ordinal or multinominal regression?

#1
This question may sound bit novice but I find it bit tricky
There are 8 continuous predictor variable and
My dependent variable is:
not used
slightly used
highly used

Is it ordinal or nominal?
Can I use multi-nominal regression in my case?
I am confused when i code my dependent variable them as 1,2 and 3 (kind of low,medium and high which gives me a feeling as if the order matters but when I say not used, slightly used and highly used it seems order does not matter.

Can anybody help?

thank you
 
#5
But with that I am getting warning there are 66% zero frequencies data as I have 10 predictor variables, very low number in category 0(10%) and only 200 cases.
Since there is warning it says my model may not be conclusive or I cannot conclude anything.So I am able to do that.
It is observed that i need around 1000 cases to avoid that warning which does not seem possible in my case as it is kind of academic thing.
 
#7
Can anybody know what your data looks like? What frequencies is it in the 3 categories?
Where did you "observed that"? Where does that comes from?
I have a model with a ordinal dependent variable: low, medium, high (coded as 1,2,3).

There are around 13% data with 1, 55% with 2, 32% with 3)

n=200

and there are 10 continuous predictor variables.

I carried out ordinal regression. But since n

is relatively small, dependent variable has 3 variations and data in one category(1) is low, it is giving some warnings (SPSS):

There are xx cells (i.e., dependent variable levels by observed combinations of predictor variable values) with zero frequencies.

The data were collected using survey. So, I cannot start data collection again and add data (as it is time bound study). And I found it is virtually impossible to eliminate those warnings with out adding cases (rows).

In goodness of fit:

pearson Sig 0.005
deviance Sig 1.00

Deviance significance is good but pearson is not.

I am trying to investigate significance (p value) and relation (effect) of each predictor variable on dependent variable.

So, what can or should I do or what can be done?

Can I ignore the warning and use the model and significance?

As an alternative: I carried out multiple regression every thing is fine. 8 variables are found the be significant. Model significance is less 0.05 (good).

Can I go with it? Totally stuck.
 
#8
It is observed that i need around 1000 cases to avoid that warning
Where did you get this from? From the software? From a friend? From....?


Maybe you could collapse category 1 and 2 into one category and run binary logit.

Maybe a LASSO-regression. Or, if there is a lot of multicolinarity, maybe create principal components from the iv:s and use that.
 
#9
Where did you get this from? From the software? From a friend? From....?


Maybe you could collapse category 1 and 2 into one category and run binary logit.

Maybe a LASSO-regression. Or, if there is a lot of multicolinarity, maybe create principal components from the iv:s and use that.
Based on those data it was suggested by one of the statistician that I may need at least 95-100 data for each predictors and I have 10.

Ya that is what I did(merged kind of two lower ones) and every thing is looking good but problem is now I cannot comment any thing with in that little bit difference in the group I merged.
I am thinking of keeping it to multiple regression(assumptions are being met) so that I can have three groups on 1,2,3.
I also see that lots of papers have done so(used multiple regression on dependent variables that has three categories).
how about that?
if nothing works I will settle with binary logistic regression(the results are almost same in multi reg and binary logistic reg). But may be people may comment on use of particular method.
Thank you
 
#10
Based on those data it was suggested by one of the statistician that I may need at least 95-100 data for each predictors and I have 10.
Interesting! I wonder where that "statistician" got that from. In previous nonsense someone said that you needed 10 observations per variable. That is also wrong. I want to remind you that in a Plackett Burman it is enough with 12 observation for 11 variables. It depends on the
data, the design, how the experimental points are arranged. On the other hand, with the very colinear Longley data set, the data are so colinear that it is difficult to estimate the parameters.

It is irrelevant to run regression on 1,2,3 data. It can be arbitrarily recoded as 1,2,4 or as 1,2,1000. They are all ordinal.
 
#11
Interesting! I wonder where that "statistician" got that from. In previous nonsense someone said that you needed 10 observations per variable. That is also wrong. I want to remind you that in a Plackett Burman it is enough with 12 observation for 11 variables. It depends on the
data, the design, how the experimental points are arranged. On the other hand, with the very colinear Longley data set, the data are so colinear that it is difficult to estimate the parameters.

It is irrelevant to run regression on 1,2,3 data. It can be arbitrarily recoded as 1,2,4 or as 1,2,1000. They are all ordinal.
Somewhere i read that "all models are bad but it is just how much until it is totally unusable". Especially academic research, I have found it is as if getting the enough backups for the claim. And some are still indecisive(you get plenty references on either sides)
As I will not be able to collect more data and I see zero frequencies data won't go away with out that(i.e. no enough data to conclude that using ordinal regression), I am going to use binary regression for now.

When you say it is totally irrelevant to run multiple regression, I think the question is again like discussion either likert scale(1-5) scale is either ordinal or continuous.Can we get a mean of that etc. I have seen so many papers (even in reputed journals) using multiple regression for dependent variable(1,2 and 3). So, I now understand why it is called RE-search. :) thank you