home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!paladin.american.edu!auvm!HILBERT.MATHS.UTAS.EDU.AU!MCPHERS
- X-Mailer: ELM [version 2.2 PL11]
- Message-ID: <9207290122.AA27829@hilbert.maths.utas.edu.au>
- Newsgroups: bit.listserv.stat-l
- Date: Wed, 29 Jul 1992 11:22:32 EST
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: "D. Glen McPherson" <mcphers@HILBERT.MATHS.UTAS.EDU.AU>
- Subject: Association and interpretation in log-linear modeling
- X-To: "stat-l list" <stat-l@vm1.mcgill.ca>
- Lines: 104
-
- Given models based on a three-way table of frequencies, where the variables
- are designated A, B and C, it is intuitively reasonable that the table formed
- by collapsing over levels of C to give an A x B table is not generally
- meaningful if there is a three-way association between variables A, B and C.
- The reason is clear - a three-way association implies that the relation
- between A and B varies across the levels of C. Hence the association
- portrayed in the collapsed AxB table is a composite measure of association.
- Not only is information about the AxB association lost by the combination,
- but the composite measure of the AxB association is dependent on such,
- commonly irrelevant, information as the relative numbers of sample members at
- each level of C.
-
- My problems arise when there is no three-way association between A, B and C,
- but the simplest model which fits the data must include all two-way
- associations, i.e. AxB, AxC and BxC. A number of authors establish that, in
- such circumstances, the parameter(s) defining the AxB association under the
- model which fits all three variables, are not the same parameters as those
- which apply when only the two-variable model for A and B is fitted to the
- collapsed table. (Technically, the three-way table is not 'strictly'
- collapsible' in this case). I am comfortable with the distinction - under the
- model based only on the AxB table, the parameters measure the association
- between A and B, whereas under the model based on the AxBxC table, the
- parameters measure the 'partial' association between A and B adjusting for
- relationships of A and B with C.
-
- My problem is how do I estimate these 'partial' association parameters? From
- my reading, none of the books which counsel against collapsing the table over
- C to estimate the parameters defining the AxB association, actually tell me
- how to estimate these parameters. It is as though most authors in this area
- never venture beyond the world of academia, and they are quite content to say
- 'Stop the analysis if you reach this situation!'.
-
- Can someone point me to a reference which might throw some light on the way
- to estimate these parameters, please?
-
- My ultimate test of the usefulness of methodology is its practical
- application. This is one example of many difficulties I have in attempting to
- develop skills in the practical application of log-linear methodology for my
- non-statistics majors, and in interpreting results of analyses for consulting
- clients. Do others have the same problem or am I missing out on what should
- be intuitively obvious? Following is my non-fashionable way of viewing model
- formation and interpretation. I would appreciate comments on the approach.
-
-
- The usual parameterizations employed with linear models are both intuitively
- appealing for non-statistics majors and can be quickly introduced. By
- contrast, standard log-linear parameterizations are not easily grasped by
- non-mathematically oriented students, and require both a lengthy introduction
- and a long absorption period. The fact that log-linear models come within the
- framework of Generalized liner models, has led to the presentation of the
- linear predictor form of presentation of the equations. It seems to me that
- the multiplicative form is more appealing since it is a simple extension of
- the notion of probabilistic independence which is easily motivated. Thus, in
- a model based only on the two variables, A and B,
-
- the independence model includes the equation
-
- Pr(A=i and B=j) = pi(A)i x pi(B)j
-
- where pi(A)i is the marginal probability for variable A, and pi(B)j is the
- marginal probability for the variable B. If the variables are not
- independent, there is need to adjust for this fact by introducing 'AxB
- association' parameters delta(AB)ij, i.e.
-
- Pr(A=i and B=j) = pi(A)i x pi(B)j x delta(AB)ij
-
- This is the analogous development to the introduction of interactions in
- linear models, and, as such, sits comfortably with students who have
- previously been introduced to the concept of interactions in additive
- equations. The test of no association becomes a test that all delta(AB)ij =
- 1.
-
- The approach extends simply and naturally to three variables. Thus, pairwise
- independence among all pairs is represented by the equation
-
- Pr(A=i and B=j and C=k) = pi(A)i x pi(B)j x pi(C)k.
-
- The inclusion of a component which allows for an association between A and B,
- but independence between the other pairs, is represented by the equation
-
- Pr(A=i and B=j and C=k) = pi(A)i x pi(B)j x pi(C)k x delta(AB)ij.
-
- Further associations are readily and obviously introduced with the inclusion
- of additional multiplicative components.
-
- The role of model comparisons in hypothesis testing is readily tied to
- components in the multiplicative equations.
-
- Given a model which fits the data, the matter of practical interpretation can
- follow. As part of this process, reparameterization can be discussed . The
- interpretive value of considering odds as ratios of 'pi' components, and odds
- ratios based on 'delta' components can be introduced - for their practical
- value rather than as mathematical artifacts.
-
- Do others share my concerns about teaching the basis and application of this
- methodology to non-statistics majors? Am I missing something - can the
- log-linear approach be made intuitive and easy?
- --
- ---------------------------------------------------------------------------
- | Glen McPherson | | _--_|\ |
- | Department of Mathematics |-----------------------------------| / \ |
- | University of Tasmania | E-Mail: | \_.--._/ |
- | Australia. | mcphers@hilbert.maths.utas.edu.au | * |
- ---------------------------------------------------------------------------
-