Methods for the Analysis of Longitudinal Social Network Data

纵向社交网络数据分析方法

• 批准号: 8377365
• 负责人: A. JAMES O'MALLEY
• 金额: $ 10.42万
• 依托单位: HARVARD MEDICAL SCHOOL
• 依托单位国家: 美国
• 资助国家: 美国
• 起止时间: 至 2013-06-30
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/8377365
• 关键词: Accounting; Adopted; Adoption; Affect; Alcoholism; Behavior; Body mass index; Characteristics; Complex; Cross-Sectional Studies; Data; Data Analyses; Development; Ego; Epidemic; Equation; Event; Genes; Genetic; Geographic Locations; Health; Health behavior; Health behavior outcomes; Heterosexuals; Homosexuals; Individual; Life; Literature; Marriage; Measurement; Measures; Mental Health; Methodology; Methods; Modeling; Monitor; Myocardial Infarction; Nature; Neighborhoods; Nodal; Obesity; Outcome; Outcome Measure; Pathway Analysis; Patients; Persons; Physicians; Play; Population; Positioning Attribute; Principal Investigator; Probability; Property; Research; Role; Series; Shock; Smoking; Social Network; Statistical Methods; Stimulus; Structure; Techniques; Testing; Time; Triad Acrylic Resin; Work; affection; base; data modeling; experience; innovation; interest; longitudinal analysis; physical conditioning; programs; social; statistics; trait; transmission process

We propose to develop new statistical methods for more precise estimation of the influence of one individual on another in a network, testing and controlling for selection effects such as homophily observed between individuals. The work is challenging because network data may contain multiple types of information, including network topology, nodal covariates, tie characteristics, and temporal change. The central problem is accounting for the complex correlation structure that arises because each actor in the network may play the dual role of an ego (rater or responder) and alter (target or stimulus) and thus may appear in the data multiple times. Furthermore, outcomes might be geographically correlated and correlated over time if subjects are followed longitudinally. Here, we focus on development of statistical methodology for longitudinal analysis as this provides the best opportunity for obtaining causal inferences; however, we also propose innovations involving cross-sectional analysis of networks. We have three specific aims: (1) To develop methodology for longitudinal analysis of egocentric data. The objective of such analysis is to determine the causal effect, if any, of an alter adopting a certain health-related behavior or experiencing a certain outcome (e.g., obesity, heart attack) on an ego adopting or experiencing a similar behavior or outcome. Because the correlations between characteristics of egos contain important information on how effects propagate across a population, such models offer the potential to further the scientific understanding of network effects. (2) To develop methods for longitudinal analysis of observations made on distinct groups of connected actors (e.g., dyads, triads). For example, suppose that distinct dyads are defined based on marriage of two individuals; it may be that a property of the tie, such as the quality of the marriage (e.g., measured by strength, mutual affection, time spent together per day), is in turn related to the actors' obesity, the occurrence of health shocks, or the obesity genes in the partners. Although there is a similarity to egocentric analysis, the dependent variable and possibly some of the independent predictors here are measured on groups of connected actors rather than the individual actors. (3) To develop methods for modeling the transition of dyadic data across time as a function of attributes of the actors and of network characteristics (e.g., clustering, transitivity). Here, the dependent variable is defined for all potential dyads whether they exist or not. For most substantive analyses, the dependent variable will be an indicator of whether a tie exists at a given time, in which case we model the transition of the dyad between connected and unconnected states. However, we will also develop methods for the case where the dependent variable is more general (e.g., a count such as the number of patients shared between any two physicians in a network, or some other continuously-valued measure).

我们建议开发新的统计方法，以更精确地估计一个人的影响 在网络中的另一个中，测试和控制选择效应，例如在 个人。这项工作具有挑战性，因为网络数据可能包含多种类型的信息， 包括网络拓扑，节点协变量，TIE特征和时间变化。中心问题 正在考虑出现的复杂相关结构，因为网络中的每个演员都可以玩 自我（评估者或响应者）和Alter（目标或刺激）的双重作用，因此可能出现在数据中 多次。此外，结果可能会随着时间的推移而在地理上相关并相关 纵向遵循受试者。在这里，我们专注于开发统计方法论 纵向分析为此提供了获得因果推论的最佳机会。但是，我们也是如此 提出涉及网络横断面分析的创新。我们有三个具体的目标：（1） 开发用于以自我为中心数据的纵向分析的方法。这种分析的目的是 确定通过某种与健康相关的行为或经历A的因果效应（如果有任何） 对采用或经历类似行为或经历类似行为的自我或 结果。因为自负特征之间的相关性包含有关如何 效果在人群中传播，这种模型为进一步的科学理解提供了潜力 网络效应。 （2）开发用于对不同群体进行观察的纵向分析的方法 连接的演员（例如，二元组，三合会）。例如，假设根据 两个人的婚姻；可能是领带的财产，例如婚姻质量（例如， 通过力量，相互感情，每天度过的时间衡量），反过来又与演员的肥胖有关 健康冲击的发生，或伴侣中的肥胖基因。虽然与 以主为中心的分析，因变量以及这里可能的某些独立预测因子是 以连接的参与者而不是单个参与者的群体进行衡量。 （3）开发用于 建模跨时间的二元数据的过渡，是参与者和网络的属性的函数 特征（例如聚类，传递性）。在这里，为所有潜在二元组定义了因变量 它们是否存在。对于大多数实质性分析，因变量将是 是否在给定时间存在一条领带，在这种情况下，我们建模了连接之间的二元组的过渡 和未连接的国家。但是，我们还将开发因变量的情况 是更一般的（例如，诸如任何两位医生之间共享的患者人数 网络或其他一些连续值的度量）。