SPSS Statistics

 View Only
  • 1.  Can't get multinomial logistic regression to work to control for dominant variable

    Posted Fri April 02, 2021 12:05 PM
      |   view attached

    Hi there, I'd be very grateful for help with a problem. I have some independent or predictor variables, and I am testing for a statistically significant relationship with various dependent or outcome variables. Some of the predictor variables are dominant - I know they have a strong influence on results. So when some of the other predictor variables are found to have a statistically significant correlation, I don't know if that is simply because they are in turn associated with the dominant variable. That is, are all the observed effects explained by the dominant variable? So I am trying to control for the influence of the dominant variables by using multinomial logistic regression and testing for interaction effects.

    The variables are mostly nominal categorical variables or dichotomous Yes/No variables. This has made it harder to find a method and get help, since most guidebooks assume that users have continuous variables.

    I ran the exercise in SPSS via Analyse > Regression > Multinomial Logistic. I only used two predictor variables to test. Before running the model, I tried to add for interaction effects by going to Model then selecting Custom/Stepwise and adding either 'Interaction' or 'All 2-way' for selected variables.

    However, when I add interaction terms, I don't get a result in the Likelihoods Ratio box for several of the variables. (I read that the Likelihoods Ratio is what I should look at to tell me if my nominal/categorical variables have a statistically significant relationship with the outcome variable.) Instead, SPSS gives a message saying 'This reduced model is equivalent to the final model because omitting the effect does not increase the degrees of freedom'.

    When I don't specify interaction effects, I do get results in the Likelihoods Ratio box for both variables. But this doesn't help me; I want to test for interaction, like I said.

    I looked online for why I am getting this message, and one forum suggests it is because the variables are limited. But the variables that I tested are fairly simple and meaningful variables: one is a simple Yes/No dichotomous variable and the other has 3 categorical options. 

    Does anyone know why SPSS is doing this? Or if there is a better way to 'control' for the effect of dominant predictor variables so I can see if any other variables are having an effect?



    ------------------------------
    Rebecca
    ------------------------------

    #SPSSStatistics


  • 2.  RE: Can't get multinomial logistic regression to work to control for dominant variable

    IBM Champion
    Posted Fri April 02, 2021 01:32 PM
    It's hard to know exactly what is happening, but the message seems to be saying that you have redundant categories in the model.  For example, if you have the equivalent of a basis of dummy variables but also have an intercept, one category is redundant and would be dropped - without a change in the df.

    To pin this down, you could run linear regression with all the categorical variables as dummies and see what gets dropped.  

    --





  • 3.  RE: Can't get multinomial logistic regression to work to control for dominant variable

    Posted Fri April 02, 2021 02:12 PM
    Thank you very much. I don't know what is an 'intercept' - could you clarify? For this particular example I only used two variables; both are dichotomous Yes/No and they are not at all related to each other. So if by 'redundant' you mean the two variables might be saying the same thing, I don't think that can be possible.
    My Forced Entry terms were: Variable 1, Variable 2, Variable 1*Variable 2.





  • 4.  RE: Can't get multinomial logistic regression to work to control for dominant variable

    IBM Champion
    Posted Fri April 02, 2021 03:01 PM
    The intercept is just the constant term, usually included automatically e

    ------------------------------
    Jon Peck
    ------------------------------