BIOST/CSSS/STAT 529

Sample Survey Techniques (3)

Winter 2008

Instructor: Thomas Lumley <tlumley@u>

Lectures: MWF 9:30-10:20am. Health Sciences Building  HST T360

Office hours: After class Monday.  Thursday 9am-10am.  HSB F656

Email: Questions by email are welcome. They will often be answered quite quickly, but this is not guaranteed. In particular, I don't always check email over the weekend.

Overview: The course will cover the statistical design and analysis of complex 

surveys, with applications in the social sciences and health sciences. 

This is an applied statistical methods course, using as little 

mathematics as I can get away with.  In addition to traditional topics in 

survey analysis we will cover data visualization, regression modelling of 

data from complex surveys, the design and analysis of two-phase samples 

from existing cohorts, and links between survey analysis and causal 

inference methods.  We will not cover item-response theory and will spend 

little time on implementation issues.

Prerequisites: Students must have taken an graduate-level introductory 

course in applied statistics, and a regression modelling course is 

recommended. Knowledge of R or Stata would be helpful.

Textbook: Lohr S (1999). Sampling: Design and Analysis. [available at the 

Health Sciences branch of University Bookstore]

Assessment: Midterm exam, final poster project, and some homeworks.

Textbook:  "Sampling: design and analysis" by Sharon Lohr.

Learning Objectives:

After successfully completing this course, students should ordinarily expect to be able to:

• Define a probability sample and explain its importance in statistics. 

• Describe situations where a probability sample can lead to greater precision  that an attempt at complete enumeration.
• Distinguish finite-population and superpopulation inference and give examples where each would be appropriate
• Determine whether a survey design uses a probability sample.
• Define common features of complex surveys --
• strata, 
• clusters, 
• unequal sampling probabilties
•              and explain how they affect the cost of the survey and precision of estimates
• Write down the Horvitz--Thompson estimator of the population total and explain to a non-statistician why it gives an unbiased estimate.
• Compute summary statistics and fit regression models to data from complex surveys  using R or Stata. Describe these analyses in language suitable for an academic paper in a health sciences or social sciences journal
• Explain why assumptions about the distribution of data are not relevant to standard survey  inference and what criteria are relevant for choosing summary statistics and models  
• Define post-stratification and raking, and explain how they can increase precision
• Describe some strategies for mitigating the bias from non-response.
• Explain the advantages and disadvantages of including sampling weights in a  regression model
• Describe case-cohort, case-control, two-phase case-control, and countermatching designs for sampling from a cohort, and how data from these design can be analysed.