I get the same BIC and -2 Res Log!?

[krokodil]

Hi guys!

I need to know why I get the same BIC and -2 Res log likelihood values. Shouldnt that be impossible when the equation for BIC is -2ReslogL + 2p, p = number of parameters.

Im using this code for a dataset of macro economic data

PROC IMPORT OUT=oil DATAFILE= "W:\oil.xlsx"

DBMS=xlsx REPLACE;

GETNAMES=YES;

RUN;

proc mixed data=oil covtest noitprint;

class country year quarter;

model hcip=oilprice interest exchange unemploy output_g/solution residual;

random intercept/subject=country;

random intercept/subject=year(country);

repeated /type=ar(1) subject=year;

run;

[ChrisHemedinger]

I moved this to for better visibility to the stats experts.

Chris

[SteveDenham]

What about the other information criteria?  Are they messed up as well?

It could be that the -2 log likelihood value is so large that adding in 2p gets swamped in the printed output. Is that a possibility?  To check, use ODS output to get the IC into a dataset, and look at the values in a long format.

Steve Denham

Message was edited by: Steve Denham Removed calculated value that depended on number of fixed effects

[lvm]

By the way, the formula for BIC is -2LL + p.ln(n). You gave the formula for AIC. Also note that p is the number of variance-covariance parameters with REML estimation (what you are doing). The fixed effect parameters have nothing to do with it (you may already know this, but other readers may not). Are all your variance-covariance parameters 0? Steve also has a good point. Would be good to see all your output.

[Rick_SAS]

I'll add that procedures that minimize some sort of information criterion (LOESS, GLMSELECT, SEVERITY, any model selection procedure) sometimes add or substract constants.  Not relevant for this problem, I don't think, but some sources might give F(n, p; x) + C, whereas SAS reports F(n,p; x).

[lvm]

Good point about statistical software. This can lead to a lot of confusion in comparing output from different programs with the same data and same model. In MIXED and GLIMMIX, however, one does get the complete likelihood (or restricted likelihood) in the information-criterion statistics, including the term(s) that does not depend on the the parameters. For REML estimation, only variance-covariance parameters are used in the calculation because the fixed effect are eliminated at each ML step.  I hope to see a compete output from the OP to help figure out this issue.