An asymptotic theory for weighted least-squares with weights estimated by replication (Q3774757)
From MaRDI portal
!
This is the item page for this Wikibase entity, intended for internal use and editing purposes. Please use the normal view instead:
scientific article; zbMATH DE number 4034905
| Language | Label | Description | Also known as |
|---|---|---|---|
| default for all languages | No label defined |
||
| English | An asymptotic theory for weighted least-squares with weights estimated by replication |
scientific article; zbMATH DE number 4034905 |
Statements
An asymptotic theory for weighted least-squares with weights estimated by replication (English)
0 references
1988
0 references
large-sample properties
0 references
generalized least-squares
0 references
multiple linear regression model
0 references
heteroscedastic errors
0 references
replications
0 references
weighted least- squares estimators
0 references
inverses of the sample variances
0 references
normally distributed errors
0 references
asymptotic distribution
0 references
sample average squared errors
0 references
efficiency
0 references
0.9152491
0 references
0.8974907
0 references
0.88068235
0 references
0.8731133
0 references
0.87106645
0 references
0 references
0 references
A multiple linear regression model with heteroscedastic errors and m replications is considered. The model equation is \[ y_{ij}=x^ T_ i\beta +\sigma_ ie_{ij}\quad (i=1,...,N;\quad j=1,...,m). \] If, as assumed, the \(\sigma^ 2_ i\) are unknown, weighted least-squares estimators \(\beta\) can be calculated by estimating the weights by the inverses of the sample variances. Usually m is small and generally the estimates are inconsistent for asymmetrical error distributions. For normally distributed errors the asymptotic distribution of \(\beta\) is given and it is shown that only for \(m\geq 6\) its mean is \(\beta\) but that the estimated sample variances in those cases cause an inflation of variance by the factor (m-5)/(m-3) over weighted least squares with known weights. A warning is given when standard statistical packages are used with estimated weights; m should be at least equal to ten. If a preliminary fit to the regression function is made and the sample average squared errors from this fit are used as a basis for estimating weights, the efficiency is higher compared with that of \(\beta\). However, even here a fair number of replicates is helpful.
0 references