1

We have conducted a survey using Likert Scales with the following options:

  • Strongly Agree
  • Moderately Agree
  • Slightly Agree
  • Slightly Disagree
  • Moderately Disagree
  • Strongly Disagree

For analyses purpose, is it possible to group responses as either agree or disagree? Strongly Agree + Moderately Agree + Slightly Agree =>'Agree' Slightly Disagree + Moderately Disagree + Strongly Disagree =>'Disagree'

user34351
  • 11
  • 1
  • 2
  • Could you tell something about the survey or the analysis? It would be easier to give related answers. – Jawa Aug 08 '13 at 07:01
  • We have developed and evaluated (usability and effectiveness) a web-based system with end users. We developed Likert-scale statements to gather data from the respondents. For the analysis purpose, I produce table to show the level of agreement and thought that I will be interesting to group the responses as agree or disagree. – user34351 Aug 08 '13 at 07:46
  • 1
    Quick suggestion not related to your question; consider using an odd number of answers. When you use even numbers you may be "forcing" users to exhibit a bias. This is a rule of thumb with some CS academics I work with. – puppybeard Aug 08 '13 at 09:34

2 Answers2

3

It totally depends from the analysis you are doing. Nobody is going to come after you if you group the responses, but that kind of invalidates the questionnaire - you could have just asked "Agree / Disagree" instead.

The purpose of Likert scale is to find the level of responder's agreement, which would be totally ignored in such a grouping.

Jawa
  • 1,791
  • 1
  • 15
  • 18
  • It might be OK to group responses for reporting purposes though. Such as "33% of people disagreed to a large or small degree", just as you could group both of the 'strongly' opinions together "28% of people had the strongest opinions". It just depends what it is you want to use the results to say. Unfortunately you can report against the statistics in whatever way you want to get the outcome you want. – JonW Aug 08 '13 at 08:27
2

You can do this, but you are probably reducing the quality of your data considerably.

In most cases* where you use a Likert scale, you have to factor in a strong ceiling effect**. Imagine asking something like "Our application is easy to use", and your customers can choose options from "Strongly agree" to "Strongly disagree". If you look at a histogram of the answers, you will see that you have much more answers in the "Agree" part than in the "Disagree" one.

There are statistical ways to discover this and deal with it, but you shouldn't need them if you are not going to publish your results in peer-reviewed journals. But here is the practical way to look at things.

  • The ratio of "Agree" to "Disagree" answers is not interesting, because you will almost always have more agrees than disagrees, even if your software is mediocre. It has to be truly evil to get the ratio reversed.
  • What is interesting is the histogram shape. With a Likert scale, you will get a somewhat skewed Gauss curve, shifted to the right (assuming positive answers are on the right) and cut off on the right. For the best applications, the curve is shifted so far to the right that you cannot see the "hump" of the bell curve and you only see the first slope of it, so you are looking at a upwards sloped line instead of a bell. If you can get your users' metrics to show this shape, you have made it big. Else, you will see the hump somewhere in the right half of the scale. Interesting metrics are how far to the right it is, what percentage of people have given answers falling below the modal answer, and, if you want to go deep into it, steepness/skewness measures. These metrics already give some information by themselves, but they really shine for comparisons (e.g. satisfaction between features of your software, to see what your users really hate. Or comparing your own software to the competition and seeing where yours has to catch up).

The more fine-grained your data, the better you can use these metrics. Two- to four-valued distributions are practically useless for them. But if you go too high, humans are not able to differentiate the own attitudes with enough precision. So questionnaires normally use five- to nine-valued scales (with a strong ideological battle between the "even-" and "odd-number-valued scale" camps). You already have this kind of data, so use it. Clumping it together to just "Agree" and "Disagree" values makes it impossible to draw the histograms and calculate the metrics which give you real information.

Another word of caution for the metrics above: Never calculate means for your data gathered with a Likert scale. Likert scale data is ordinal; treat it this way. Methods like an arithmetic mean are created for cardinal data, and while you will get a numerical result with them, it will have no real meaning, and any reasoning applied on it will be misleading.

[*] You are not saying what you are measuring. I have experience with measuring satisfaction, and related concepts like usability, etc. - generally answers determined by users' attitudes to a product, and I will assume that on this site, you are measuring something similar (besides, this is the canonical use of Likert scales, they were developed for attitudes). I don't know how much the answer applies to some completely different use of the Likert scales.

[**] For the especially interested: Peterson, Robert A., and William R. Wilson. "Measuring customer satisfaction: fact and artifact." Journal of the Academy of Marketing Science 20.1 (1992): 61-71.

Rumi P.
  • 6,224
  • 2
  • 20
  • 31