Assignment 5: Causal Discovery

Your task is to verify the findinings in [Druzdzel & Glymour 1994]
[Marek J. Druzdzel and Clark Glymour. Causal inferences from databases:
Why universities lose students. In Clark Glymour and Gregory F. Cooper
(eds), Computation, Causation, and Discovery, Chapter 19, pages 521-539,
AAAI Press, Menlo Park, CA, 1999, an older version is available at:
http://www.pitt.edu/~druzdzel/ftp/kdd94.pdf.]
concerning student retention in US colleges.  The original study
was performed on 1992 US News and World Report data.  The data
that you will be studying is for the year 1993.

There are some small differences in variables under study between the
1992 and 1993 data sets.  Because freshmen retention and senior
graduation rates differred so little (most students who drop out do
so in their freshmen year), freshmen retention has been eliminated.
What used to be "apgra" is in the 1993 data set called "apret" (this
is the percentage of student retention over the four years).
(We rerun the 1992 data set and verified that the conclusion
regarding the impact of the quality of incoming students on the
retention rate is the same.)  Otherwise all variables are the
same (but the measurements are for the year 1993).

Do the 1993 data support Druzdzel & Glymour's conclusions?  If not,
why not.  Can you find anything else going on in the data (i.e.,
is there any structure in the data?)?  What are the causal graphs
suggested by GeNIe?  What causes student retention?  Please, feel
free to help GeNIe using your common sense knowledge of interactions
between the variables included in the data set (not necessarily the
knowledge inputed by Druzdzel & Glymour).  For example, you might
know something about the time precedence.  Perhaps there are pairs
of variables for which we are reasonably sure that there is/there
is not an arc between them?  Check the sensitivity of your result
to the significance level.  Druzdzel & Glymour made the assumption
that the data are normally distributed and linearly dependent (this
is an implicit assumption used in the PC algorithm in GeNIe).
Please, feel free to check whether the assumption is indeed valid
in the data using the graphing capabilities in GeNIe.

You don't need to make this a lot of work, but please, run the PC
algorithm at least a few times and look critically at the results.
Describe your observations and conclusions.

I have included the data file on the CourseWeb.

The software that you should use in this assignment is GeNIe,
available for download at the following location:

  http://genie.sis.pitt.edu/

The on-line help for learning/causal discovery functionality may
not be comprehensive but here is in brief how you can use it:

(1) Open the data file through File menu/Open Data File.
(2) Choose Learn New Network from Data menu.
(3) Choose the PC algorithm and set the significance level
    (PC algorithm uses classical statistical independence
    tests and here is where you set the significance level
    for these tests).
(4) You can enter the background knowledge by clicking the
    Background Knowledge button.  Here you can force and
    forbid causal connections and also order variables in
    temporal tiers.  A variable that is in a later tier
    cannot cause variables in earlier tiers.
(5) The result of the algorithm is a "pattern," i.e., an
    equivalence class of causal graphs that are compatible
    with the data.  The true graph is one of the graphs
    represented in the pattern.