Friday, April 12, 2019; 2:00 PM - 4:00 PM

Session Slides:

Session Audio/Video Recording (Access restricted to registered students):

Required Reading:

Start with the Murray and Wawer Readings regarding cluster randomized trials.  They were recommended by Dr. Oldenburg.  The other four papers skim for now; they will be very relevant as we move into more about multilevel modeling.  

1. Singer, J. D. (1998). "Using SAS PROC MIXED to fit multilevel models, hierarchical models, and individual growth models." Journal of Educational and Behavioral Statistics 24(4): 323-355.
2. Arcaya M (2013) "Effects of Proximate Foreclosed Properties on Individuals’ Weight Gain in Massachusetts, 1987–2008"  Am J Public Health
3. Lee (2012) "Length of Inpatient Stay of Persons With Serious Mental Illness: Effects of Hospital and Regional Characteristics" Psychiatric Services 63, pg 899.
4. Greenland S (2000) Principles of multilevel modeling. Intl J Epidemiology. 29: 158.
5. Murray et al (2018)  Design and analysis of group-randomized trials in cancer: A review of current practices. Preventive Med 111. Pg. 241.
6. Wawer et al (1999) Control of sexually transmitted diseases for AIDS prevention in Uganda: a randomised community trial 353. pg 525.
   

Singer et al is a classic and brilliant article by one of the great popularizers of multilevel models.  It is worth reading several times.  Arcaya and Lee are examples of common applications of multilevel models to illustrate the types of questions people approach with these models.  Lee in particular illustrates how the lowest unit of observation does not need to be an individual. 

Greenland's framing is unusual but extremely helpful because it makes the link between multilevel models and Bayesian frameworks. 

Optional Reading:

Assignment: Find any article using clustered data and describe: the unit of clustering; the hypothesized effects and the level at which the exposure is measured (is it a characteristic of the cluster or the observation within the cluster); and the statistical model used to estimate the effect.  Describe whether there are any other statistical models that might be appropriate and whether they would be preferable (e.g., GEE vs mixed).

Additional Assignment for Data Application IS: 

1) Register to use data at: https://usa.ipums.org/usa/.  

2) Download a data set from the 2000 Census 5% sample, including at a minimum basic demographics, year of birth, education, state (FIPS), and all available disability variables.  Feel free to include anything else you'd like.  You can also restrict by age or another variable to make the data set smaller (use sample case selection). 

3) Open the data set in the statistical software of your choice.  I suggest that you write your code to pull a 1% sample of the data, so you can manipulate it.  There are many ways to complete this selection, for example generate a random number with a uniform distribution and keep only those observation with random number <.01.  You may find that you cannot open the data set at all on your computer.  If so, go back to the IPUMS data set and select "customize sample sizes" and change the requested density.  

4) Create a variable that is the % of people with <=6 years of education in each state (state low education).  

5) Estimate a linear regression (not accounting for clustering by state) using education and state low education predicting self-care disability. 

6) Estimate a mixed model with random intercepts for state and no fixed effects.  

7)  Estimate a mixed model with random intercepts for state and own education as a fixed effect.  

8)  Estimate a mixed model with random intercepts for state, own education as a fixed effect, and state low education as a fixed effect.

9)  Next week we will ask you to write a summary of what you have done.  For this week, just try it and figure out all of the problems you will have getting the data, handling the data, specifying and interpreting the models.