Multilevel/Mixed Models Using Stata
This two-day course is an introduction to using Stata to fit multilevel/mixed
models. Mixed models contain both fixed effects analogous to the coefficients
in standard regression models and random effects not directly estimated but
instead summarized through the unique elements of their variance–covariance
matrix. Mixed models may contain more than one level of nested random effects,
and hence, these models are also referred to as multilevel or hierarchical
models, particularly in the social sciences. Stata’s approach to linear
mixed models is to assign random effects to independent panels where a
hierarchy of nested panels can be defined for handling nested random effects.
The course will be interactive, use real data, and offer ample opportunity for specific research questions and for working exercises to reinforce what is learned.
The course will be taught in five parts. During the first four parts, the discussion will be confined to linear mixed models for continuous responses. The fifth part will focus on binary and count responses.
Request this course
- Part I — This part looks at the classic random-intercept linear model. We will discuss several approaches for fitting this model, along with the associated benefits and assumptions of each approach.
- What constitutes a linear mixed model?
- The random-intercept model
- The within estimator versus the generalized least squares (GLS) estimator; the Hausman test
- Maximum likelihood and restricted maximum likelihood
- Using the xtmixed and xtreg commands for the random-intercept model
- Part II — This part will focus on random coefficients and the various covariance structures that can be imposed with multiple random-effects terms.
- Adding random coefficients
- Specifying models hierarchically
- Covariance structures for random effects
- Growth curves
- Linear transformations of covariates in a random-effects setting
- Likelihood-ratio (LR) tests
- Part III — This part's theme can best be described as tricks of the trade, covering various methods for fitting more complex models, including crossed-effects models, growth curve models, and models with complex and grouped constraints on covariance structures.
- Multiple-level models
- Crossed-effects models
- Using Stata’s “R.” factor notation for mixed models
- Complex and grouped constraints on variance components
- Heteroskedastic residual errors
- Alternate residual-error structures
- Part IV — This part will consist of predictions, model diagnostics, and other postestimation tasks.
- Best linear unbiased predictions (BLUPs)
- Fit diagnostics
- Diagnostic plots
- Cataloging and comparing mixed-model results in Stata
- Part V — This part will focus on models for binary and count responses. During this part of the course, you will learn that most of what is discussed for linear mixed models can be applied equally to mixed models with noncontinuous responses.
- Binary and count responses
- Estimation via adaptive Gaussian quadrature
- Model building using the Laplacian approximation
- Predictions and other postestimation tasks
Basic knowledge of standard linear regression and a working knowledge
of Stata and the Do-file Editor.
This is a two-day course. All training courses generally run for eight hours per day and include morning and afternoon breaks and a lunch break. You can arrange a convenient schedule with your instructor and training coordinator.
Request this course