Poly-Matching Causal Inference for Assessing Multiple Acute Medical Managements of Pediatric Traumatic Brain Injuries

摘要 临床效果研究(CER)在与紧急医疗服务相关的研究中发挥着核心作用 儿童(EMSC)。它是在广泛地评估干预措施时认真使用可用的最佳证据。 被定义为可能导致健康改善的医疗、健康政策或实践模式 护理质量和患者结局。观察性数据更多地用于医疗保健的评估 与随机对照试验(RCT)相比，由于实际或伦理原因，系统或复杂的临床实践。 然而，利用观测数据进行因果推断面临着挑战：(1)重要的协变量可能是分布的 不同的治疗方案不同；(2)常规统计分析缺乏对未测量的控制 令人困惑。当干预是二分法时，基于倾向分数的调整被广泛使用来减少 由观察到的协变量通过匹配、分层或加权引入的混杂偏差。匹配 是研究人员的流行选择，因为它创建了类似于RCT的数据结构，易于解释，并且 对结果建模中的错误规范具有健壮性。但是，有一个严重的方法论差距阻碍了对 有多个(两个以上)治疗选择时的匹配设计。这是因为缺少好的 匹配算法生成匹配良好的集合，增加了匹配后推理的复杂度。 我们的首要目标是开发统计上有效的匹配设计(称为PMD)和后续 与复杂的观察性医疗保健数据库一起使用的因果推断程序，其中有多个 治疗武器或治疗跨越多个时间点。具体目标：(1)设计一个创新的PMD用于研究 随着时间的推移有多个治疗臂或治疗方法；根据PMD的 评估未测量混杂的潜在结果框架和敏感性分析策略 (2)评估接受不同创伤护理的重型颅脑损伤(STBI)患者的因果死亡率影响。 创伤中心类型(PL1-1级儿科、AL1-1成人和ML1-1混合创伤中心；PL2- 2级儿科、AL2-2级成人和ML2-2级混合创伤中心)；(3)评估4个 1-2级医疗管理/治疗(颅内压监测、头部CT扫描、脑脊液引流、 去骨瓣减压术)对sTBI患者死亡率的影响；(4)：评估疾病中心的依从性 不同类型医院轻型颅脑损伤控制与预防(CDC)头部CT指南。 这项研究有望通过扩展常用的二分法来填补EMSC研究中的一个关键空白 具有多个治疗组的复杂观察性研究的配对设计。这个项目意义重大， 我们提出的方法是创新的，因为它们既包括观察到的混杂调整，也包括未测量的 令人困惑的评估。我们设想，这种通用方法将得到广泛应用，并能够 使政府机构、政策制定者以及社会科学和健康科学研究人员受益，在这些领域， 数据经常被用于比较结果研究和方案/政策评估。

Abstract Clinical effectiveness research (CER) plays a central role in research related to emergency medical services for children (EMSC). It is the conscientious use of the best available evidence in evaluating interventions, broadly defined as medical treatments, health policies, or practice patterns, which could lead to improvements in health care quality and patient outcomes. Observational data are more often used in the evaluation of healthcare systems or complex clinical practice than randomized controlled trials (RCTs), due to practical or ethical reasons. However, causal inference with observational data faces challenges: (1) Important covariates may be distributed differently between treatment options; (2) Conventional statistical analysis lacks control for unmeasured confounding. When the intervention is dichotomous, propensity score based adjustment is widely used to reduce the confounding bias introduced by observed covariates, through matching, stratification or weighting. Matching is a popular choice among researchers, as it creates data structures similar to RCTs, is easy to interpret, and robust to misspecifications in outcome modeling. But there is a critical methodological gap hindering the use of matching design when there are multiple (more than two) treatment options. This is due to the lack of good matching algorithms to generate well matched sets and the increased complexity of post-matching inference. Our overarching goal is to develop a statistically valid matching design (referred to as PMD) and subsequent causal inference procedures for use with complex observational healthcare databases, where there are multiple treatment arms or treatments over multiple time points. Specific aims: (1) Devise an innovative PMD for studies with multiple treatment arms or treatments over time; Develop causal inference strategies for PMD based on the potential outcome framework and sensitivity analysis strategies for assessing the unmeasured confounding effect; (2) Evaluate causal mortality impact of severe TBI (sTBI) patients who received trauma care at different type of trauma centers (PL1-level 1 pediatric, AL1-level 1 adult, and ML1-level 1 mixed trauma centers; PL2- level 2 pediatric, AL2-level 2 adult, and ML2-level 2 mixed trauma centers); (3) Assess the effectiveness of 4 Tier 1-2 medical management/therapies (ICP monitoring, head CT scan, cerebrospinal fluid drainage, decompressive craniectomy) on sTBI patient mortality; (4): Evaluate compliance with a Centers for Disease Control and Prevention (CDC) head CT guideline for mild TBI (mTBI) patients by different types of hospitals. This study is expected to fill a critical gap in EMSC research by extending the commonly used dichotomous matching design to complex observational studies with multiple treatment groups. This project is significant and our proposed methods are innovative as they include both observed confounding adjustment and unmeasured confounding assessment. We envision that this general-purpose methodology will be widely applicable and can benefit government agencies, policy makers, and social and health science researchers, where observational data are often utilized for comparative outcomes research and program/policy evaluation.