Excel is a very friendly and convenient platform for displaying data
tables and performing simple summary statistics and graphs. Second
moment statistics like VAR, STDEV, etc. do not have a numerically good
implementation in Excel versions prior to 2003, but you can substitute
DEVSQ(data)/(COUNT(data)-1) for VAR(data), ... to get around this.
Moreover, the numerical problems with older versions will rarely be an
issue in univariate calculations unless you have a very large data set,
or a very small coefficient of variation (called relative standard
deviation in some circles).
LINEST will do multiple linear regression, polynomial fits, etc., but
Excel versions prior to 2003 can have numerical difficulties. Stats
packages will include a number of diagnostic tests for linear (in the
unknowns) models that are not native to Excel.
Fitting a polynomial trendline on an "XY (Scatter)" chart is incredibly
good (numerically) in all versions of Excel. It is more accurate than
SAS, S-PLUS, R, Minitab, etc. (sorry, I don't use SPSS). If all you
need are the coefficients, and not the diagnostics, Excel is great for
this purpose.
You can do ANOVA with LINEST, but only if you know enough about the
underlying linear algebra to force it.
Beyond Linear Models, Excel has very little capability that you would
find in stats packages.
Excel's probability functions are probably adequate for simple
hypothesis testing, but neither as accurate nor as robust (working over
a broad range of input parameters) as what is provided in stats
programs. Excel 2003 is a big improvement, but didn't go far enough.
Ian Smith's VBA library of probability functions
http://members.aol.com/iandjmsmith/examples.xls
is actually better than what is provided in any stats package I know of.
Overall, Excel is a reasonable choice to distribute specific statistical
calculations (that are within its capabilities) to a number of users
that have Excel, but do not have (or do not know how to use) stats packages.
If you want to do a variety of analyses and have access to a stats
package, that would be a better choice. If you don't have access to a
stats package, and the cost is prohibitive, consider R,
http://www.r-project.org
which is a free-ware implementation of the S language.
Jerry