Are there Issues with Separation of Values in Ordinal Logistic Regression

Click For Summary

Discussion Overview

The discussion revolves around potential issues related to the separation of values in ordinal logistic regression, particularly in the context of 3-valued or higher-valued logistic regression. Participants explore the implications of separation, its causes, and possible solutions, focusing on both theoretical and practical aspects of the modeling process.

Discussion Character

  • Debate/contested
  • Technical explanation
  • Exploratory

Main Points Raised

  • One participant questions whether issues of separation in binary logistic regression also apply to 3-valued logistic regression, suggesting a connection to the behavior of the S-curve.
  • Another participant clarifies that "Separation of Points" typically refers to complete or quasi separation, which can occur in small or miscoded datasets, and notes that this issue does not disappear in more general cases.
  • A participant explains that separation occurs when there is a threshold value of the independent variable beyond which all outcomes are either successes or failures, providing an example related to smoking and cancer.
  • One participant agrees that separation is indeed a problem for higher orders and mentions that statistical packages often alert users to this issue, suggesting penalizing maximum estimators or hidden logistic models as potential solutions.
  • A follow-up question is raised about the appropriateness of slightly altering data values to avoid separation, with the participant expressing concern about preserving the intrinsic properties of the dataset while modeling probabilities.

Areas of Agreement / Disagreement

Participants generally agree that separation is a relevant issue in ordinal logistic regression, including higher orders. However, there is no consensus on the best approaches to address this problem, and differing opinions on the implications of altering data values are present.

Contextual Notes

Participants express uncertainty regarding the definitions and implications of separation, and there are unresolved questions about the effects of modifying data values to mitigate separation issues.

WWGD
Science Advisor
Homework Helper
Messages
7,806
Reaction score
13,120
Hi all , just curious if someone knows of any issues of Separation of Points in Ordinal 3-valued
Logistic Regression. I think I have an idea of why there are issues with separation in binary
Logistic -- the need for the S-curve to go to 0 quickly makes the Bo term go to infinity. Are there
similar issues with 3-valued (or higher-valued) Logistic Regression?
 
I'm not entirely clear what you mean by "Separation of Points". Whenever I hear "Separation" with regards to logistic regression, it deals with complete separation or quasi separation, which tends to occur with small dataset/miscoded datasets. The problem that causes this (MLE not existing) doesn't disappear in more general cases.

There's ways around that (sometimes), but I feel that we may be talking about two different things.
 
MarneMath said:
I'm not entirely clear what you mean by "Separation of Points". Whenever I hear "Separation" with regards to logistic regression, it deals with complete separation or quasi separation, which tends to occur with small dataset/miscoded datasets. The problem that causes this (MLE not existing) doesn't disappear in more general cases.

There's ways around that (sometimes), but I feel that we may be talking about two different things.
Hi thanks for replying. Separation happens when there is a value Xo of the independent variable (obviously this applies to cases with numerica; variables) such that for all X>Xo all trials (Bernoulli or multinomial) are fails or all trials are successes. e.g., if Y dependent was "has Cancer" and X is number of cigarettes smoked per week, then X is separated if for, e.g., X>10 all are fails, i.e., everyone who smoked more than 10 cigarettes got cancer.
 
Ok, then I think we are talking about he same thing. Then yes, separation is a problem even for higher orders. Most statistical packages are good at notifying you when this happens. One way around this is by using a penalizing the maximum estimator. I'm personally a fan of using a hidden logistic to overcome this when necessary.
 
Just a followup on this: would it be reasonable, in the sense of not affecting "intrinsic" properties of a data set with separation of values with smallish size each, say in the range [0,5] , to slightly alter ; increase/decrease some of the data values , so as to overcome this issue, i.e., so that the values beyond a certain number are not monotone? Say my cutoff point for this data set within the [0,5] range is 3 and I have several points with value 3. Then I could change the data set to replace , in some cases, 3 by 3.02, in other cases 3 would be replaced by, say 2.98 , in order to avoid this problem? I just want to be able to model the probability of success by doing this; obviously, I would think, most of the properties of the data would be preserved by doing this?
 
Last edited:
  • Like
Likes   Reactions: Greg Bernhardt

Similar threads

  • · Replies 4 ·
Replies
4
Views
1K
  • · Replies 21 ·
Replies
21
Views
3K
  • · Replies 14 ·
Replies
14
Views
2K
  • · Replies 10 ·
Replies
10
Views
3K
  • · Replies 1 ·
Replies
1
Views
2K
  • · Replies 7 ·
Replies
7
Views
2K
  • · Replies 13 ·
Replies
13
Views
2K
  • · Replies 15 ·
Replies
15
Views
2K
  • · Replies 6 ·
Replies
6
Views
3K
  • · Replies 1 ·
Replies
1
Views
2K