On Tue, 19 Apr 2016 17:39:15 -0000 (UTC), Herman Rubin <
hrubin@skew.stat.purdue.edu> wrote:
On 2016-04-19, Kashif Beg <kashifbeg90@gmail.com> wrote:
Ques 1: After fulling all the assumptions, the result of rgression model >>shows that all the variables in model are insignificant. However, for good >>regression model 50% of variables must be significant, is this compulsary?
After 20 years of reading the stats groups each day, this seems
to be a brand new rule of thumb. However -- I gather that you
may be coming from econometrics, and that has never been a
big topic for any of the folks who answer in this group, or for
the folks who have posted questions to us.
The quality of a regression model is in its R^2; if this is low, the
model predicts poorly. This can happen even if each variable tests >insignificant; from an action viewpoint, significance is essentially >meaningless.
Herman - You wote too fast. Low R^2 says it predicts "poorly" in the
sense of variance; odds-ratios might suggest good prediction.
With large N, small R^2 can be "significant". With large R^2, every
variable can still test not-significant if there is confounding.
The /lack/ of significance, overall, says that you don't have
something reliable. But even "strong significance" does not
say that you necessarily have something very useful, so your
conclusion is fairly apt -- that "significance" is fairly meaningless
unless you have a an appropriate quesiton.
Ques 2: In order to check consistency, i applied Polled ols, fixed
effect and random effct models of panel data, i have shown this in
similar manner as given below in result and discussion chapter, but >>interpretation is based on most appropriate model. Is this appropriate?
The meaning of consistency that you are using is not the standard one;
I am not sure what it means.
I assume that Polled should be Pooled. I have never had panel data
that I considered treating that way, nor have I read about the
problems or benefits. I assume that some econometricians will be
familiar with your terminology (which, I am not).
Ques 3 Consistency check or Robustness check is same or different?
Again, I am not sure what either term means as you are using them.
Ques 4 In order to deal with problem of heteroscedasticity
and autocorrelation, I ran the model with heteroscedasticity and >autocorrelation consistent robust standard errors (HAC) errors. But
the durbin watson values are not under the prescribed limit of 1.5 to
2.5. Is it a matter of concern. what is the solution?
Yeah, time series of data are a real bitch. "Panel" seems to describe something different from the economists' other sort of long series,
daily or monthly or yearly cross-sectional estimates. "Solution"
depends on precisely what the data are, and what the questions
are.
What assumptions are you making, and which of them are you testing?
Given a fuller description, we may or may not have useful things
to say. But I think your best bet is to find an econometrician who
is familiar with data such as yours.
--
Rich Ulrich
--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)