Challenges and Strategies for Analysis of Complex Survey Data When Statistical Methodology and Software Package Are Underdeveloped

• Activity Number: 98
• Type: Roundtables
• Date/Time: Monday, August 5, 2013 : 7:00 AM to 8:15 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Yan Ma*+
• Companies: Hospital for Special Surgery--Weill Medical College of Cornell University
• Keywords: complex survey data ; logistic regression ; missing data ; propensity score ; model selection ; goodness-of-fit

Abstract

Federally funded large-scale health surveys offering ample information about comparative effectiveness of treatments and costs of health care have become increasingly popular in epidemiological and medical research. A substantial amount of funding opportunities seeking methodology development or encouraging the use of large-scale surveys is available (e.g., PA-09-070:AHRQ Health Services Research). These databases are often built from complex survey data collected through multiple-stage sampling designs. However, the standard statistical analytic methods do not take into account the complex survey designs and may produce invalid inference. This roundtable will discuss the challenges in analysis of complex survey data due to limited statistical methodology and software packages from the perspective of my experience. In particular, I will focus on logistic regression in analysis of complex survey data. Under this framework, topics include propensity score matching and methods for analyzing matched pairs, missing data, model selection, test of goodness-of-fit, etc. Existing statistical methods and software packages will be introduced, and limitations in these methods will be discussed.