PUBH 8400/1, Spring 2010:
Some ideas for class projects
1/18/10, Jim Hodges
These are only
suggestions, to give you some idea of what might be a suitable subject for your
class project.
In the list, each item has a label referring to the detailed syllabus, e.g., I A 2 refers to Part I
("Mixed Linear Models, per RWC"), Section A (Mixed linear models in the standard
formulation), number 2 (Doing statistics with this model).
(1) (IA1) For the vocal folds data, we have five outcomes analyzed with identical design
matrices;
for 4 of them, the REML maximization converges, but for the other one it doesn't.
Why?
(2) (IA2a) Standard REML-based software often produces zero estimates and does not give a
standard error.
This is the wrong answer.
For a simple case like the balanced one-way random
effects model, propose an (approximate) standard error for use with a zero estimate, and do a
small designed simulation experiment to examine its properties.
(3) (IA2a) Standard REML-based software often produces zero estimates and does not give a
standard error.
This is the wrong answer.
Do a literature review of solutions that have been
proposed for this or related problems (if any).
(4) (IA2b) Do a literature review of prior distributions that have been proposed for variance-
structure parameters in MLM, hierarchical, or related models.
(5) (IA2b) Do a literature review of the effect of prior distributions on Bayes factors in MLM.
(6) (IA3) Pick a software system, e.g., SemiPar, and work a lot of different problems in it,
preferably using datasets from the course web site.
Compare the results of different fits obtained
in SemiPar, e.g., for different bases or different ways of selecting the smoothing parameters (CV,