…Standardized mean differences (SMDs) were estimated to assess the balance of covariate distribution. 23 A SMD ≥ 0.1 was considered a sign of imbalance in the two groups.…”

“…To minimize the effects of confounding due to baseline covariates, we conducted a 1:1 exposure‐driven propensity score matching 23 . The propensity score was estimated using an adjusted logistic regression model, with prenatal and postnatal antibiotics exposure as the independent variable and all the baseline variables presented in Table 1 as covariates.…”

“…We used a “greedy nearest‐neighbor matching algorithm” to match children in two groups with random selection and without participant replacement within specified caliper widths (0.001 standard deviations). Standardized mean differences (SMDs) were estimated to assess the balance of covariate distribution 23 . A SMD ≥ 0.1 was considered a sign of imbalance in the two groups.…”

“…In order to reduce potential confounders, we additionally conducted the propensity score-matched cohort. 23 For the propensity score-matched cohort, we matched children with and without antibiotic exposure in a 1:1 according to the following factors: maternal age at delivery, infant sex, region of residence, birth season, income, nulliparous, delivery year, maternal history of food allergy, atopic dermatitis, and asthma, Severe Maternal Morbidity (SMM) score, 24 delivery mode, preterm birth, low birthweight, the number of antibiotic prescriptions within a year before pregnancy, the number of inpatient admissions of the mother within a year before pregnancy, and the number of outpatient contacts of the mother within a year before pregnancy. Finally, we included a total of 1,639,408 children (n = 819,704 for children with antibiotic exposure, and n = 819,704 for children without antibiotic exposure) in the propensity score-matched cohort.…”