Why do logistic regression coefficients in SAS produce signs opposite to those of other statistical software packages?
The SAS PROC LOGISTIC procedure models the probability of
the smaller value of the dichotomous response. Assuming a dependent
variable Y has been coded to have the values 0 or 1, SAS by default
estimates a model of probability of 0.
More specifically, SAS estimates log{(1-p)/p} =
-X'b, where 1-p is the probability of Y=0,
b is the parameter vector, and X is the
vector of independent variables.
In contrast, the .logit command in Stata and
the LOGISTIC REGRESSION command in SPSS model the
probability of the larger value of the dependent variable. In Stata
and SPSS, the estimated logit is log{p/(1-p)} =
X'b. This difference results in signs opposite to those from SAS.
The DESCENDING option in the PROC LOGISTIC
statement enables users to model the probability of Y=1 in
SAS. The statement is as follows:
Last modified on March 24, 2011.
