topic Re: Conditional logit using proc logistic vs proc mdc in Statistical Procedures

Conditional logit using proc logistic vs proc mdc

Demographer — Sat, 15 Sep 2012 20:46:19 GMT

Hi,

I try to perform a conditional logit with SAS. Suppose that z is an alternative's characteristic coded 0-1 and x an individual's characteristic coded 0-1 too. The dependant variable CHOICE is of course binary too. I want to modelise the interaction between z and x, so I create 3 dummy variables as follow:

xz1=0; if x=0 and z=1 then xz1=1;

xz2=0; if x=1 and z=0 then xz2=1;

xz3=0;if x=1 and z=1 then xz3=1;

The database is then like this:

SUBJECT	MODE	CHOICE	X	Z	XZ1	XZ2	XZ3
1	1	0	0	1	1	0	0
1	2	0	0	1	1	0	0
1	3	0	0	1	1	0	0
1	4	0	0	1	1	0	0
1	5	0	0	0	0	0	0
1	6	0	0	0	0	0	0
1	7	0	0	1	1	0	0
1	8	1	0	1	1	0	0
1	9	0	0	1	1	0	0
1	10	0	0	0	0	0	0
2	1	0	1	1	0	0	1
2	2	1	1	1	0	0	1
2	3	0	1	1	0	0	1
…	…	…	…	…	…	…	…
1000	9	0	0	1	1	0	0
1000	10	0	0	0	0	0	0

When I use the procedure PROC MDC, all seems correct.

proc mdc ;

id SUBJECT;

model CHOICE = XZ1 XZ2 XZ3/ type=clogit choice=(MODE) ;

run;

The results are as follows:

Erreur Prob.

Paramètre DDL Estimation type Valeur du test t Approx. > |t|

XZ1 1 -0.8324 0.0996 -8.36 <.0001

XZ2 1 -0.2845 0.0546 -5.21 <.0001

XZ3 1 0.4596 0.0546 8.42 <.0001

However, when I use the procedure PROC LOGISTIC, results are weird:

proc logistic;

model CHOICE(event='1') = XZ1 XZ2 XZ3 ;

strata SUBJECT;

run;

Erreur Khi 2

Paramètre DDL Estimation type de Wald Pr > Khi 2

XZ1 1 -0.8324 0.0996 69.8063 <.0001

XZ2 1 -0.7442 0.1092 46.4685 <.0001

XZ3 0 0 . . .

As you notice, the parameter of the last interaction variable is set to 0. This message appears in logs:

Note: The following parameters have been set to 0, since the variables are a linear combination of other variables as shown. XZ3=-XZ2

I first think it was a problem with the database, but since the results with the same variables using the PROC MDC are corrects, I think my database is ok. In addition, I notice that the parameters for XZ2 and XZ3 from the PROC MDC give the parameter for XZ2 from the PROC LOGISTIC by doing XZ2-XZ3 (-0.2845-0.4596 = -0.7442).

The same problem as the one in the PROC LOGISTIC occurs to when I try PROC PHREG.

Some has an idea on the reason of this problem?

Re: Conditional logit using proc logistic vs proc mdc

Reeza — Mon, 17 Sep 2012 20:14:24 GMT

Its saying that your model has two variables that are perfectly collinear, ie Xz3=-XZ2 . You only need of the two variables in the model, because the second provides no extra information and can be calculated separately.

Re: Conditional logit using proc logistic vs proc mdc

Demographer — Mon, 17 Sep 2012 20:56:33 GMT

Thank for your answer, but I'm pretty sure this message from SAS is wrong. I need 3 variables since I want the interaction of 2 binary variables (so 4 possible cases, one of them is the reference). Moreover, it works with 3 variables by using PROC MDC or Stata.

Re: Conditional logit using proc logistic vs proc mdc

Reeza — Mon, 17 Sep 2012 22:07:40 GMT

xz2=0; if x=1 and z=0 then xz2=1;

xz3=0;if x=1 and z=1 then xz3=1;

Those variables are the inverse of each other, when xz2=1 then xz3=0 and when xz2=0 then xz3=1.

Run a proc freq xz2*xz3 to see....

If you want the interaction in SAS you can use the class statement instead:

proc logistic data=have;

class x z/param=ref;

model choice(event='1')= x z x*z;

run;

Re: Conditional logit using proc logistic vs proc mdc

Demographer — Tue, 18 Sep 2012 00:46:01 GMT

You're right, thank a lot...

But there is something I don't get then. How can I estimate the parameter for xz3? There is 4 possibilities and only 2 parameters.

Re: Conditional logit using proc logistic vs proc mdc

Demographer — Tue, 18 Sep 2012 00:51:15 GMT

Reeza a écrit:

proc logistic data=have;
class x z/param=ref;
model choice(event='1')= x z x*z;
run;

I'm not totaly sure, but I can't do that since it's a conditional logit (with strata statement). The individual's variables (x) can't be there alone. They have to be in interaction with an alternative's variables.