I have ten factors to test with logistic regression, 4 of them are an intrinsic factor, and six are extrinsic (those are independent variable). Should I run all of them at one step or run with the intrinsic alone then the extrinsic.
Depends on your research question an whether or not your goal is prediction or explanation. See attached. Best, D. Booth
@Salah
I am not 100% sure about your query. But when you are performing logistic regression to get adjusted odds ratio then you should use one dependent variable and all the relevant independent variables to the dependent variables.
As David Eugene Booth points out, this depends on your research question.
And, as David also points out, it depends on whether your research question regards explanation or prediction. To this end, I would point out that I see nothing about the criterion – two groups, more than two groups, etc.
However, I will address prediction for now, as that is the goal I most often see in a LR analyses. This regards how accurately your LR model classifies Ss into the criterion groups. Such hit-rate analyses have been SOP in discriminant analysis for many years. In addition, a cross-validated estimate of such accuracy is usually preferred. This, then, has spawned use of some matrix algebra simplifications to create such cross-validated statistics through the work of Lachenbruch (U method, based on the matrix identity due to Bartlett) and Huberty (Leave-One-Out), also more generally mentioned by Mosteller and Tukey. In the case of LR, the same round-robin subject deletion and validation is possible, but no matrix simplifications are available for the maximum-likelihood estimator used. Yet one can still do the repeated prediction of each of N subjects from the remaining N-1 subjects, there is just more arithmetic. If your research question regards whether classification hit-rate (overall, and/or for each separate group) is incremented by one variable subset in respect to another (for instance, do the intrinsic variables increment classification accuracy significantly over that available from the extrinsic variables, or, of course, asymmetrically, the obverse), Carl Huberty and I described process and code for DA for this here. The test is identical for LR or any classification algorithm.
Article Full Versus Restricted Model Testing in Predictive Discrimin...
Similar code, including that for LR, although for the comparison of LR and DA, was described here:
Article Predictive Discriminant Analysis Versus Logistic Regression ...
If explanation is you goal, the answer is, perhaps more complicated, although there is one, or more answers for that.
My best on your research,
John
Try to run each of the response variables in a separate model (explained by the independant variables). So, in case the other responses are assumed to be explanatory variables for one or more responses, then include them together with the six.
To get the crudes odds ratio you can run the model with each of your dependent variable.
But, for the adjusted odds ratio you need to run the model with all of yours independent variables.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS