[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: Model Building
Nick Cox <email@example.com>
Re: st: Model Building
Tue, 26 Apr 2011 16:43:48 +0100
I would put it differently from Justina. Statalist certainly allows
statistical questions as well as Stata questions. But in practice,
many members are reluctant on various grounds to answer questions that
look like "How should I analyse my data?" or "What should I next do in
Otherwise put, Statalist is a discussion list, not a help line
guaranteeing support for the lost or clueless! There has to be a tacit
contract, i.e. someone else wants to discuss your question, for an
exchange to occur.
That said, there is a Stata question at the heart of this. Fractional
polynomial techniques do not require that you generate the fractional
polynomial beforehand. On the contrary, the whole point is to generate
that for you.
On Tue, Apr 26, 2011 at 1:14 PM, Justina Fischer <JAVFischer@gmx.de> wrote:
> Dear Stefan,
> your question is a bit out of the scope of this list - we discuss specific methodological issues only.
> You could select important variables through a general-to-specific approach.
> However, in general it is better to make a theory-based selection.
> Good luck
> -------- Original-Nachricht --------
>> Datum: Tue, 26 Apr 2011 14:07:38 +0200
>> Von: Stefan Nijssen <firstname.lastname@example.org>
>> An: email@example.com
>> Betreff: st: Model Building
>> Dear Statalist members,
>> I am trying to find the best indicators (independent variables,
>> corporate fundamental data) to predict the dependent variable (the
>> risk factor; 'oas'). I want to create a model with as few variables as
>> possible to best predict 'oas'. So I need to make a selection of all
>> variables available, see below. However, first thing to do seems to be
>> finding the correct form with which the independents interact with the
>> dependent. Plotting the variables separately with 'oas' I find some of
>> them to be linear, some clearly quadratic and others fractional
>> polynomial. To me it seems logical to converge the variables into
>> their correct form as to best work with them, puzzling which to
>> include and which not. I thought of using the .reg function to test
>> the function. I do not see how to create a quadratic variable (other
>> than generating var*var), and fractional polynomial will be even more
>> difficult, before using them in the regression. Can anyone provide
>> suggestions on what to do?
>> After doing so, I was thinking of using Multiple Discriminant Analysis
>> to reduce the number of variables.
>> Stefan Nijssen
* For searches and help try: