Skip navigation

Category Archives: Likert Scale

Article by Jamieson, Susan (2004) in Medical Education, 38.

 

Many of my colleagues trapped in the middle of a debate whether or not Likert scale (ordinal[1]) data can be analyzed parametrically[2]. Some of them (who stick to the theory that ordinal can not be treated as interval on whatever reason and basis) even ‘aggressively’ criticized my other colleagues who did so in their research.

 

I did explain to them few times that there are two school-of-thoughts pertaining to this issue. The purist (merely my term) argued that ordinal is ordinal, nothing in this world can make it interval. For this school-of-thought, ordinal must always be analyzed non-parametrically. The other school-of-thought argued that to a certain extent, ordinal and interval can not be distinguished as the ‘border line’ between them is very vague[3] (Abelson (1995) called it ‘fuzzy’[4]). Based on this premise, the second school-of-thought argued that ordinal data can be analyzed parametrically if (and only if) that the data is ‘sufficiently close’ to normal distribution[5].

 

Participating in such old and endless debate is, for me, a waste effort. As a matter of fact, both have proven contributed significantly to the body of knowledge and theory development. As long as we realize the existence of the school-of-thought, and as long as we know which school-of-thought we are in, I believe it is sufficient for amateur researchers like me to continuously active in research. Time and resources should be focused on conducting research where our contribution is possible, rather than busy looking for pitfalls in the other camp’s research.

 

[Note: the first three paragraphs is my personal view, it was not taken from this paper.]

 

This paper discussed the same issue as I described above. The author presented the two school-of-thought by referring to two different views, one by Knapp (1990) and the other one by Kuzon (1996). No conclusion has been made at the end of this paper on whether or not the second school-of-thought is applicable. The author leave the core issue of the debate as it is before, makes it remained as a ‘never-ending story’.


[1] There are 4 different types of measurement levels – ratio, interval, ordinal and nominal. For thorough definition, please refer to Steven, Stanley Smith. (1946) “On the Theory of Scales of Measurement”, Science, 103, pp 677-680. For quick reference, please refer to wikipedia.

[2] There are two types of statistical analysis available, i.e. parametric and non-parametric. Which one to choose is very much dependent to the data properties. For quick and reliable reference on what is parametric and non-parametric tests, go to Wikipedia or just click here and here respectively.

[3] Interesting argument about this issue can be found in Miles, J. and Shevlin, M. (2001), Applying Regression & Correlation: A Guide for Students and Researchers. London: Sage Publication, p. 61.

[4] Abelson, R.P. (10051995). Statistics as Principled Argument. Hillsdale, NJ: Erlbaum.

[5] How ‘close’ is ‘sufficiently close’ to normal? Is there any possibility that we can ‘rest’ from the normality assumption? Please read Sirkin, R. Mark. (2005). Statistics for the Social Sciences. Sage Publication particularly on p. 245. I plan to include my note (here in this weblog) on part of chapter 8 of this book (pp. 244-246), I’m now looking for free time to revise my note before I can post it here.

 

Article by Lozano, L.M., Garcia-Cueto, E., and Muniz, J. ( 2008 ) in Methodology, 4:2.

 

This paper reports the effect on psychometric properties (reliability and validity) of a scale if the number of response alternatives [Note: ‘response alternatives’ here means the choices of answers in the Likert scale] varies.

 

The core findings of this study are:

  1. Regardless to the sample size of the study and the correlation between items in the scale, the reliability of Likert scales will drop if the number of options is reduced. This study found the exception for the case of 3 options Likert scale. Sound explanation is given but I’m not interested to put it here since for so long ago, I’ve adhered to the idea that 3 options Likert scale is not advisable;
  2. The factorial validity (assessed by the percentage of variance explained) has the same behaviour as reliability regardless to the sample size and the correlation between items;
  3. Considering the two psychometric properties, i.e. reliability and validity, the best number of options for Likert scale is between 4 and 7. More than 7 options will give better psychometric properties, but one has to make sure that it will not exceed the discriminative capacity of the respondents. It’s important to note here that a larger number of options will permit the respondents to more clearly express their point of view.
Follow

Get every new post delivered to your Inbox.

Join 28 other followers