I was reading the literature and I found it that there is a debate about what test to use on likert scale (five point scale). If I want to test if their is a significant difference or relationship of the data to the profile of respondents (e.g.age, sex, yr level).
Dear Lloyd,
If you are computing the total score by adding up the scores on the individual items, then you will have a scale variable that has too many values, and thus can be treated as a continuous variable. Therefore, you can chose any of available analyses to answer your study questions.
Best regards.
Mutasem Akour
To expand on Jochen's answer. If you have a 1-5 rating then a model that treats this outcome as ordinal will be more appropriate than one that assumes that the response can vary freely from minus infinity to infinity (e.g., a t tests, ANOVA or linear regression). The latter models break down particularly for interval estimates and prediction or if the majority of values are close to the upper or lower bound. Under certain circumstances (hypothesis testing with the distribution of ratings fairly symmetrical and unimodal) approaches like t tests can work OK.
To expand on Thom's answer :)
"approaches like t tests can work OK" - these are tests on the random variables constructed from the Likert values. Likert values are not numeric. By assigning numeric values to the Likert values we construct a numeric random variable. This assignment is completely arbitrary, we may assign any number to any Likert value (we may, for instance, assign 127.5 to "somewhat agree" and -31.8 to "fully disagree"). The assignment should, however, make sense in the context of the research question (that's outside the realm of statistics that is concerned only with the properties of the constructed random variable, and not with the rules for its construction). Because researchers often have no idea how to meaningfully assign numeric values to Likert values, they often use the order number of the Likert value (like 1...5 or -2...+2 or something similar). Whatever rule is used to construct the random variable, the test is about this random variable and not about the Likert data. This must be clear for the interpretation of the results.
Ordinal logistic regression is the (almost only) tool that analyzes Likert data. (Almost) all other tools analyze a constructed random variable (with arbitrary construction rules that may or may not make sense!).
Since everyone is "following up" in this thread, I'll follow up on Matusem's answer. ;-)
Lloyd, it is not clear to me whether you are analyzing Likert-type items, or Likert scales. For more on the distinction, see the link below. HTH.
http://www.john-uebersax.com/stat/likert.htm
I'll add my follow up to Bruce Weaver's point.
If you have only a single Likert-scored item to analyze, then I would agree with treating that as an ordinal-level variable. But if you have a set of Likert-scored answers, then I would recommend combining them to form a scale, which you could use as an interval-level variable.
Bruce,
Thank you for pointing that out (whether Likert type or Likert scale)..I found this blog article that explains the difference of the two.. https://www.joe.org/joe/2012april/tt2.php
My follow up question would be, is it okay to sum all the means for a Likert type or Likert scale which have unequal number questions. For example, I would like to ask my respondents on their perception on the following aspects: OBJECTIVES, MONITORING AND EVALUATION, PERSONNEL, ACTIVITIES, FACILITIES. Under each aspects has unequal number of questions. 5 questions for Objectives, 8 for monitoring, 10 for evaluation, 8 for personnel, 9 for activities and 5 for facilities.
This is where I got confused. Is it okay to sum all of the questions for each of the aspects or just leave it as Likert type data.
Thanks for all your reply.
https://www.joe.org/joe/2012april/tt2.php
Because there have been so many questions about creating scales from Likert-scored items, I have collected a set of resources on this topic:
https://www.researchgate.net/post/How_can_one_help_people_create_simple_scales_from_Likert-scored_items
David,
Thanks for the thread.. I just realized it now that it is needed to conduct first PCA or factor analysis before incorporating one Likert item to a single Likert scale.. thanks a lot..
Just to give a heads-up, most of the time PCA and factor analysis (Exploratory FA) are used interchangeably, which is not correct. This confusion may be because of the fact that in SPSS both of these tests are done using same menu options.
- Badges
- Science method