I am planning on doing an ordinal regression with -obviously- ordinal variables. The ordinal variables are 'overarching variables/ concepts'. So, I took the mean of multiple questions answered on a Likert scale to make a new variable which is the concept.
My question : should I round the averages of the overarching concepts before I carry out the ordinal regression?
Otherwise you will have averages like 1,3 ; 1,4 ; 1,7 ; 2,3 ; 2,5 etc. which seems unworkable..
Greetings,
The procedure you followed is not statistically correct, you cannot take mean values out of likert type questions since the difference between the categories is not constant, or measurable since it expresses opinions. Better it would be to use factor analysis and use the factors as variables in the ordinal regression model.
There's no need to round the values. Ordinal regression shouldn't care what the ordinal levels are called, whether it's 1, 2, 3 or 1.1, 7.6 ,21.17, or "happy", "sad", "miserable", as long as it knows the order.
Of course, your software may feel differently.
Hi Marieke,
You should consider a Structural Equation Modeling approach.
First, you should conduct an Exploratory/Confirmatory Factor Analysis for categorical data to identify the latent variables that your items are measuring. Then, you can do the substantive analysis using Structural Equation Modeling based on the latent measurement model that you identified in step 1).
Roberto
Thank you all so much for taking the time to answer!
The different questions are measured on exactly the same scale. Is it really not possible to just take the mean? I have no idea what to do with a factor analysis in SPSS... Also, I've read on multiple sites that you cannot perform a Confirmatory Factor Analysis in SPSS
@Salvatore so it is oke to have a table with all these different answers? I am not sure how to interpret it..
Maybe I can treat the averages as continuous independent variables , since I have this large range of answers
Hi Marieke,
A more traditional approach that would avoid the use of CFA and SEM would be:
1) If the number of categories of your items is five or more, you could assume they can be modeled as they were continuous.
2) Conduct an Exploratory Factor Analysis by using Principal Axis with an oblique rotation (here you can use Promax or Oblimin).
3) Based on the results obtained in 2) you have to decide if the set of items can be considered unidimensional (one factor). If not, you have to identify the items that define the factor(s) that is(are) more relevant to your research question.
4) Compute a factor scores based on the factor(s) that you identify in 3).
5) Now, you can use the factor scores as the dependent(s) variables in your analyses.
Roberto
There are two assumptions you have to make in order to take the average of Likert item data: 1) The categories represent an underlying continuous variable; 2) The categories are equally spaced.
#1 is probably almost always reasonable for Likert items. #2 I would argue is probably reasonable for the categories used in most Likert item questions. Think about how you would convert those categories (SD, D, N, A, SA) to a number on a 1 to 100. My tendency is to divide them into equally spaced categories.
* * *
If your results have many different values, it might be just as well to treat them as a continuous variable. Just understand what assumptions you have applied to get to that point.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression