Further Statistical Methods -HT06
Practical - week 2
MIM and graphical models.
MIM has both a command interface and a menu based interface. Every item on
the menu typically writes a specific command on the command interface. For
information about these commands and the possibilities, use the help facility in
MIM. The help facility is invoked either by writing help on the
command line, or by using the help menu directly.
- Start MIM
- Download this file from a study of
risk factors for coronary heart disease among 1841 car-workers in the former
Czechoslovakia and look at the file. The car-workers were participating
in a health study and these are the risk factors at the entrance time of the
study. See e.g. Edwards (2002) for a more detailed description of the study.
- Read the file into MIM.
- The file specifies the saturated model. Display its graph and fit the
model.
- Test all conditional independences of any pair of variables, given the
others. Check also whether the asymptotic results seem to conform with
Monte-Carlo p-values. Use e.g. the command
testdelete with appropriate options or the "Test" menu. Try also to
use the "Select" menu with one step only. Note which conditional independence
relations are rejected.
- Find a well-fitting log-linear model by using MIM's unrestricted version of
backward stepwise selection, and display the dependence graph of the final
model.
- Display also the factor graph (interaction graph in MIM) of the final
model using the "Type" menu in the graphics window. Is this more informative than the dependence graph in this case?
- Give a verbal interpretation of the model, in particular in terms of
conditional independence.
- Experiment with different stepwise procedures and observe the different
results.
- For this size, a global model search using AIC or BIC is just feasible. Use global search
to find the best model using AIC. This does take a little time, as about 33,000 models must be fitted
and compared. How many exactly and why?
- Display the graphs of the model found by AIC, and compare the model to the
one found above under item 6, both in terms of interpretation and fit. Test
the overall fit of each of the models (use the command
fit) and the fit of the smaller
model, when the larger is assumed. For the latter, specify and fit the larger model,
make it base, then specify the smaller model and
use test.
Back to course overview.
Last updated:
Monday, 23 January 2006 15:39.
Steffen L. Lauritzen