Beyond knowing mean, median, mode, and standard deviation, what else do you hope to learn?
Dear Readers,
Currently, I am conducting a study to investigate the need for assistance with a certain process. I used a survey including Likert-items for my study. In this way, I want to be able to rank the aspects of the process from 'most assistance needed' to 'least assistance needed'.
For example (simplified):
Question 1: I want assistance with writing
Question 2: I want assistance with speaking
Question 3: I want assistance with hearing
All these questions use a Likert-scale (1 being strongly disagree to 7 being strongly agree). I want to use a statistical method to rank these several aspects (writing, speaking, hearing). Thus, I want to be able to state that 'the need for assistance for writing was the highest, followed by hearing and speaking required the least assistance'.
Although I think the means would give a valuable overview, I believe that this is not the best way to rank these aspects.
So does anyone have an idea which statistical method I can use best to be able to rank Likert-items (preferably with references)?
Thank you in advance!
Beyond knowing mean, median, mode, and standard deviation, what else do you hope to learn?
MvZ (05-15-2017)
Keep in mind that the mean isn't usually appropriate with ordinal data like this. The true, theoretical Likert items were supposed to be interval level data, I believe, but most "Likert scales" that people create don't actually fit the criteria for interval data, meaning that they are at most ordinal data, and only the median and mode are applicable, but not typically the mean.
MvZ (05-15-2017)
You should use Median Test. Please check on Conover : Nonparametric Statistics.
Good Luck
MvZ (05-15-2017)
I disagree with the blanket statement that mean is not appropriate when considering Likert scale responses. That's not to say that parametric tests are appropriate, but I think a quick mean can be helpful when thinking about what Likert data are telling you.
I think one should always first ask "what am I trying to learn from the data" and THEN ask "what tests can I run?"
GretaGarbo (05-12-2017), MvZ (05-15-2017)
Thank you so much for your responses. I was aware of the fact that I will be using 'simple' descriptive statistics such as median and mean.
However, the Friedman test seems as a plausible method, doesn't it?
This is a classic example of Friedman:
'n wine judges each rate k different wines. Are any of the k wines ranked consistently higher or lower than the others?'
I want to know if n students rank several aspects consistently higher or lower than others, allowing me to make a ranking order. Will the Friedman test do this?
In addition, I want to compare groups. So I want to be able to split the ranking between for example 6th years students and 1st years students?
Furthermore, I would like to know if for example experience would influence the result of the ranking and how much its influence is?
Could you guys help me out with this as well, because that would be so helpful.
Thank you in advance!
I swore I posted this a few days ago, but it was either deleted, or I was mistaken.
I also disagree with the blanket statement, but if you read my post, you'll notice I said the mean isn't usually appropriate. There is a fair bit of consensus among statisticians that the mean is not appropriate for these kinds of scales because they're not actually interval data. The true Likert item would be interval in nature, but most often what people call "Likert" does not fulfill the properties of interval data, so the mean is less appropriate but the median and mode are better to look at.
MvZ (05-15-2017)
Ondansetron -- Agreed!
MvZ (05-15-2017)
MvZ -- I do not think a Friedman test is appropriate to your data. Friedman is for repeated measures. If you are going to ask the same people the same questions again on a different issue, then maybe Friedman is appropriate. But if you're only asking each person each question once, then you don't meet the test's assumptions. In that case, I would suggest a chi square test instead.
MvZ (05-15-2017)
Thank you everyone for helping me out with this issue.
I think I know what I have to do now.
Some questions use a Likert-format (SD to SA), but are Likert-items that are not combined. For these questions, Likert is treated as ordinal data and likely non-normally distributed and as such median and mode are the best descriptive statistics. In addition, I will use the Mann-Whitney U test to analyze differences in medians between the groups (6th years vs. 1st years).
Some other questions are used as Likert-scale and combined they measure a single latent variable. As such, this data will be treated as interval data and likely non-normally distributed and the median or mode will be the best descriptive for this variable. In addition, I will use a Kruskal-Wallis test and/or a Jonckheere-Terpstra test to analyze differences in the medians between the groups.
If you still have any comments or feedback, I would be glad to hear.
I would find it OK to calculate the mean for each variable (question) and compare them from highest to lowest.
Many suggest to use the median, but that can be quite boring for data like this since the median will often be just the middle value. Boring and not very informative. That is because there are so many ties (the same value). It would have been OK if the scale had been a continuous scale so that there could have been an infinite number of values. Here there are only 7 possible values.
I can't remember anyone say in the statistical education that it is not "allowed" to use means and I don't know of any statistician that has published that in a statistical journal. It would be interesting to hear. It is that from Stevens but that seems controversial.
An other possibility would be to plot the (accumulated) empirical distribution function. (If the data are approximately normal it would look S-shaped.) Variables with many high values would be far out to the right, and low values to the left. If you are interested in the median you can draw a horizontal line from 0.5 and look which variables are high and which are low.
But who has said that it would be the centrality parameter (mean median of mode) that would be of primary interest? Maybe it is where the most satisfied (highest scoring) are (for example the 80 percent percentile) or what variabels have the most angry and dissatisfied people (say, the 20 percent percentile). Statistics has nothing to say about what is the most relevant to this researcher. It is her own thing to decide.
I can't understand when some people say that it is not "allowed" to use the mean for a single Lickert item. But that they say that if you sum (say) ten items into a Lickert scale, then it would be OK to use the mean. I don't understand that.
Tweet |