# Corrected AIC in Proc Reg

#### noetsi

##### Fortran must die
Some call this AICc. Supposedly, some say, Proc Reg will generate this statistic, but I could not find it in the documentation. It is produced in proc logistics by specifying FITSTAT

Does anyone know if this exists in Proc Reg?

#### hlsmith

##### Not a robit
does it exist in proc glm? reg is pretty useless.

its aic a MLE thing?

#### noetsi

##### Fortran must die
Its a modification of AIC for small samples. Because AIC assumes a large sample size. I am uncertain how it is generated.

PROC REG is not useless. Is not is not..

Given that PROC GLM does not generate many of the key diagnostics that PROC REG does to test the regression assumptions [it generates almost none of them] you have to use proc reg to test these and then just assume the results would be the same in proc glm. Which I always worry about.

Its why after spending a lot of time learning PROC GLM I never use it. I think things like White's SE are also not built for it.

#### Con-Tester

##### Member
Not directly, as far as I know. It seems to be limited to fitting penalised b-splines (proc transreg). But you can easily calculate or estimate AICc from AIC if you know how many data points were used and how many parameters are involved, including those of variable transformations. You can specify AIC output as part of the “model” statement in proc reg.

One must be careful when counting parameters because some parameters can be rolled into a single one. For example, a linear model includes the term m·x’. If the transformation x’ = ln((x+a)/(x+b)) was applied, the term m·x’ involves three parameters, namely m, a and b. On the other hand, if the transformation x’ = ln(x^a/(x+b)) was applied, the term m·x’ involves only two parameters, namely m’ = m·a and b because the exponent a can be extracted from the logarithm and combined with m into a single parameter m’.

#### noetsi

##### Fortran must die
Strangely it is calculated for PROC Logistics and some of the time series modules. It apparently is calculated even for some methods where its use has not been validated.

What do you specify in the model statement in PROC REG to use it? I have not found this. Obviously its an option, which one?

#### Con-Tester

##### Member
I don’t have access to SAS right now, so I will only be able to confirm tomorrow.

I think you just need to specify “AIC” in the <Options> list after the “/”, like so:
Code:
PROC REG...;
:
:
MODEL <Dependent> = <Independent> /<Options>;
:
:
RUN;