-----Original Message-----
From: Arthur J. Kendall [mailto:[EMAIL PROTECTED]]
Sent: Monday, April 08, 2002 5:05 AM
To: David Heiser
Cc: [EMAIL PROTECTED]; [EMAIL PROTECTED]; [EMAIL PROTECTED]
Subject: Re: Auxiliary regression
I think Quattro-Pro, Lotus, and Excel can be very useful to do exercises
that
demonstrate what goes on in some statistical processes. I personally use
QuattroPro, Lotus, and Excel for spreadsheet purposes. In exercises we
pretend
that data arise in full bloom like Venus from the mind of Zeus, but in
practical
applications there is a lot that goes before the data is ready to be used.
However, I have seen many instances where people have used a spreadsheet in
place of a statistical package. This resulted in:
more person-time to do the analysis;
much more person-time to review and validate the analysis,
failure to treat different kinds of missing data differently,
lack of data definition for others to read and follow,
ambiguity about what transformations were made, etc.,
_rework_.
Much depends on the context. If the purpose is to help in understanding
stat
constructs by showing some computation, spreadsheets are very useful.
Likewise
if the process does not need to be reviewed, or if there are not missing
data
considerations.
When quality assurance, data quality, and data sharing are part of the
project
as they are in evaluation (applied social science), and basic research, it
is,
IMO, preferable to learn a stat package. I consider clarity of syntax,
consistent command structure across procedures, consistent and complete data
definition, etc. of major importance especially when the study (evaluation)
is
in support of public policy decision making.
----------------------------------------------------------------------------
I agree here with you.
Excel is not a proper final tool to use in academic level research. It may
however be a useful supplementing tool. Excel data entry is easy, and the
files can be read by most of the big software packages.
Missing data is a special problem that requires extensive considerations and
software to "work-around". I recommend that Excel only be used when there is
no missing data.
The preprocessing of data is an advantage in Excel, since one can trace the
transformations. One can also use multiple workbooks to hold a variety of
trial transformations and subset collections.
However, computers are programmed in a language (C++, FORTRAN, VB, ....) and
like all languages the "words" are used in many different ways. I don't see
how any program can achieve clarity of syntax when there are so many
different ways, and so many different programmers writing modules (objects),
and they all have intricate subroutine call paths and their own peculiar way
of doing things.
My point is that there is only a small fraction of all students who take all
introductory statistics courses, that end up doing research. Excel should
work for these not doing research, since it is available.
DAHeiser
David Heiser wrote:
> -----Original Message-----
> From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]]On
> Behalf Of Jay Warner
> Sent: Friday, April 05, 2002 4:29 PM
> To: [EMAIL PROTECTED]
> Cc: [EMAIL PROTECTED]
> Subject: Re: Auxiliary regression
>
> Amen to that.
>
> the 'problems' that Excel has with statistics seems to appear with
> various types of regression on less-than-well-cornformed data. any
> regression analysis deserves as many tests of 'legitimacy,' robustness
> etc. as you can give it. SPSS is one way to get a heck of a lot more
> such tests than Excel.
> ----------------------------------------------------------------------
> Not entirely true. A lot of the comments against EXCEL are like
folk-tales,
> like the folk tale about someone putting a JATO unit on a car (The Darwin
> Awards....) someone heard it, passed it on, etc...
>
> I have been doing a paper on the problems with Excel. It is very hard to
get
> specifics with data so that I can verify the problem. Even McCullough made
> some incorrect statements.
>
> I can get Excel to do an accurate solution to FILLIP with an LRE of 10.2
on
> the coefficients, 14.5 on the F statistics and 14.3 on the standard error.
> Both McCullough and Altman report that the software they tested did not
> solve FILLIP.
>
> The MINVERSE function does a pretty good solution with near singular
> matrices. I got it to work with data having X correlations of 0.999999.
>
> Like any tool, you have to learn how to use it first. Excel is complicated
> and Microsoft does not make any effort to facilitate learning it. The
> commercial manuals really gloss over the stat area, because the authors do
> not have stat backgrounds. The commercial books/manuals "Statistics using
> Excel" etc. are grossly simplified and aimed as a primer to anybody who
> wants to learn how to use the Stat add-on package.
>
> Excel does have problems, but there are workarounds and fixes.
>
> DAHeiser
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
