Generalised version of LINEST using QR Decomposition

Will wrote...
>Does anyone know if a generalised version of the Excel 2003 LINEST function
>using QR decomposition is available, beyond that posted in the knowledge base
>article(http://support.microsoft.com/kb/828533)? That shows the way in which
>this can be done with 2 regressors but not for n regressors and I need to
>know how this can be extended. Ideally, I think this should be available as a
>free bug fix in Excel, as Microsoft has acknowledged that this is a bug.
>Also, the Knowledge Base article doesn't show how the standard error of each
>coefficient can be determined, hence it's impossible to derive the t-stats
>for each coefficient.

It's fixed in the sense that if it's really important to you, you can
upgrade to Excel 2003 (or 2004 for Macs). There's not going to be a
free fix for earlier versions of Excel.

As for the QR decomposition, it's an explanation of how it works. From
skimming it, it could be extended to more than two independent (X)
variables *if* you know the underlying linear algebraic equations. If
you don't, you shouldn't be messing with this.

Thanks Harlan, but the biggest problem is that I work at a firm where any
rollout of new software is done on a corporate-wide basis and takes several
years to occur. If I could take the upgrade path then I would, but this isn't
available to me for several years.

With regards to the given solution, some steps are unclear, for example, why
a column of 1s is added and why the columns are switched around, hence the
explanation is not sufficiently clear in order to generalise. I am trying to
generate an Excel-based solution for some econometrics tests (specifically,
the Augmented Dickey-Fuller test which is specifically designed to account
for correlations in the lagged variables, and hence is guaranteed to generate
problems with multi-collinearity) and I would rather not be forced down the
route of switching to using Matlab instead of Excel.

Does anyone know if a generalised version of the Excel 2003 LINEST function
using QR decomposition is available, beyond that posted in the knowledge base
article(http://support.microsoft.com/kb/828533)? That shows the way in which
this can be done with 2 regressors but not for n regressors and I need to
know how this can be extended. Ideally, I think this should be available as a
free bug fix in Excel, as Microsoft has acknowledged that this is a bug.
Also, the Knowledge Base article doesn't show how the standard error of each
coefficient can be determined, hence it's impossible to derive the t-stats
for each coefficient.

Will wrote...
>Does anyone know if a generalised version of the Excel 2003 LINEST function
>using QR decomposition is available, beyond that posted in the knowledge base
>article(http://support.microsoft.com/kb/828533)? That shows the way in which
>this can be done with 2 regressors but not for n regressors and I need to
>know how this can be extended. Ideally, I think this should be available as a
>free bug fix in Excel, as Microsoft has acknowledged that this is a bug.
>Also, the Knowledge Base article doesn't show how the standard error of each
>coefficient can be determined, hence it's impossible to derive the t-stats
>for each coefficient.

It's fixed in the sense that if it's really important to you, you can
upgrade to Excel 2003 (or 2004 for Macs). There's not going to be a
free fix for earlier versions of Excel.

As for the QR decomposition, it's an explanation of how it works. From
skimming it, it could be extended to more than two independent (X)
variables *if* you know the underlying linear algebraic equations. If
you don't, you shouldn't be messing with this.

Thanks Harlan, but the biggest problem is that I work at a firm where any
rollout of new software is done on a corporate-wide basis and takes several
years to occur. If I could take the upgrade path then I would, but this isn't
available to me for several years.

With regards to the given solution, some steps are unclear, for example, why
a column of 1s is added and why the columns are switched around, hence the
explanation is not sufficiently clear in order to generalise. I am trying to
generate an Excel-based solution for some econometrics tests (specifically,
the Augmented Dickey-Fuller test which is specifically designed to account
for correlations in the lagged variables, and hence is guaranteed to generate
problems with multi-collinearity) and I would rather not be forced down the
route of switching to using Matlab instead of Excel.