Binary Logistic Regression with categorical predictors and interactions.

[SonofDrogo]

I'm new to logistic regression and the project I'm working on requires a model that involves two predictors:

• Experimental condition (categorical variable with 8 categories)
• Score on a measure (continuous variable)

So, one categorical predictor and one continuous predictor. But, these two things should interact! So, we get:

• Experimental condition
• Score
• Experimental condition * score

I have one dependent variable that is binary.

When I run my analysis in SPSS I get a Variables in the Equation table that has a lot of information and I'm not sure which information is important. I seem to be getting a B-value for every experimental condition, but no B-value for experimental conditions overall. I'm also getting a B-value for every interaction between the scores and the experimental conditions, but no overall B-value for that interaction.

In other words, I'm not sure what B-values to report when I write up my results.

I was using Andy Field's book Discovering Statistics using SPSS for Windows to walk through logistic regression, but he doesn't explain how to interpret a situation like this.

Sorry if my question seems silly, I am new to this procedure and was not formally educated on how to use it!

Any help would be appreciated!

 

[noetsi]

I only know logistic regression in SAS so these suggestions might not help. SAS has a type III analysis of effects which is essentially a test of the categorical variable as a whole (which implictly test that at least one level of the categorical variable is different from another - like a F test in ANOVA). You should look to see if SPSS has this in its own table (it will probably have a variable named exactly what your categorical variable is named).

SAS like SPSS generates paramater estimates for each level of the categorical variable interaction with the interval variable. But again in the type III analysis it has the interaction as a whole - with a signficance test. Again I think you need to look for a specific table that does this. SPSS probably does what SAS does in that to do this analysis you have to declare the categorical variable is a categorical variable. This will be when you drop your independent variable into the slot in the initial data run.

Your questions are not silly. I have spent weeks working through this recently. I have a bunch of threads on this issue on the SAS board and in the regression board. I don't know if they will help (the code is very different) but you might get suggestions on the type of places to look. One thing that is critical IMHO. If you have a signficant interaction effect your should not focus on the main effects which are difficult to interpret when interaction is occuring. You should instead look at simple effects - the impact of the IV on the DV at a specific (theoretically important) value of the other IV it is interacting with. SAS does this with ODDSRATIO and AT statements - SPSS will do this although I don't remember how (its not obvious where it is in the software).

The Paramater estimates are difficult to explain given that the logit is not an obvious concept to most. Odds Ratios are much easier to understand and you should certainly comment on them.

Wish I could be of more help - but I work most of the time in SAS. UCLA has a series of comments on SPSS and logistic regression. You might look there.

 

[noetsi]

I only know logistic regression in SAS so these suggestions might not help. SAS has a type III analysis of effects which is essentially a test of the categorical variable as a whole (which implictly test that at least one level of the categorical variable is different from another - like a F test in ANOVA). You should look to see if SPSS has this in its own table (it will probably have a variable named exactly what your categorical variable is named).

SAS like SPSS generates paramater estimates for each level of the categorical variable interaction with the interval variable. But again in the type III analysis it has the interaction as a whole - with a signficance test. Again I think you need to look for a specific table that does this. SPSS probably does what SAS does in that to do this analysis you have to declare the categorical variable is a categorical variable. This will be when you drop your independent variable into the slot in the initial data run.

Your questions are not "silly" or easy to answer. I have spent weeks working through this recently. I have a bunch of threads on this issue on the SAS board and in the regression board. I don't know if they will help (the code is very different) but you might get suggestions on the type of places to look. One thing that is critical IMHO. If you have a signficant interaction effect your should not focus on the main effects which are difficult to interpret when interaction is occuring. You should instead look at simple effects - the impact of the IV on the DV at a specific (theoretically important) value of the other IV it is interacting with. SAS does this with ODDSRATIO and AT statements - SPSS will do this although I don't remember how (its not obvious where it is in the software).

The Paramater estimates are difficult to explain given that the logit is not an obvious concept to most. Odds Ratios are much easier to understand and you should certainly comment on them.

Wish I could be of more help - but I work most of the time in SAS. UCLA has a series of comments on SPSS and logistic regression. You might look there.