Dear all,
I noticed that some articles use ANOVA for Likert scale data and also test normality for the results and take the mean for the data?! as we Know Likert scale is an Ordinal Scales tell us relative order, but give us no information regarding differences between the categories so we don’t have real zero and don’t have mean. I feel there is mistake?! So what are the best tests should be done to Likert scale survey in addition to descriptive statistics.
Distinguish between DVs and IVs. For IV Likert scales are fine for ANOVA/OLS regression models. As a DV a Likert scale won't work in ANOVA/Regression very well. In that case you have generalized linear models ( eg various forms of logistic regression) which will work quite well. See the notes below for details. Best, David
Data from individual Likert items are generally considered to be ordinal in nature. That is, as you say, they are ordered categories, but the spacing between categories is not known. If numbers are assigned to responses, it is for convenience. That is, we do not know that “Strongly agree” is twice as far from “Neutral” as “Agree” is.
So, for individual Likert items, analyses appropriate ordinal data should be used, and I think this norm is usually followed.
When several Likert items are combined into a Likert scale, traditional parametric statistical tests are often used.
You can see the theoretical problem here. If Likert item responses are ordinal, and Likert scales are made up of Likert items, then Likert scales at best can be ordinal.
However, there are a few things to consider to justify the switch.
First, the only assumption that needs to be made to make Likert items into interval data is that the categories are equally spaced. I don’t know if people apply this assumption intentionally. But for most common Likert items, I think it is possible to reasonably apply this assumption. You can try it yourself, by, say, taking a 5-point Likert and try to translate in to ranges of a 100-point scale, as you see it.
Second, we probably shouldn’t be too strict in interpreting the distribution of data. No real is normally distributed. Repeat, no real data is normally distributed. It is just a matter if our data follow a conditionally-normal distribution well enough to make using a certain set of tests reasonable. Responses from a well-constructed scale may have a normal-like distribution.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA