This is the home page of Pop 510: Multilevel Models, as offered in the Spring of 2016 (Session II). The course registrar's page is here. For Pop 509: Survival Analysis, click here. For my own research on multilevel models click here.
This half-course, offered in the second session of the spring term, provides an introduction to statistical methods for the analysis of multilevel data, such as data on children, families, and neighborhoods.
We review fixed- and random-effects models for the analysis of clustered and longitudinal data before moving on to multilevel random-intercept and random-slopes models.
We discuss model fitting and interpretation, including centering and estimation of cross-level interactions.
We cover models for continuous as well as binary and count data, reviewing the different approaches to estimation in common use, including Bayesian inference.
A list of my papers on multilevel models including abstracts and links to JSTOR where available, as well as the Rodríguez-Goldman data, may be found at http://data.princeton.edu/multilevel.
We start with simple variance-component models using data on language scores from Snijders and Boskers. Part 2 has random intercept and random slope models, and Part 3 has a model with a level-2 predictor, where the random intercept and slopes depend on school SES.
We illustrate growth curve models by replicating an analysis by Goldstein of the heights of school boys measured on nine occasions between ages 11 and 13. In addition to random intercepts, slopes and curvatures, we consider serial correlation among residuals.
The next example deals with 3-level linear models using growth curve data on 1721 students in 60 schools. We start with simple variance-component models and hen move to growth curves with random intercepts and then random slopes. Our analysis concludes with a comparison of growth curves in schools that differ in observed and unobserved characteristics.
We next move to multilevel logit models, starting with an application to longitudinal data on union membership from the NLS, focusing on a comparison of marginal and subject-specific models and the calculation of intraclass correlation for latent and manifest outcomes.
We continue with an application to contraceptive use in Bangladesh, where we consider random-intercept and random-slope models. We also illustrate the estimation of random effects using maximum likelihood and posterior Bayes estimates.
For a three-level logit model consider the analysis of immunization in Guatemala. The data are available on the multilevel section of the website and the book by Rabe-Hesketh and Skrondal has a substantial analysis.
The notes on how to run multilevel logit models using winBUGS are here, with a link to a compound document that can be run from WinBUGS. See also part 2, showing how to run WinBUGS in batch mode, and how to import CODA output into Stata for further analysis.
I also recommend you have a look at the MCMC feature in MLwiN, as demostrated in class. This is probably the easiest way to estimate multilevel models using MCMC procedures.
We now have a sample run of Stan, the latest on MCMC estimation using Hamiltonian Monte Carlo and the No U-Turn Sampler (NUTS), applied to the hospital delivery data right here.
Stata can fit some multilevel models using Metropolis-Hastings combined with Gibbs sampling. We illustrate the procedure using the same hospital delivery data used with WinBUGS and Stan and compare resuts of all methods here.
Our final application follows the analysis of infant and child mortality in Kenya which you will find in my chapter in the Handbook of Multilevel Analysis.