I am completely in over my head with logistic regression at the moment, so what follows is probably very basic and silly questions. But I would appreciate it hugely if anyone took the time to respond nevertheless!

I will be conducting a cross-sectional analysis involving data from a large cohort study. My goal is to look at the impact of one categorical (binary) predictor variable on five different categorical outcome variables, all of which can be considered different proxy measures for the same concept (hence why I am looking at all five, as no direct measure was collected in this data set - in essence, I am trying to corroborate my results by looking at the impact of the predictor variable on five different outcome variables supposedly measuring of the same/related concept(s)). As mentioned above, all of these variables are categorical. My outcome variables have different numbers of categories, ranging from 2 to 4. The 3- and 4- category variables are ordinal. All of these variables (both outcome and predictor) are basically items on self-report questionnaires.

Does this sound way over-complicated and mad? If this sounds like a terrible plan, I could potentially get rid of some of the outcome variables and just use fewer ones instead, although ideally I would use all five. I am currently completely lost with regards to how to analyse this. The only way I can think of would be to run separate logistic regressions for all of these outcome variables.

Any advice on how to analyse data like this would be hugely appreciated!


2 Answers 2


If the 5 outcome variables are meant to measure the same latent concept, the questions you have to ask is:

From a research perspective, are you interested in assessing the impact of the predictor variable on this latent concept (which is measured indirectly via the 5 outcome variables)? Or are you interested in assessing the impact of the predictor variable separately on each of the five outcome variables?

How you analyze your data will depend on the answer to these questions.

  • $\begingroup$ Thank you very much for the response. I am mainly interested in the impact of the predictor variable on the overall concept. But ideally I would also be able to analyse whether there are differences between the outcome variables, in terms of the extent to which the predictor is associated with each of them. I hope that makes sense. $\endgroup$
    – Amelia M.
    Commented Mar 31, 2018 at 17:46
  • $\begingroup$ Then you may need to consider a modeling framework called Structural Equation Modelling (SEM). For a reference, see "Introduction to Structural Equation Modeling using IBM SPSS Statistics and AMOS" By: Niels J. Blunch ( 2013 | Second Edition). $\endgroup$ Commented Mar 31, 2018 at 17:56
  • 1
    $\begingroup$ Or latent response models. $\endgroup$
    – Björn
    Commented Apr 1, 2018 at 5:13

You say your outcomes are categorical: Are the 3- and 4-category variables ordinal or nominal?

In the nominal case, you can do this in Mplus. Declare the outcomes as nominal and regress them all on the predictor. Then test the constraint that all the regression coefficients are equal to zero, using WLSMV difference testing.

I don't know of a canned package other than Mplus that can do this for nominal variables, but lavaan can also do it in the ordinal case.

  • $\begingroup$ Thank you so much your reply. I should have been more clear - they are ordinal (I will edit my original question now). I am also a total statistics newbie it appears as I only have access to SPSS. $\endgroup$
    – Amelia M.
    Commented Mar 31, 2018 at 17:22
  • $\begingroup$ I can't speak to whether this can be done in SPSS. lavaan is part of R (free) but might be intimidating to jump to straight from SPSS. $\endgroup$ Commented Mar 31, 2018 at 17:25
  • $\begingroup$ At this point I am ready to cry with trying to work this out so I would give R/lavaan a try (I have used R briefly in the past a few years ago). Would you be able to give any advice as to how to go about doing this (of course I am not expecting you to write it all out for me, but any tips would be welcomed). Thanks again! $\endgroup$
    – Amelia M.
    Commented Mar 31, 2018 at 17:36
  • $\begingroup$ This is a lavaan tutorial by the creators: link . There is a page for declaring data to be ordered. This defaults to the WLSMV estimator. The "Model Syntax 2" page shows how to create a constrained model. $\endgroup$ Commented Mar 31, 2018 at 17:44
  • $\begingroup$ The last answer got away from me too quickly--there's a bit more information now! $\endgroup$ Commented Mar 31, 2018 at 17:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.