-----Original Message-----
From: markleeds at verizon.net [mailto:markleeds at verizon.net]
Sent: 2008?10?7? 23:10
To: Hsiao-nan Cheung
Subject: RE: [R-SIG-Finance] Statistically significant in linear and
non-linear model
Hi: it's an interesting question and probably something that you
should
send to R -help also. One reasonis the following but
I don't know if it's a big one ?
if one has the linear model log(y) = X*beta + epsilon, then, this can
clearly be transformed to y = exp(Bx) and minimized using
nls. But, I think it's possible that beta can turn out significant in A)
and not in B) because of the assumption about the error term.
In A) the error term is assumed to be additive and this assumption is
used HEAVILY in standard OLS theory.
In B) The error, term if A is true, is multiplicative and, however, nls
works out the standard errors ( I guess it's estimates the Hessian
and uses that ) could cause Beta to be not significant.
And, the argument can also probably go the other way so that, one could
have significance in nls bt not OLS.
As I mentioned, I don't know what the constribution of above is to
your
question but it's a thought. If you get any offline replies,
could you send them to me because i'd be interested. I bet you would
get
a lot of responses from R-help if you sent it there also.
On Tue, Oct 7, 2008 at 10:38 AM, Hsiao-nan Cheung wrote:
Hi,
I have a question to ask. if in a linear regression model, the
independent
variables are not statistically significant, is it necessary to test
these
variables in a non-linear model? Since most of non-linear form of a
variable
can be represented to a linear combination using Taylor's theorem, so
I
wonder whether the non-linear form is also not statistically
significant in
such a situation.
Best Regards
Hsiao-nan Cheung
2008/10/07
[[alternative HTML version deleted]]