Ordinal Logistic models and the importance of reading manuals

[Paul Johnson <>]

Toward the end of the last semester, the answer I kept giving everybody 
was "you have to read the manual for the software you use, and if 
there's no manual, get different software."  I was reminded of the 
problem when I ran 2 ordinal logit models this morning, and the signs 
came out exactly backward.  One is predicting Y<1 and the other predicts 
Y>1.  If you aren't looking closely, you can really get in a tail spin.

Here's some R output from 2 procedures, one vglm (from VGAM package) and 
the other lrm (from the Design package). Note the sign reversal and also 
the fact that vglm reports the t stat, while lrm gives Wald Chi Square

> summary(vglm(punish~Vict,cumulative(parallel=T),data=dat))

Call:
vglm(formula = punish ~ Vict, family = cumulative(parallel = T),
    data = dat)

Pearson Residuals:
                   Min       1Q   Median      3Q     Max
logit(P[Y<=1]) -1.2173 -0.43163 -0.39514 1.33643 1.51802
logit(P[Y<=2]) -1.3670 -1.23322  0.43162 0.47613 1.41605
logit(P[Y<=3]) -2.1809  0.20569  0.31432 0.81117 0.85266

Coefficients:
                     Value Std. Error   t value
(Intercept):1    -0.947516    0.40285 -2.352023
(Intercept):2     0.016581    0.38347  0.043239
(Intercept):3     1.373316    0.43149  3.182743
VictMale          0.247066    0.47125  0.524283

Number of linear predictors:  3

Names of linear predictors: logit(P[Y<=1]), logit(P[Y<=2]), logit(P[Y<=3])

Dispersion Parameter for cumulative family:   1

Residual Deviance: 166.386 on 179 degrees of freedom

Log-likelihood: -83.193 on 179 degrees of freedom

Number of Iterations: 4
> lrm(punish~Vict,data=dat)

Logistic Regression Model

lrm(formula = punish ~ Vict, data = dat)


Frequencies of Responses
 0  1  2  3
19 14 17 11

Frequencies of Missing Values Due to Each Variable
  punish             Vict
           0           24

       Obs  Max Deriv Model L.R.       d.f.          P          C 
  Dxy
        61      2e-11       0.28          1     0.5997      0.525 
0.049
     Gamma      Tau-a         R2      Brier
     0.102      0.037      0.005      0.214

                  Coef     S.E.   Wald Z P
y>=1               0.94752 0.4049  2.34  0.0193
y>=2              -0.01658 0.3810 -0.04  0.9653
y>=3              -1.37331 0.4257 -3.23  0.0013
Vict=Male         -0.24707 0.4710 -0.52  0.5999


--
Paul E. Johnson                       email: pauljohn_AT_ku.edu
Dept. of Political Science            http://lark.cc.ku.edu/~pauljohn
1541 Lilac Lane, Rm 504
University of Kansas                  Office: (785) 864-9086
Lawrence, Kansas 66044-3177           FAX: (785) 864-5700