Inference from Multilevel Versus Single-Level Methods with Complex Survey Data: An Example Using the National Youth Risk Behavior Survey

• Activity Number: 215
• Type: Invited
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Caucus for Women in Statistics
• Author(s): Emily Olsen*+ and Shannon L. Michael and Sherry Everett Jones
• Companies: CDC and CDC and CDC
• Keywords: multilevel ; complex survey data ; government statistics

Abstract

The national Youth Risk Behavior Survey (YRBS) is conducted by the Centers for Disease Control and Prevention (CDC) biennially to monitor six categories of priority health risk behaviors among a representative sample of high school students. The complex sample was not designed to be representative below the national level; however, researchers are often interested in the effects of higher-level characteristics (e.g., a state law or policy) on lower-level characteristics (e.g., student behaviors). To simply add higher-level characteristics to a logistic regression model with a student-level outcome may lead to naïve or incorrect inference. The purpose of the study is to compare modeling procedures to analyze state-level characteristics on student-level outcomes using YRBS data. We explore the differences in multilevel models in M-Plus and single-level (i.e., non-multilevel) models using SUDAAN to analyze state tanning bed laws on student tanning bed use. We hypothesize the inference from each model may differ and aim to identify the simplest model that maintains inferential accuracy.