Comparative Analysis of Propensity score?Based Methods for Estimating Factors Associated with Delayed Cerebral Ischemia after Aneurysmal SAH using Multicenter Observational Data = 다기관 관찰자료를 이용한 동맥류성 지주막하출혈 후 지연성 뇌허혈 관련 요인 추정을 위한 성향점수 기반 방법들의 비교 분석

다국어 초록 (Multilingual Abstract)

Analyzing treatment effects in real-world multicenter data poses substantial methodological challenges, particularly when treatment assignment is highly imbalanced and confounding is severe. This study aimed to demonstrate the implementation, performance characteristics, and practical considerations of four propensity score (PS)–based methods—matching, stratification, inverse probability of treatment weighting (IPTW), and covariate adjustment—using an observational dataset of patients with aneurysmal subarachnoid hemorrhage. Data from a prospective multicenter registry were analyzed. PSs were estimated via logistic regression incorporating demographic, clinical, and center-specific covariates. Four PS-based approaches were applied: nearest-neighbor matching with multiple ratios and caliper widths to optimize covariate balance; stratification using exposed-based strata with varying numbers of quantiles to control for confounding; IPTW with stabilized weights and multiple trimming procedures to mitigate extreme propensity values; and covariate adjustment including the PS, with and without additional temporal covariates. Across all methods, treatment effects on delayed cerebral ischemia (DCI) were estimated with adjusted odds ratios (ORs), 95% confidence intervals (CIs), and p-values, and covariate balance was assessed using standardized mean differences. Effect estimates varied with methodological choices. PS matching ORs ranged from 0 to 1.09 depending on matching ratio and caliper, with reduced sample sizes limiting precision. Stratification produced consistent but imprecise estimates (ORs 0.75–0.79), while IPTW results were highly sensitive to trimming strategy (ORs 0.46– 1.85). Covariate adjustment yielded similarly wide CIs. None of the four methods demonstrated statistically significant associations. This study illustrates the practical limitations of PS-based methods in low-incidence, heterogeneous multicenter observational datasets. Method selection, balance diagnostics, and sensitivity analyses are critical for valid causal inference when dealing with extreme PS distributions, sparse strata, and small sample sizes.

목차 (Table of Contents)

• - Table of Contents -
• I. Introduction 1
• II. Method 3
• 2.1 Review of PS methodologies 3
• 2.1.1 PS matching 4
• 2.1.2 Stratification 5
• 2.1.3 IPTW 6
• 2.1.4 Covariate adjustment 7
• 2.2 Data source 8
• 2.3 PS Estimation and Implementation of PS based methods 9
• 2.4 Balance assessment and diagnostics 11
• 2.5 Statistical analysis 11
• III. Results 12
• 3.1 Patients’ demographics 12
• 3.2 Multivariable Logistic Regression 14
• 3.3 PS distribution 16
• 3.4 PSM 17
• 3.5 Stratification 19
• 3.6 IPTW 19
• 3.7 Covariate adjustment 22
• IV. Discussion 23
• Reference 28
• Abstract in Korean 30
• - List of Tables -
• Table 1. Patients’ demographics 12
• Table 2. Adjustment effect estimates of Multivariable logistic regression 14
• Table 3. Standardized mean differences for covariates after 3:1 propensity
• score matching with caliper 0.2 17
• Table 4. Standardized mean differences for covariates in untrimmed and
• common-range trimmed IPTW models 20
• Table 5. Adjusted Treatment Effect Estimates of PSM, Stratification, IPTW,
• and Covariate Adjustment 22
• - List of Figures -
• Figure 1. The distribution of propensity score in the treated and untreated
• groups 16