Does anyone know how you can calculate the significance using the 2LL ?
With a chi² test but how?
Im using SPSS and it is for a binary logistic regression where i have an interaction in the second block. An interaction with 4 dummie combinations under it. So I want to see if the general interaction model is significant, and the seperate dummie combinations.
So far as I know, you evaluate how much the -2LL changes from model with*out* the term and model *with* the term. How ever many terms you add (e.g., n terms), that's the number of degrees of freedom you use for testing the change in -2LL as a chi-square.
Change_in_-2LL = chi-sq(n)
Best wishes,
Damian
If I add an interaction between 2 categorical variables with each 3 categories, how many terms would this be?
Instead of terms, think in coefficients in the underlying linear model; you can visualize them by making a two-way table of the means (or what plays the role of means: odds in logistic regression, and so on) in the differents combinations of your categorical variables, and see how many coefficients you need to model it. Degrees of freedom will be the number of estimable/linearly independant coefficients you need to caracterize the model.
3 * 3 = 9 coefficients for interactions (9 cases in the 2-ways contingency tables of your 2 predictors), but 1 line and 1 column are already used by the « main effects » terms (don't forget them...), remains 2 * 2 = 4 free coefficients in the model = 4 degrees of freedom for the test. May be more if you are using multinomial or ordinal logistic regression.
I do not undestand well what you mean by « So I want to see if the general interaction model is significant, and the seperate dummie combinations », so check carefully that your models are suited for likelihood ratio tests (are they nested?) and what coefficients in your model are involved between the two models.
4.
The interaction is modelled by 4 indicator variables and 4 coefficients. The model without interaction will thus have 4 coefficients less then the model with interaction.
Generally: if you have 2 factors, one with k levels (categories) and one with m levels, then the interaction of these factors will have (k-1)(m-1) coefficients in the model.
@ Jochen: if main effects terms are present, and to be precise, the number of coefficient depends on the coding of the model (which I forgot to mention in my answer), but not the number of degrees of freedom (if main effects are present), except eventually in some incomplete designs (some empty cells).
For a « reference » coding, like in R, there are 4 coefficients, 4 degrees of freedom. For a « complete » coding like in some SAS procedures, there are the 9 coefficients, but only 4 are independant/estimable if main effects are present, hence still 4 degrees of freedom.
I don't know which coding uses SPSS by default.
Counting number of degrees of freedom is not such an easy thing :(
Thanks
"so check carefully that your models are suited for likelihood ratio tests (are they nested?)"
Well I have first the control variables and the main variables. 4 controls, 3 main ones.
And then in a seperate block I have the interaction between 2 of the main variables, which creates 4 dummie combinations.
Should be OK if you are testing the 4 dummies at once and not trying some fancy combinations...
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS