If you fit an ordinal logistic model using the DIST=MULT and LINK=CLOGIT options, and if you give only a partial specification of an estimable ESTIMATE or CONTRAST statement, then the complete set of coefficients that PROC GENMOD constructs can be incorrect producing a nonestimable result. Construction of a nonestimable specification can occur when it involves the intercept for any logit other than the first. For example, the following partially specified ESTIMATE statements attempt to estimate the LS-mean for the first level of A in each of the three cumulative logits when Y has four levels. PROC GENMOD completes the specification by adding coefficients for the B and A*B effects. However, the coefficients that it constructs are correct only in the first ESTIMATE statement for the first logit. The coefficients that it constructs for the B effect are incorrect in the last two ESTIMATE statements and are nonestimable. The coefficients of the completed specifications are displayed by the E options.

proc genmod;
  class A B;
  model Y = A B A*B  / dist=mult link=clogit;
  estimate "A1 @ Y1" Intercept 1 0 0   A 1 0  /  e;
  estimate "A1 @ Y2" Intercept 0 1 0   A 1 0  /  e;
  estimate "A1 @ Y3" Intercept 0 0 1   A 1 0  /  e;
  run;

One way to avoid the problem is to provide a complete specification:

  estimate "A1 @ Y2" Intercept 0 1 0  A 1 0  B .3333333 .3333333 .3333333 A*B .3333333 .3333333 .3333333 0 0 0 / e; 

Alternatively, you can use the LSMEANS statement:

  lsmeans A / e;

Another way around the problem is to fit the model in PROC LOGISTIC. The following statements fit the same model and the single ESTIMATE statement with the CATEGORY=SEPARATE option provides the estimates for all three logits:

proc logistic;
  class A B / param=glm;
  model Y = A B A*B;
  estimate "A1" Intercept 1  A 1 0 / e category=separate;
  run;

Operating System and Release Information