Requirements for using ordinal logistic regression

[Alexander Greiner]

Hello everyone!

I have a data set with N=900. All IVs and DVs are ordinal-scaled (4-point assessments of a questionnaire).

I want to determine how the assessment of a set of specific characteristics (IVs) affects the assessment of some general trends (DVs):

Example Variables:
DV: Assessment of the influence of digitalization on the compatibility of private and work life (Ordinal scale, 4 points)
IV1: Assessment of the importance of flexible working hours (Ordinal scale, 4 points)
IV2: ...
IV3: ...

Can I use ordinal logistic regression with the data as is, or is a transformation of the IV variables (dummys/nominal factors) necessary or useful? Different sources report a need for a certain scale level of the independent variable to use ordinal logistic regression.

Hoping that my comments are precise enough and allow for an assessment: Are there other statistical tests or Stata commands other than one-sample non-parametric tests that might be of interest to me?


Thanks so much in advance, any help is highly appreciated.

Best regards,

Alexander

[Andrew Musau]

As long as the dependent variable is ordered, there is no restriction on the independent variables. Scaling is a completely different matter. If the coefficient on a variable is too small, rescaling the variable may make sense. When there are large differences in the values of a variable across units and the variable takes on strictly positive values, e.g., GDP in a macroeconomic panel, log transformation is common. But the choice of estimator largely depends on the nature of the dependent variable.

[Carlo Lazzaro]

Alexander:
welcome to this forum.
As and aside to Andrew's helpful advice, as far as ordered predictors are concerned, you might be inerested in -pca- (see also https://www.theanalysisfactor.com/pr...l-scale-items/).

[Nick Cox]

On a different level,

1. Indicators (you say dummies) can’t be usefully transformed any way. Any one to one transformation whatsoever is linear and so does not help fit. (Any such transformation is summarized by two distinct points on a scatter plot and thereby linear.) Sometimes people prefer say -1, 1 coding to 0, 1 coding but the fit is equivalent.

2. How would you transform nominal factors that are represented by such variables any way? If even an ordering is not implied, there is no scope for genuinely different transformations.

Can you give references for this strange idea? A duty to achieve something that is impossible is no duty at all.

[Alexander Greiner]

Andrew Musau Since both IV and DV are assessments on a 4point scale, there are no large differences in value. Thank you for the useful tips.

Carlo Lazzaro Thank you! I looked into the article you referenced

For pca based on polychoric correlations, it describes that this is a perfect fit for Likert-scale-items that "assume the variables are ordered measurements of an underlying continuum."

The data collected are not typical Likert-scale items, as the middle category "neutral/don't know" was not included in the survey.

Instead, there are the expressions: [positive / rather positive / rather negative / negative]. I assume that the distance between [positive-rather positive] is smaller than the distance between [rather positive-rather negative], which makes it more difficult to assume an underlying true continuum.

Anyway the article you referenced also states: "They don't need to be truly continuous ..."] so I will dig further into that. Thanks again


Nick Cox Thank you for your response.

I meant whether it makes sense to transform the ordinal-scaled predictors into binary dummies, or whether I can use the 4point-scaled predictors as they are for an ordered logistic regression.
I have expressed myself unclearly here, sorry.

I am reading up on a number of different approaches to dealing with ordinal variables and really appreciate the help of the forum.

[Nick Cox]

You were clear. I was misreading you. Sorry.

[Richard Williams]

Given that the IVs are all ordinal, you may want to consider whether it is ok to treat them as continuous. See

https://www3.nd.edu/~rwilliam/stats3...ndependent.pdf

Better yet, see the revised published version at

https://methods.sagepub.com/foundati...dent-variables

If your library doesn't provide free access and you want the article, PM me with your email address,

[Weiwen Ng]

...

The data collected are not typical Likert-scale items, as the middle category "neutral/don't know" was not included in the survey.

Instead, there are the expressions: [positive / rather positive / rather negative / negative]. I assume that the distance between [positive-rather positive] is smaller than the distance between [rather positive-rather negative], which makes it more difficult to assume an underlying true continuum.

...

I meant whether it makes sense to transform the ordinal-scaled predictors into binary dummies, or whether I can use the 4point-scaled predictors as they are for an ordered logistic regression...

Can I use ordinal logistic regression with the data as is, or is a transformation of the IV variables (dummys/nominal factors) necessary or useful? Different sources report a need for a certain scale level of the independent variable to use ordinal logistic regression.

First, a minor detail: I don't believe that Likert items require a neutral point. And technically a Likert scale is a collection of responses to several Likert-type items (item can be read synonymously with question), but most people do conflate Likert scale and item.

Then responding to the second and third points in the quote above (3rd point is pulled from the original post): You stated that you have one dependent variable. It's an ordered categorical variable. Thus, your logical first choice is either ordered logistic regression or generalized ordered logistic regression. The latter model has some less restrictive assumptions than the former. Read up on Richard Williams' site if you like.

As to the independent variables, you can just enter all of them as categorical variables, e.g.

ologit DV i.IV1 i.IV2 i.IV3 ...

Stata creates the dummies for all the levels of the IVs automatically. No transformation needed, as already explained to you.

If your IVs are multiple items measuring the same underlying construct, e.g. various different aspects of how important flexible working hours are, then it might be justified to combine the applicable items/questions into a scale. If the team originating this survey chose to use an established scale, then do that. If it didn't, then unless you're willing to put in the time and effort in to learn about this sort of thing, I think it's probably best to just leave the IVs as they are for now. There's nothing wrong with that.