Suomen Biostatistiikan Seura ry      



John Aspegrén
Orion Oyj Orion Pharma
PL 425
20101 Turku

sihteeri @


SAS®/GLIMMIX for Mixed Models
and Generalized Linear Mixed Models

May 23-25, 2012
University of Turku, Finland

The Finnish Society of Biostatistics (FSB) is organising a 2-day lecture course + optional 1 day computer workshop on SAS®/GLIMMIX for Mixed Models and Generalized Linear Mixed Models. There are limited number of seats available in the workshop and they will be assigned in a first come first serve basis.

The course is given by Prof. Walter W. Stroup, PhD, University of Nebraska, Lincoln.

This course surveys generalized linear mixed model methods and their implementation with SAS® PROC GLIMMIX. Our focus will be on model construction, estimation and inference in studies with various forms of clustering, including repeated measures, split-plot and multi-location studies. Mixed models for normally distributed data will be presented to review important concepts, but our focus will be on generalized linear mixed model for non-normal data (e.g. proportions and counts) The course will include use of mixed model methods for planning, design and sample size determination as well as data analysis. Planning and design are especially important because much of the conventional wisdom about design follows from normal theory; designs well-suited for normally distributed data are often poorly suited for non-normal data. All example will use SAS procedures, mostly PROC GLIMMIX, but some PROC NLMIXED (for certain GLMMs that cannot be implemented using GLIMMIX) and PROC MIXED and GENMOD (mostly for comparison with GLIMMIX). Attendees should have background in design and analysis of experiments.

The course is supported by the Academic Program of SAS Finland and the University of Turku.

The first announcement can be found here.


The lectures (23-24.5.2012) will start at 9:00 and end at 17:00. The computer exercises (25.5.2012) will start at 8:00 and end at 16:00. See venue below for details on the lecture rooms.

    Lectures May 23-24, 2012
     9:00 -  9:30 Registration (day 1) + coffee/tea
     9:30 -  9:40 Opening words, Timo Hurme
     9:40 - 10:30 Lecture
    10:30 - 10:45 Break
    10:45 - 12:00 Lecture
    12:00 - 13:15 Lunch
    13:15 - 14:30 Lecture
    14:30 - 15:00 Coffee/tea
    15:00 - 17:00 Lecture

    Workshop May 25, 2012
     8:00 -  9:30 Workshop
     9:30 - 10:00 Coffee/tea
    10:00 - 11:30 Workshop
    11:30 - 12:30 Lunch
    12:30 - 14:00 Workshop
    14:00 - 14:30 Coffee/tea
    14:30 - 16:00 Workshop

Outline as PDF

I. Essential Elements of a Statistical Model: What a Linear Model Must Do
  A. Historical setting
   i. “General” linear model once meant
   ii. Challenge I: random model effects
   iii.Challenge II: distribution of y not Gaussian
   iv. Challenge III: correlation among model effects
  B. How a model is constructed.
   i. Old approach: observation = systematic + random
   ii. New: The “probability distribution” form
   iii. Essential processes the model accounts for & how this works.
   iv. The ANOVA-modeling connection: “What Would Fisher Do?” (WWFD – a process I’ve developed to aid model construction)

II. Estimation and Inference Essentials
  A. Essential inference background
   i. Issue unique to generalized linear models: model vs. data scale
   ii. Issue unique to models with random model effects: broad vs. narrow inference space
   iii. Issue unique to models with “generalized” and “mixed” both present: conditional vs. marginal inference
  B. Estimating equations
   i. pseudo-likelihood
   ii. Integral approximation: Laplace and Gauss-Hermite quadrature
   iii. REML and ML covariance parameter estimation
  C. Inference
   i. estimable and predictable functions
   ii. standard errors and test statistics
   iii. degree of freedom and bias correction issues
    1. Satterthwaite’s approximation
    2. the Kenward Roger correction
    3. sandwich (empirical)(robust) estimators
   iv. inference for covariance parameters: likelihood ratio tests, fit statistics and interval estimation

III. Power and Sample Size
  A. use of GLMM to compare competing designs
   i. different design, same sample size often => very different power characteristics
   ii. it’s not just how many observations but how wisely you deploy them
  B. use of GLMM to assess power for designs intended to be used with non-Gaussian data
   i. how it works
   ii. examples when primary response variable with be binomial and count
   iii. take-home fact: design requirements when the data are non-normal are often very different from the requirements for normally distributed data

IV. Applications with Specific Types of Response Variables and Distributions
  A. Rates and Proportions
   i. binary data
   ii. binomial data
   iii. multinomial data
   iv. continuous proportions: the beta distribution
  B. Count data
   i. Poisson data
   ii. potential issues with Poisson data: focus on overdispersion
   iii. Negative Binomial data
   iv. Too many zeros: Zero-inflated and Hurdle models
  C. Time to Event Data
  D. What we know and don’t know: Some gray areas and research work in progress

V. Repeated Measures & Spatial Correlation.
  A. Review of Mixed Model methods for normally-distributed repeated measures and spatial correlated error data
  B. GLMMs for non-normal repeated measures data –focus on Binomial and Poisson case here
   i. “PROC MIXED analog” approach => R-side (GEE is a special case).
   ii. “What would Fisher do?” approach => G-side repeated measures model.
  C. R- and G-side models are not the same
   i. what the differences are
   ii. why it matters
  D. Choosing a covariance model
  E. Choosing the standard error
   i. model-based vs. sandwich estimators
   ii. bias control
  F.Spatial GLMMs
  G. Alternative to covariance modeling: radial smoothing

VI. Overdispersion and Temporal & Spatial Correlation Revisited
  A. Overdispersion
   i. review of basics
   ii. the scale parameter and two-parameter non-normal distributions (e.g. negative binomial, beta, gamma)
   iii. What we do and do not know: future research
  B. GLMMs for Temporal and Spatial Correlation
   i. review of basics
   ii. issues unique to two-parameter non-normal distributions
   iii. What we know and do not know: future research

Course dinner

The course dinner is at 19:00 on Wednesday 23.5.2012. The name of the restaurant will be announced before start of the course. Please enter your preference (meat/fish/vegetarian) when registering to the course.


The registration deadline is May 11, 2012. To register please go to the registration page.

Participant's fees include
- admittance to the course,
- course materials,
- coffee/tee (two coffee breaks/day),
- course dinner (23.5.2012).

NOTE! The early bird registration deadline is April 30, 2012.

Registration fees: The Finnish Society of Biostatistics (FSB) members** Non members
On or before April 30, lectures 150 EUR 175 EUR
On or before April 30, lectures + workshop 225 EUR 275 EUR
After April 30, lectures 175 EUR 200 EUR
After April 30, lectures + workshop 250 EUR 300 EUR

** FSB = Suomen biostatistiikan seura


The course will take place at the University of Turku. The seminar and workshop will take place in the Publicum building (T50 in upper part of the map), Assistentinkatu 7. The lectures will be held in room PUB1 (1st floor) and the workshop in room ATK409 (4th floor). Publicum is situated near the other Universities in Turku and Turku University Hospital, within walking distance from Kupittaa railway station (1.5 km) and about 3 km from the central trainstation and bus stations (see map).

For more information about the course, please contact:
   - John Aspegrén, sihteeri @
   - Timo Hurme, timo.hurme @


Updated: May 9, 2012 | sihteeri @

Suomen Biostatistiikan Seura ry
Biostatistiska Föreningen i Finland rf
The Finnish Society of Biostatistics
© Copyright Suomen Biostatistiikan Seura ry   |   |   sihteeri @