Markus Quandt wrote:
>
> First, thanks for your reply!
>
> Bob Hayden schrieb:
>
> > This may not fully answer all your questions, but the various formulas
> > for inference for regression have a place where you plug in THE
> > variance of the points around the regression model, i.e., they treat
> > this as a constant. If it is not a constant, the results of the
> > formulas will not be correct. Specific ramifications would depend on
> > what formual you are interested in, the exact manner in which the
> > variance varies, etc.
>
> Yes, I'm aware that non-constant variances let the usual computational rules break
>down if you really want to
> incorporate them in the model. But my problem is: Why do I want to consider them?
> As stated below, there are really two questions:
> 1) Why do the standard errors *increase* -- is there a intuitive explanation (more
>intuitive than the proof
> that the estimate of b is not efficient)? So, I think the specific formula of
>interest is either the
> calculation of b or -- more probably -- of its SE. Subquestion: Why does OLS fail to
>reflect this tendency to
> have larger SEs?
This is a read Herring. Increase compared to what? OLS applied
heteroskedastic data vs OLS applied to homoskedastic data? This is a
pointless comparison. The data has different properties and the
estimates of b will have different variances - smaller or greater
depending on how you "cook up" the heteroskedasticity.
What matters is that, *if* the data is heteroskedastic, we can do better
with WLS. For heteroskedastic data the variance of b_OLS is greater than
b_WLS.
This inefficiency is really the only problem with the OLS estimate of b.
And this is in most cases a minor problem.
The real problem is the associated estimate of the variance of b_OLS
*when* calculated under the assumption of homoskedasticity. (Which is
what any regression routine will do unless you explicitly tell it to do
something different.) This estimate of the variance will be biased and
inconsistent. The bias could be positive or negative depending on the
form of the heteroskedasticity and how it interacts with the explanatory
variables (Kmenta's equation 8.16a).
Note that this only matters if we want to conduct tests on b or
calculate confidence intervals.
In many cases (but not all) the error variance tend to be larger for
large values of the dependent variable. If b is positive the error
variance is large for large values of the explanatory variables, they
are positively correlated and the second term in 8.16a is positive. The
first term in 8.16a approximates the estimate of the variance of b_OLS
assuming homoskedasticity. That is, the standard estimate of the
variance underestimates the true variance of b_OLS in this (common)
case. This is probably where the (misconceived and common) notion of
increasing variances comes from.
> 2) Regarding Kmenta's argumentation: Inefficiency of OLS is proved by relating it to
>WLS' better performance.
> This *assumes* that weighting down the observations with higher residuals is
>legitimate. How do I justify this
> strong assumption, apart from the case when I know that the larger residuals result
>from higher *true* random
> disturbances? In other words: If residual diagnostics show uneven distribution of
>residuals, shouldn't be the
> first step to search for unobserved predictors?
Yes, that is always a good thing to do.
And if I cannot measure these, isn't OLS as good a guess at the
> true relationships of the observed variables as WLS would be?
This we will never know. Both will in all likelihood be wrong, but which
is most wrong and are they wrong to the extent that it matters?
With a resonably large sample a reasonably safe procedure is to estimate
b with OLS and use a heteroskedasticity robust estimate of the variance
of b_OLS. This foregoes efficiency but inference is (asymptotically)
valid.
A common robust variance estimator is the one of White (Econometrica,
1982?). White also has a test for heteroskedasticity and one can base
the decision on which variance estimator to use on the outcome of this
test.
/Sune
>
> (Because Mr. Hayden started a new thread, I do not delete his quotation of my
>original message, in order to
> retain the context.)
>
> Regards, MQ
>
> > ----- Forwarded message from Markus Quandt -----
> >
> > Hello all,
> >
> > when discussing linear regression assumptions with a colleague, we
> > noticed that we were unable to explain WHY heteroscedasticity has
> > the well known ill effects on the estimators' properties. I know
> > WHAT the consequences are (loss of efficiency, tendency to
> > underestimate the standard errors) and I also know why these
> > consequences are undesirable. What I'm lacking is a substantial
> > understanding of HOW the presence of inhomogeneous error variances
> > increases the variability of the coefficients, and HOW the
> > estimation of the standard errors fails to reflect this.
> > I consulted a number of (obviously too basic) textbooks, all but
> > one only state the problems that arise from het.sc. The one that
> > isn't a total blank (Kmenta's Elements of Econometrics, 1986) tries
> > to give an intuitive explanation (along with a proof of the
> > inefficiency of the estimators with het.sc.), but I don't fully
> > understand that.
> > Kmenta writes:
> > "The standard least squares principle involves minimizing
> > [equation: sum of squared errors], which means that each squared
> > disturbance is given equal weight. This is justifiable wehn each
> > disturbance comes from the same distribution. Under het.sc.,
> > however, different disturbances come from different distributions
> > with different variances. Clearly, those disturbances that come
> > from distributions with a smaller variance give more precise
> > information about the regression line than those coming from
> > distributions with a larger variance. To use sample information
> > efficiently, one should give more weight to the observations with
> > less dispersed disturbances than to those with more dispersed
> > disturbances." p. 272
> >
> > I see that the conditional distributions of the disturbances
> > obviously differ if het.sc. is present (well, this is the
> > definition of het.sc., right?), and that, IF I want to compensate
> > for this, I can weight the data accordingly (Kmenta goes on to
> > explain WLS estimation). But firstly, I still don't see why
> > standard errors increased in the first place... And secondly, is it
> > really legitimate to claim that OLS is 'wrong', if it treats
> > differing conditional disturbances with equal weight?
> >
> > Assume the simple case of increasing variances of Y with increasing
> > values of X, and therefore het.sc. present. With differing
> > precision of prediction for different X values, the standard error
> > (SE) of the regression coefficient (b) should become conditional on
> > the value of X, the higher X, the higher SE, with E(b) constant
> > over all values of X - correct? Then, isn't the standard error as
> > estimated by OLS implicitly an _average_ over all these conditional
> > SEs (just following intuition here)? How can we claim that the
> > specific SE at the X value with the lowest disturbance is the
> > 'true' one? (Exception: het.sc. is due to uneven measurement error
> > for Y - I can see that the respective data points are less
> > reliable.)
> >
> > Regarding the first question: Can this be answered at all without
> > the formal proof?
> >
> > ----- End of forwarded message from Markus Quandt -----
>
> --
> ________________________________________________________________
> Markus Quandt
--
Sune Karlsson | Fax: + 46 8 34 81 61
Stockholm School of Economics | Phone: + 46 8 736 92 39
Box 6501, 113 83 Stockholm, Sweden | http://www.hhs.se/personal/SuneK/
http://www.hhs.se/stat/ | http://swopec.hhs.se/
===========================================================================
This list is open to everyone. Occasionally, less thoughtful
people send inappropriate messages. Please DO NOT COMPLAIN TO
THE POSTMASTER about these messages because the postmaster has no
way of controlling them, and excessive complaints will result in
termination of the list.
For information about this list, including information about the
problem of inappropriate messages and information about how to
unsubscribe, please see the web page at
http://jse.stat.ncsu.edu/
===========================================================================