GLM, cov for different dispersion

[maxchen]

How the covariance are estimated in GLM with different settings of dispersion?

Code: Select all

%ex7 = "\Example Files\EV7 Manual Data\" %ch27 = "Chapter 27 - Generalized Linear Models" %wf = @evpath + %ex7 +%ch27 +"\strike.wf1" wfopen %wf 'how the cov are estimated equation eq060.glm(family=poisson,disp=pearson,covinfo=bhhh,c=1e-8) numb c ip feb equation eq061.glm(family=poisson,disp=unit,covinfo=bhhh,c=1e-8) numb c ip feb equation eq062.glm(family=poisson,disp=deviance,covinfo=bhhh,c=1e-8) numb c ip feb equation eq063.glm(family=poisson,dispval=3,covinfo=bhhh,c=1e-8) numb c ip feb

my questions are
1. compare eq061 with eq063, with unit and 3 as dispersion value respectively, however, give the same cov estimation
2. compare eq060 with eq061, what are the difference when estimating cov
3. compare eq060 with eq062, the std errs are scaled by eq060.@deviancestat/eq060.@dispersion

[maxchen]

It seems that the rules are
(a) if given dispval, the cov are estimated by disp=unit
(b) if by disp=pearson or disp=deviance, the cov are estimated by disp=unit scaled by @dispersion/!N*(!N-!K)