RE: st: statistical significance of cut points in ordered logit

[Richard Williams <richardwilliams.ndu@gmail.com>]

At 08:43 AM 5/17/2010, Grace Jessie wrote:
> Additionally, does the statistical significance of cut points in 
> ordered logit matter?
> I found there are no z or P>|z| for cut points, though I could get it.

If you want p values for the cutpoints, you can get them with -oglm-, 
available from SSC.  If no options are specified, oglm produces the 
same estimates as ologit (albeit more slowly).

I'm not sure what you do with the p values once you have 
them.  Indeed, there is no reason that one of the cutpoints can't be 
zero.  I suppose if the confidence intervals for two cut points 
overlapped, you might want to examine whether the N is too small in 
some categories (of course you may want to examine that 
regardless).  If Ns in some categories are very small, you might want 
to combine some adjacent categories.  For example, I don't like using 
the auto data for ordinal regression problems because the Ns are way 
too small both overall and for individual categories of rep77, e.g.

webuse fullauto
ologit rep77 foreign length mpg
tab1 rep77

The latter command yields

-> tabulation of rep77

     Repair |
Record 1977 |      Freq.     Percent        Cum.
------------+-----------------------------------
       Poor |          3        4.55        4.55
       Fair |         11       16.67       21.21
    Average |         27       40.91       62.12
       Good |         20       30.30       92.42
  Excellent |          5        7.58      100.00
------------+-----------------------------------
      Total |         66      100.00

You are supposed to have at least 100 cases for maximum likelihood 
methods, and I would think even more for something like ologit.  Toss 
in the extremely small Ns at the 2 extremes and I don't know how much 
you can trust these estimates.

Incidentally, if you now run the -brant- command (part of Long & 
Freese's spost package) you get

. brant
not all independent variables can be retained in all binary logits
brant test cannot be computed
r(999);

So again, my big concern is a small overall N and/or small 
subcategory Ns.  Cutpoints that don't significantly differ from each 
other might be a result of such problems, but there are other ways to 
check too, e.g. just look at your frequencies and/or run a brant test.


-------------------------------------------
Richard Williams, Notre Dame Dept of Sociology
OFFICE: (574)631-6668, (574)631-6463
HOME:   (574)289-5227
EMAIL:  Richard.A.Williams.5@ND.Edu
WWW:    http://www.nd.edu/~rwilliam

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/