Statistical methods for survey data in the social sciences

点击翻译按钮获取中文摘要

I will pursue research in three areas related to complex survey methodology, with common themes of latent variables and linkages. The first area is latent variable modeling in the complex survey context. Latent variables are hypothesized but unobservable variables, giving structure and meaning to a statistical model. For example, a random effect in a mixed or multilevel model is often regarded as latent; structural equation models use latent constructs to explain the relationships among observed psychometric variables; in item response theory (IRT) models used in educational testing, item difficulty and test taker ability are often taken to be latent. I will extend previous work on multilevel models, and apply estimating function approaches to robust estimation of latent variable model parameters. The second area is the analysis of data from sampling on social networks. A network is a collection of nodes, some of which are joined by links. My collaborators and I have worked on “patchwork sampling”, wherein links from a probability sample of nodes are followed to a certain distance, with multiple inclusions recorded. Applications include a collaboration network of scholars; a network of telecom customers who call each other; or a contact network with links for disease transmission. I plan to extend our preliminary work to the estimation of connectivity and complex dependences. The third area is multi-frame methods and data linkage. A frame is a database or list of members of a population, from which samples can be drawn. The members may or may not be contactable, and the frame may or may not be “rich” in the sense of providing useful information about the members. Typically the databases or samples overlap considerably, and proper linkage of the records can make analysis more efficient. If the identifiers are not unique, or if erroneous or missing identifiers make linkage uncertain, it is desirable to have techniques for automatic estimation of links. I will explore relationships among traditional multi-frame methods, record linkage methods, and sampling using linkage information. The work on latent variable models addresses two problems facing social science researchers: the need for accurate multi-level analysis when the sampling design is structurally informative and/or the models are non-normal; and the need for robustly estimable latent variable models. The rest of the proposal concerns trends in survey methods arising from the increasing availability of vast amounts of rich and linkable data on people and establishments, and the increasing difficulty of carrying out designed studies such as traditional sample surveys due to non-response. The list of potential application areas includes not only official statistics and social research methods but also forensic uses such as detection of unusual activity in communications traffic. Supervision of graduate students will be included in each area.