BIC after GLM

[Dan Palmer]

I've read the documentation from Stata about the computational differences between BIC values as reported by glm versus those reported by estat ic. With those differences in mind, I'm curious to know if anyone here would be willing to share which report (i.e., via glm or via estat ic) she or he typically uses to assess BIC and why. It seems to me that the BIC estimate given from the glm command is problematic given that it directly reflects deviance, which is scale dependent in some respects.

[Dan Palmer]

I ask this because after running two GLM models for employee salary (the first being gaussian-identity, the second being gamma-power-1) I'm seeing conflicting information from the two different BIC measures. The BIC indicator reported from the glm command shows a drastic reduction from 1.70e+11 to -8884, while the BIC indicator reported from the estat ic command shows an increase from 27441 to 30,886. So, the two indicators of BIC seem to be pointing me in two different directions.