Hello,
For this project I have been tasked with emulating an old Dynamic Linear
Modelling paper's results in the R programming language with the same data.
The majority of the work (creating the model, filtering, smoothing, forecasting,
etc.) has been done via the dlm package, and I have been successful in at least
mimicking the old project's plot and coming within reasonable range on the
Mean Squared Errors and Median Absolute Deviations (for goodness of fit testing,
etc...).
The strange thing has been in calculating the Log Likelihood. Running it through
the dlm package's dlmLL method spits out an unusually large number compared
with the old projects results (which was run on this Bayesian Analysis of Time
Series (BATS) program). This is specifically for a second order polynomial DLM
(linear trend model), whose filtered values were produced from a method
dlmFilterDF (DF = discount factor) written by Giovanni Petris.
A constant trend model will return values much closer to what I have expected
(100s vs 4000-5000s). Same with the linear trend model if you plug in just the
filtered plot values (i.e. if you ran the program, I ran filteredmodel$m[,1] for
the Log likelihood).
What I am hoping to do is get an accurate read on the LL, which will then be
wrapped in a logLik object and used to calculate the Akaike and Bayesian
Information Criterion for further analysis under the stats4 package.
Those who I am working with and a stats professor have told me I might be
getting raw LLs, where as the old paper might have neglected to mention a
constant they may have applied. Personally I would rather get the measurement
correct given the environment.
So in general, if you have played around with the dlm R package, I would like to
know if you have ran into this situation and/or what could help in making the
LLs a little more reasonable (parameters within the model that need
manipulation, perhaps I am using the wrong thing to calculate the LL...).
FWIW: my math background is heavy in calc, decent with linear algebra,
introductory stats yet I have learned much in the ways of Bayesian Statistics
and parameters that drive the models I am running here for the project, so if
you explain something in that latter realm, a small summary of details will
suffice.
Many thanks!!!
-Joeseph P. Smith
