Nimble, simple, and not so simple solutions to complex survey problems: Practical experience from tobacco control

Public health has benefited greatly from valuable information provided by national large-scale probability health surveys. For efficiency these large scale surveys use complex sampling designs. The analysis phase requiring sophisticated software must take into account these designs. Software development has progressed making now commercially available some complex survey software such as SUDAAN and WESVAR. More recently, SAS has developed procedures to address common analytical needs. However, despite this progress there are issues that arise that can not be easily addressed using commercially available software. For these issues simple and nimble solutions are needed. This presentation will use the National Cancer Institute's experience with its Tobacco Use Supplement (http://riskfactor.cancer.gov/studies/tus-cps/) to the Census Bureau's and Bureau of Labor Statistics' Current Population Survey to illustrate a number of issues that have arisen in its attempt to evaluate its ASSIST project and to further address important tobacco control research, monitoring, and evaluation questions. For example, the presentation will include how to overcome the problem of modeling multi-level data for which there is not commercially available software. Other issues that arise with complex surveys to be presented include: how to handle weights and their corresponding variance estimates when merging years of data (1992-2007) with varying number of replicate weights; how to adjust weights and variance estimates for partial longitudinal follow-up of cross-sectional survey samples; and how to bridge race estimates when there are secular changes in the number of race responses accepted. Analysis of mixed mode designs will be briefly mentioned.