SATURDAY, JUNE 7 - 10:00 a.m. - 5:00 p.m.

1. Introduction to Hierarchical Modeling for Health Services Research

Faculty: Sharon-Lise Normand and Alan Zaslavsky, both from Harvard University

Hierarchical models (also known as multi-level models) are an increasingly popular method for analysis of clustered data (e.g., patients grouped by doctors who are grouped by hospitals) and longitudinal data (multiple observations on a subject over time). In this course, speakers will introduce basic principles of hierarchical modeling. Using examples from health services research, they will illustrate use for estimation of individual and contextual effects, profiling of health care units, and attributing variation in outcomes to levels of a multi-level system. Approaches to sample size calculations will be described. Software available for fitting these models (SAS and MLwiN) will be introduced.

Level: Familiarity with regression modeling at the level of a master's degree in quantitative methods.

2. Introduction to Missing Data Methods

Faculty: Recai Yucel, University at Albany, SUNY

Statistical analyses of multivariate data are often complicated by arbitrary missing values. Advances in the computational and theoretical statistics over the last two decades have resulted in a numerous procedures with a sound statistical basis for analyzing multivariate data with missing values. Despite these advances, many misconceptions and unsound practices exist among practitioners analyzing multivariate data missing values. The first part of this talk will frame the missing-data problem, and review old and new methods. Methods specifically tailored towards missing-data problems (e.g., Multiple imputation approach) and available software (e.g., SAS PROC MI, IveWare, NORM, PAN, MLwiN etc.) will be presented in the second part of the talk. Throughout this talk, the relevant concepts and software will be demonstrated using examples from health services research.

Level: Intermediate, familiarity with data analysis at the level of a master's degree in quantitative methods.

3. Analysis of Complex Survey Data

Faculty: Brady West, University of Michigan

This seminar will provide participants with an introductory overview of issues frequently encountered when conducting secondary analyses of data collected from sample surveys with complex multi-stage designs [e.g., Panel Study of Income Dynamics (PSID), National Health and Nutrition Examination Survey (NHANES), National Comorbidity Survey (NCS)], including design-based weight determination, software choice, and proper analysis methods. The seminar is not intended for participants looking to design a survey, but rather for participants who have a desire to analyze complex sample survey data. Complex sample design concepts that motivate the analytic methods will be introduced, but are not the main focus of the seminar. Topics to be covered include recognizing a sample with a complex design, calculation of sample weights based on sample designs/non-response/post-stratification, calculation of new weights for subgroups/longitudinal analyses, weighted vs. unweighted analyses, calculation of correct confidence intervals for sample estimates, hypothesis testing based on sample estimates, design Effects, current software packages capable of complex sample survey data analysis, common analysis methods (linear modeling, descriptive statistics), and interpretation of results, and approaches to handling missing data. Several examples will be considered using real survey data sets, and software code in a variety of packages that can be used to replicate the examples will be provided. All data sets will be publicly available, and a comprehensive list of references will be provided to participants.

Level: For health services researchers looking to conduct a thorough analysis of a data set collected from a survey with a complex sample design.

Prerequisite: Some experience using the SAS or Stata statistical software packages for data analysis is desirable, but not required. A background in applied statistical analysis, including multiple regression analysis and hypothesis testing, is recommended.