Mauro Talevi wrote:
> Hi Phil,
>
> thanks for reviewing the multiple linear regression implementations
> and setting up the R/NIST data tests. I finally got around to
> installing R and can now run them too.
>
> Phil Steitz wrote:
>> While clear and elegant from a matrix algebra standpoint, the
>> "nailve" implementation in OLSMultipleLinearRegression has bad
>> numerical qualities. It is well known that solving the normal
>> equations directly does not give good numerics. I just added some
>> tests to actually verify parameter values, using the classic "Longly"
>> dataset, for which NIST provides certified statistics. This is a
>> "hard" design matrix. R was able to get to within 1E8 of the
>> certified parameter values. OLSMultipleLinearRegression can only get
>> 1E1.
>
> The OLS implementation has been added as a simple byproduct of the
> GLS case  which is the main one I have needed for hypothesis testing
>  as it came "for free" with unitary covariance.
> True  the emphasis was on clarity and formulaic simplicity. And also
> following the old Donald Knuth maxim "optimization is the root of all
> evil". But it seems like there is a need for refinement of the
> implementation  the devil raised his head :)
>
Yes, and I would distinguish performance optimization from numerical
accuracy. From my perspective, we can release a ".0" with room for
performance improvement, but at least decent numerics are required.
>> We have talked in the past about providing an implementation based on
>> QR decomposition. Anyone up for using the QR decomposition that we
>> now have to do this? I really think we need to do it (or something
>> else to improve numerics) before releasing this class. I will get to
>> it eventually, but am a little pegged at the moment. I will review
>> and apply patches if someone is willing to do the implementation. I
>> can also explain here or offline how the R tests and NIST datasets
>> work, as these are useful in validating code.
>
> I'd be happy to improve the impl. I'm getting my head around R and
> NIST, but perhaps a chat offline would not hurt!
I may be hard to catch synchronously, as my dayandnight job is a
little demanding, but I would be happy to answer questions (with maybe a
little latency ;)
>
>> Another thing that we should think about before releasing any of this
>> stuff is the completeness of the API. Many standard regression
>> statistics are missing. If we are going to stick with the Interface
>> / Implementation setup, we need to get the right stuff into the
>> interface. It is also awkward to have to insert "1"'s in the design
>> matrix to get an intercept term computed. This is convenient for
>> implementation, but awkward for users. A more natural setup (IMHO)
>> would be to expose a "noIntercept" or "hasIntercept" property for the
>> model.
>
> No problem with adding other statistics  let's just decide on what is
> the stardard regression API.
Here are some initial ideas on what should be included in the multiple
regression API. Other suggestions welcome!
1. Coefficients should be accompanied by standard errors, tstatistics,
twosided t probablilities (can get these using t distribution from
distributions package) and ideally confidence intervals.
2. F, Rsquare, adjusted Rsquare, F prob (again can use distributions
package to estimate)
3. ANOVA table (Regression sum of squares, residual sum of squares)
4. Residuals
R, SAS, SPSS and Excel all represent (or in the case of R, can
construct) these basic statistics in some way in their output. We
should model them in classes representing properties of the computed
model.
>
> And finally, how do you see the no/hasIntercept model working?
As a configurable property  noIntercept means the model is estimated
without an intercept. The point I was making was more how the data is
supplied via the API. It is awkward to have to fill in a column of 1's
to get the linear algebra to work to estimate a model with intercept
(which should be the default).
I would recommend that we have setData or "newData" provide a n x m
matrix, where n is the number of observations and m1 is the number of
independent variables. Then either a) have the constructor take another
argument specifying which column holds the dependent variable b) assume
it is the first column c) support column labels and some form of model
specification such as what R provides (a lot of work) d) split off the y
vector, so setting data requires separate x and y vectors. Probably a)
is easiest for users, who will most often be starting with a rectangular
array of data with the dependent variable in one of the columns.
Phil

To unsubscribe, email: devunsubscribe@commons.apache.org
For additional commands, email: devhelp@commons.apache.org
