🚀 go-pugleaf

Phillip Phan says:

>> looking for a nice applied reference on survival analysis.

Singer and Willett have a very good introduction to survival
analysis in the latest (Summer 1993) issue of the Journal of
Educational Statistics.  Other articles by them are

   Singer and Willett (1991), Psychological Bulletin, 101(2),
     268-290.

   Willett and Singer (1991), Review of Educational Research,
     61(4), 407-450.

These focus more on applications of survival rather than statistical
aspects; any would provide an excellent introduction to use of
survival analysis in social science research.

>> why one would do path analysis using OLS regression rather than go to
>> a package like LISREL for causal modelling?

I think there might be two issues here.  One is path analysis vs.
structural equation modeling, and the other is OLS vs. maximum
likelihood (ML) or weighted least squares (WLS) estimation.

Let's take the second issue first.  OLS will usually give you similar
results to other estimation procedures.  The main advantage I can
think of with ML or WLS (or "asymptotic distribution free" methods) is
that the statistical properties of the resulting goodness-of-fit test
are better defined.  With non-OLS estimators, the chi-square
statistics for alternative models of the same data can be compared
statistically (provided the models compared are nested).  That is, the
difference in the chi-square for two nested models is itself a
chi-square statistic.  This can be very useful for testing hypotheses
such as "is the measurement errror for variables A and B the same?"
My experience is that many research hypothesis boil down to the
comparison of such nested models, rather than the absolute fit of any
particular model, so this property of non-OLS estimation is
important to know.

That having been said, I should also say that if there is no concern
for comparing nested models, I am of the humble opinion that one would
do just as well with OLS as with non-OLS methods for estimating a
given model; note that OLS does not make any assumptions about
multivariate normality or the like.

The second issue relates to the difference between standard
path/regression analysis and structural equation modeling.
I use the former term to refer to models where each variable
has only a single indicator, and the latter to refer to models
where one or more variables have multiple indicators.  Recall
that the standard definition of the LISREL model is that it
contains a structural (i.e., path or causal) model *and* a
measurement model (which gets into multiple indicators).

>> If I did a factor analysis  of my scales and then used the
>> resulting factors for the path analysis aren't I getting the same
>> thing as an output from LISREL?

You might not get the exact same thing, but it should be very
close.  There are some issues with correlated measurement error that
the method you describe might not be sensitive to.  However, I believe
the approach you describe is basically sound and perhaps in some ways to
be preferred, if what you are mainly interested in is the "structural"
part of the model and not the "measurement" part.

--John Uebersax
  Bowman Gray Medical School
  Winston-Salem, NC  27157
  uebersax@phs.bgsm.wfu.edu