Hi Jan,
Am 19.08.2009 18:28, student09(a)web.de schrieb:
However, both variables actually refer to count data. So I thought
that doing a VAR time-series model for poisson-distributed variables
might be appropriate - does anybody know any literature/examples in
this regard? I am quite unexperienced in time series models but have
a background in SEM/Multilevel modelling, so I am just guessing which
kind of models might be appropriate here...
without having any practical experience with count data myself,
intuitively this doesn't sound like a good idea. I mean, would it make
any sense to get non-integer predicted values? Because that's what you
would get if the coefficients are freely estimated, right? And imposing
coefficient restrictions so that it doesn't happen doesn't look trivial
to me.
Also - and also not directly related to Gretl - I realized that after
differencing, the distribution of the residuals of the variables is
actually gaussian. Wouldn't it then make sense to refrain from any
specific distributions for count data and 'simply' go ahead with the
usual estimation procedures for normal-distributed variables???
again, without being an expert on this issue and without having thought
about it much, I'm sceptical. You mean that normality tests cannot
reject the null of normality, correct? I can easily imagine that it can
happen, depending on the parameters of the underlying (Poisson)
distribution and on the sample size, but in general this doesn't sound
right.
well, actually I could have saved ourselves my vague comment, googling
is probably more informative, for example a search for "count data time
series" turns up this seemingly very relevant top result:
http://opus.zbw-kiel.de/volltexte/2005/3194/pdf/EWP-2005-08.pdf
good luck,
sven