Problem with results in ordinal regression...

#1
Hello and thanks in advance for any response...
I have 60 likert-type questions (5 items) (and their responses of course...) and age. Age as raw data is continuous. On a duplicate file I have categorized age in 5 age groups (ordinal) as I want to test whether age can affect the adoption of digital tools in the learning process according to some specific groups (Gen X, Y, Millenials, etc). When I ran Spearman's rho for age as ordinal data and likert responses I got statistically significant results only for 3 combinations (of age and likert responses). However, when I ran the ordinal regression (age as continuous, likert as ordinal) it gave me 22, at least the p-value in the model fitting information was <.05 (unless I shouldn't have looked at this...). In addition, all 22 results have warning from the SPSS that 65% (more or less) cells (i.e., dependent variable levels by combinations of predictor variable values) are with zero frequencies. Should I take the results of the ordinal regression into account? Am I doing sth wrong? Should I just keep the Spearman's rho results? Thank you (and sorry for my english)
 

CB

Super Moderator
#2
Can you explain more about what you're doing with the Likert data? I.e., are you looking at the correlation between age categories and each of the single Likert items?
 
#3
Thank you for your response...
Yes, exactly... I need a correlation between age and each of the single likert items. Another thing is that I've got a lot of missing values that I don't know how to handle. The questionnaire is actually divided in 6 parts. Each part begins with an yes/no question; if the respondent answers yes, then he is 'eligible' to answer the rest of the questions of each part; if the answer is no, she/he has to overlook these questions and go to the next part; however, the questionnaire is poorly designed (it was given us from the faculty), as there are duplicate questions, poorly stated questions with no clear meaning, etc, and as a result, many respondents who answered no, kept on completing the rest of the questions, plus there are some incomplete from those who answered no and those who just didn't answer a question for their own reasons. I know that given the data, there is no much I can do, but there is an assignment I have to submit in any case...
 

CB

Super Moderator
#4
Thank you for your response...
Yes, exactly... I need a correlation between age and each of the single likert items.
I'm sorry but that doesn't sound right. Each of sixty different items, really? That'll be a huge analysis with great risk of Type 1 error. Surely the items are intended to measure some defined construct, or set of constructs? Usually we could collate multiple items into a single scale.

Another thing is that I've got a lot of missing values that I don't know how to handle. The questionnaire is actually divided in 6 parts. Each part begins with an yes/no question; if the respondent answers yes, then he is 'eligible' to answer the rest of the questions of each part; if the answer is no, she/he has to overlook these questions and go to the next part; however, the questionnaire is poorly designed (it was given us from the faculty), as there are duplicate questions, poorly stated questions with no clear meaning, etc, and as a result, many respondents who answered no, kept on completing the rest of the questions, plus there are some incomplete from those who answered no and those who just didn't answer a question for their own reasons. I know that given the data, there is no much I can do, but there is an assignment I have to submit in any case...
Missing data is a complex topic - this book is a good reference. Aside from that you should probably talk to whoever designed the survey and find out what they intended to be done with the data given these design features.