Sorry, Mahmoud, but I think you're missing the key point I made twice in
earlier posts -- or perhaps hoping to believe that it can't be true.
There is some work on bootstrapping for time series -- but the commands
you are using do not implement that.
There is some work on t-tests for time series -- but the same comment
applies.
Independence versus dependence is not a small technical point that can
be solved by a trick. If it were, there would not be large areas of
statistical science focusing on
time series and spatial series.
If you apply procedures intended for independent observations to data
that are dependent, then P-values are the first to go. The P-values you
get are basically garbage. Standard errors are wrong too. This is not
well covered in introductory texts or courses, but it is true.
Eva's comments do not address this point either.
So, as far as I can see what you are doing is in principle wrong. How
far that matters in practice will depend on the correlation structure of
your time series, on which I naturally have no information.
It moves me not a bit that other people may have done what you are
proposing. If so, they were wrong too. Correctness in this field is not
decided by vote.
This is of course not a personal opinion. I've already given a reference
to Box, Hunter and Hunter, which explains the point very well.
Your reply to Tony Lachenbruch indicated that you have a paired data
structure. If so, the thing to do would perhaps be
1. Plot means for one set versus means for the other set.
2. Count the numbers on each side of equality and report the fractions.
In other words, descriptive statistics and graphics are what I
recommend.
Nick
[email protected]
Mahmoud Abd-El-Aal
i am just trying to see how many of the bootstraped
samples mean of a certain variable is larger than the mean of another
variable, and therefore i need to know from the tests that i presented
earlier is it larger or smaller since we reject the null hypotheses:
var2-var1=0; Tony, i am assuming they are paired because variable 2
depends on componenets from variable 1
the variables are the log returns of closing prices (var1) and the
moving
average rules of this return series (var2), I know there is
autocorrelation in them, but that is why am using just two forms of
processes that could take into account some of this, random walk with a
drift and the AR (1) process........
The point is all this has been done before in an academic paper, i tried
sending to the authors but no reply......i am totally aware of the
problems accompanying this kind of tests but there must be a way around
it
somehow.
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
