And the moderators also have more than 2 categories
Hi, looking at all the other topics, this is a very basic question, but for me it is really a brain cracker..
I have a conceptual model;
Dependent variable
- Opening rate direct mail (0=not opened, 1=opened)
Independent variables (all categorical)
- Day of delivery (days of the week)
- Month of delivery (months of the year)
- Moment of delivery (calendar event, industry specific moment etc)
Moderators (both moderating all three relations between the independent variables and dependent variable)
- Education
- Occupation
Now I want to test the conceptual model with my dataset. In SPSS I use a binary logistic regression model for the relation between the dependent and independent variables. This is clear to me, but how can I test en interpret the effect of the moderators in SPSS? Thanks
And the moderators also have more than 2 categories
Hello km88,
Just as you would with moderators in linear regression, you can test the interaction between two variables (a and b) by adding a new term that multiplies the original terms together (a*b). This term is just added to the model (make sure to also include the original terms a and b), and the t-test will indicate if it is significant or not. The B coefficient is the change in one of the original predictors' B for a one unit increase in the moderating predictor.
This is going to be kind of annoying for you, as you have so many categorical variables, because its going to mean a lot of interaction terms (due to dummy coding splitting each categorical variable up into multiple dummy variables).
If you are using SPSS, one nice feature is that you don't have to actually compute the interaction terms for logistic regression, you just highlight both variables and add them together. Don't forget to indicate which predictors are categorical.
Lots of thanks! And if I want to know if education and occupation are a significant moderator as a whole, is this procedure correct?:
1: First build a logistic regression model with only the independent variables
2: Add to education to the model, and see if it increases the R2, and see if it changes the B's and p values of the independent variables. If not -> no moderating effect
3. Same for occupation.
Mmm not quite. First of all you won't have R^2 in logistic regression because it uses maximum likelihood estimation. To see if adding or removing a block of predictors to the model has a significant impact on the model's fit you look at the difference in the -2log(likelihood) between the two models, which is your step chi-square statistic (df is the difference in df between the models). In SPSS just enter your blocks step by step and the step chi-square will be shown.
Secondly, you have to actually add interaction terms, which are the two variables multiplied together, not just the moderators. Moderation and interaction are different terms for the same thing, but you need to add the moderated variable (a), the moderator variable (b), and interaction term (a*b). All three of these terms must be in the model to properly test for moderation.
So here are the steps:
1. Build model with independent variables
2. Enter block of moderators: education and occupation (the step chi-square tests if these have main effects)
3. Enter first block of all interaction variables with education, such as education*month. The step chi-square for this block tests whether the moderating (interaction) effects of education are significant.
4. Repeat step three for the occupation interactions.
Make sure, again, that you indicate which predictors are categorical.
Last edited by marchhare; 05-25-2011 at 09:09 PM. Reason: Clarification of terminology
Ok, thanks! I think I've got it now.
Hello everybody,
I have a similar qustion as above, but then for ordinal regression analysis.
My dependent variable is ordinal scaled with four categories (lets call him Y)
Then I have the following independent variables:
Gender (male/female)
Age (ratio in years)
Health status (healthy/unhealthy)
My moderator variable (lets call him M) is measured on a ratio scale (score of susceptibility)
In my research, I would like to determine 1) the relations between Gender, age and Health status, and the ordinal Y. 2) whether there is a moderating effect of M.
How do I solve this puzzle? In the solution above you simply used an interaction term of a*b. However, some literature suggest to use 'centered variables'. Can somebody help me to tell me the right steps? I'm analysing my data with SPSS.
Thanks in advance,
Roberto
Regarding the detailed explanation in the previous posts, it is not
clear what is still puzzling you. Moderation is modeled as interaction.
So you add M to the model and for any suspected moderation you add
the M*predictor interaction term to the model.
Kind regards
K.
Thanks for your response! Well I have no idea how to do moderator analyses when your dependent varaible is measured on an ordinal scale and, thus, ordinal regresion had to be used. How can I do this SPSS?
Second, sometimes individuals determine any suspected moderation by M*predictor, but I also have seen that people first 'centered the variables', then multiply the centered variables (multiplication is a new variable) and, finally, add the new variable to the model. Do you understand what I mean?
Last edited by Roberto; 04-04-2012 at 05:41 AM.
Tweet |