# Which statistical test for these data ?

#### Balanin

##### New Member
Hi, I'm new and french, so excuse my bad english, I'll do my best anyway.

Here is my problem, I have three different kind of data and I need to test a time or a product effect (factor with only 2 modalities : 0 and 1).

1) I have a questionnaire with a Lickert scale, can I transform my data with a numeric and ordinal scale (like Agree = 4, Disagree = 0...) ? If I do this, is there a sense to test the normality of distributions in order to use a student test for example. For me these data are like ranks because it's ordinal and not really numeric, so I think a Wilcoxon/Mann-Whitney test is the best thing to do here. Am I right ?

2) I have another questionnaire, but this time it's a only an ordinal scale, the items are numeric only and they translate a feeling (from 0 to 9), there is no link between an item and a sentence like in a Lickert scale. The only thing we know is : 0 -> the worth and 9 -> the best. These data are scores, but because they are numeric, in that case I think I can check the normality of distribution in order to use a student t test, are you ok with that ?

3) The last questionnaire is in fact a large horizontal scale to translate, here again, a feeling, a subject thicks a cross on the horizontal scale, then we take a measurement of the distance from the origin and the cross (in centimeters). In that case I also think I have to check normality of distributions and use if possible a parametric test. Do you agree with that ?

Thank you all for reading a desperate french guy, looking for a solution to his rookie question..

Have a nice day!

Balanin

#### Dr.D

##### New Member
Hi, I'm new and french, so excuse my bad english, I'll do my best anyway.

Here is my problem, I have three different kind of data and I need to test a time or a product effect (factor with only 2 modalities : 0 and 1).

1) I have a questionnaire with a Lickert scale, can I transform my data with a numeric and ordinal scale (like Agree = 4, Disagree = 0...) ? If I do this, is there a sense to test the normality of distributions in order to use a student test for example. For me these data are like ranks because it's ordinal and not really numeric, so I think a Wilcoxon/Mann-Whitney test is the best thing to do here. Am I right ?

2) I have another questionnaire, but this time it's a only an ordinal scale, the items are numeric only and they translate a feeling (from 0 to 9), there is no link between an item and a sentence like in a Lickert scale. The only thing we know is : 0 -> the worth and 9 -> the best. These data are scores, but because they are numeric, in that case I think I can check the normality of distribution in order to use a student t test, are you ok with that ?

3) The last questionnaire is in fact a large horizontal scale to translate, here again, a feeling, a subject thicks a cross on the horizontal scale, then we take a measurement of the distance from the origin and the cross (in centimeters). In that case I also think I have to check normality of distributions and use if possible a parametric test. Do you agree with that ?

Thank you all for reading a desperate french guy, looking for a solution to his rookie question..

Have a nice day!

Balanin
Some people have argued that Likert scales should be treated as ordinal (hence nonparametric tests like Mann-Whitney can be used) or they should treated as interval/continuous (whereas student t-test can be used). What I recommend, as you would agree with, is to do normality tests for all scale scores to determine whether data are normally distributed, and apply the test to suit for each of the three sections on the questionnaire. The normality tests would give you better evidence to make such a judgment which may otherwise be made subjectively or based on personal preference.

#### Balanin

##### New Member
Thank you for your quick answer, I will thus check the normality of my data from now! :tup:

#### PPDoc

##### New Member
Ordinal data

Your data are ordinal and should be analyzed as such. Using rank-based methods, such as Spearman's correlation coefficient or the Mann-Whitney test are good.

Alternatively, you may want to consider ordinal logistic regression. The advantage of a regression approach is that you can look for combinations of predictors of your outcome (i.e. you can do a multivariate analysis). There are several types of ordinal regression and several of these are implemented in R.

PPDoc