I am studying student learning outcomes from an on-location learning experience as a longitudinal design focused on autobiographical memory.
I would like to split my subjects (N=191 at pretest and posttest, and N=71 at follow-up) into meaningful groups, to analyze differences on predictor variables for memory qualities and experience for those that gain from the isolated event, and those that don't, with a simple ANOVA or t-test. I am also splitting them according to how much they report using an episodic memory strategy (7-item 5-point likert scale) to recall information from the learning experience.
What are the conventions for this kind of split?
I'm thinking cutting the cakes 3 ways. I have considered using everyone with an average of 3+ (agree and strongly agree to the items) or the top and bottom quartiles versus the middle ones.
Any other conventions/tricks? What would you recommend?
(I have a pretty complex hierarchical data structure, with far too many predictors in the mix, and some ridiculously complex analyses of the dataset as a whole. This is my attempt at doing something more simple, to isolate the results for students who actually gained from the experience or who seem to qualitatively differ in their recall strategy.)
Hi Andreas - your idea about splitting per the quartiles is something often done.
When I've dealt with Likert scale responses in various survey-based data collection situations, if I've wanted to categorize participants based on high versus low scores, I've typically performed either a median split - or - grouped according to quartiles.
I see where it makes theoretical sense to group participants based on their response averages relative to the middle of the Likert scale (value of 3), but in my experience participants as a whole rarely respond with a purely normal distribution that results in an average that coincides precisely with the middle of the Likert scale used. The median split helps to get around this and categorizing participants relative to the overall median also helps to limit the impact outliers have on your categorization method.
As with anything in statistics, there's considerable controversy with the suggestions I've made. While it's common practice in most social sciences to create categories from data by grouping per a median split or quartiles, there are good arguments against these practices. For example, some suggest using regression or other methods instead of ANOVA procedures to avoid having to use categorical grouping.
To follow up a bit.
Likert responses are not normally distrubuted so you will necessarily run into various statistical issues. Basically there really is not an equivalent to the "standard deviation" in the Likert scale (yes you can compute one mathematically but it doesn;'t mean anything)
Regression, in my experience, is of limited value since the Likert scale has very little dynamic range.
I do think that forming categories based on medians is helpful and more importantly, can be done in a consistent manner.
I concur, with G. Bothun's idea about medians - but the convention for working with Likert scale data varies from field to field. In many social science fields, for example, Likert scale data is typically thought of as interval - rather than ordinal. As a result, it's not uncommon to work with the mean and standard deviations.
Treating Likert scale data as interval is advantageous for a variety of inferential procedures, especially ANOVA and MANOVA. Yes, there are non-parametric statistics that can work with ordinal data (and use the median for computation). However, these can be more cumbersome to work with depending on the statistical software one is using.
When it comes to Likert scale data, my advice is to always follow the conventions established in one's specific field.
I concur with the prior comments on selecting the cuts. I have used both the high vs te low scores and the upper 1/2 vs the lower 1/2.. One problem you have witth a lLikert scale is that you have the non comitted group. If you do not use them you lose considerable sample size and it appears you do not have a large sample. If I understand correctly you have 191 where you have both pre and post test scores, but only 71 with follow-up data. Thus you are goiong to be limited in your power to determine whether longitudinally you had any effect of your intervention. Thus any cut you make should include all participatents. I would suggest then that you look at your agree vs all other responses. Those who had a neutral response are then treated the same as those with disagree responses. This is not an unreasonable assumeption that they were not impacted by your program.
You mention that you have a lot of predictor variables. The more you have if treated individually the higher the chance that you will find a relationship, whether it is meaningful or not. I would consider a factor analysis to see if you could group some of your responses into categories to reduce the number of variables. Without knowing what you have it is impossible to say how to do such groupings. Usually you can form scales consisting of a number of questions that are closely related to forms stronger predictors. Factor analysis is fairly standard with many software programs available. Might be worth a try. Good luck!.
Dear\ Andreas Lieberoth
pls read this article :-
- http://ijqr.net/journal/v12-n2/8.pdf
- https://www.st-andrews.ac.uk/media/capod/students/mathssupport/Likert.pdf
- https://www.uth.tmc.edu/uth_orgs/educ_dev/oser/L1_2.HTM
Good luck
- Badges
- Science topic
- Similar topics
- Psychology