postestimation in glm

[misidro]

Hi Statalisters,

I am fitting a glm model with this form:

glm PS101 Yb wall urb, family(poisson) link(log) lnoffset(Meancount)

I am trying to generate the predicted value "manually" (pred2 below) so I can use it in a simulation where I am going to vary my betas and the value of my offset. However I am getting a predicted value that is not equal to mu (i.e., predict pred1, mu).

Here is an example of my codes: global xvars "Yb wall urb"
glm PS101 Yb wall urb, family(poisson) link(log) lnoffset(Meancount)
global px=e(df_m)
matrix beta=e(b)'
predict pred1, mu

global px1=$px+1
gen pred2=0
forvalues i=1/$px {
local v : word `i' of $xvars
replace pred2=pred2 + Meancount* exp(beta[$px1,1]+ beta[`i',1]*`v')
}
If anyone could point out what is wrong with my pred2 formula or could provide a resource on the formula behind the GLM post estimation calculation for mu when you have Poisson distribution and there is an offset that would be really great!

Thanks in advance!
-Maris

[Scott Merryman]

The offset should be within the exponential function.

For example:

webuse rod93,clear
generate logexp=ln(exposure)
qui glm  deaths cohort, offset(logexp) family(poisson) link(log)
predict mu,mu
gen my_mu= exp( [deaths]_b[cohort]*cohort + [deaths]_b[_cons] + logexp)
l in 1/2

[Richard Williams]

As a sidelight, while the lnoffset option still works, it has been replaced by exposure. So Scott's code could be rewritten as

webuse rod93,clear
qui glm  deaths cohort, exposure(exposure) family(poisson) link(log)
predict mu,mu
gen my_mu= exp( [deaths]_b[cohort]*cohort + [deaths]_b[_cons] + ln(exposure))
l in 1/2

[misidro]

Thanks for your help! Really appreciate it.