Which test to use?

[Helena]

Hello everyone!

I need your help. I have conducted a study with three intervention and one control group. Now, I want to analyze the (ordinal scale) data for pre-post effects.
What do you think, which test would be appropriate here?
I have thought about the Friedman test since there are more than two dependent groups and the data is ordinal. However, I am not sure how exactly to conduct the Friedman test in this case since I have three intervention and 1 control group. So not all the data is paired, it's just the pre-post results.
Should I conduct a Wilcoxon matched pair sign test for every group separately and would i need to perform for example a bonferroni correction in this case?
I am interested in your thoughts about this.

Thank you

 

[Karabiner]

I have thought about the Friedman test since there are more than two dependent groups and the data is ordinal.

What you described were 4 independent groups and 2 dependent measurements (pre vs. post).
So, yes, Friedman is not useful here.

Exactely how did you measure your dependent variable?

Should I conduct a Wilcoxon matched pair sign test for every group separately and would i need to perform for example a bonferroni correction in this case?

No. You are interested in the comparison between groups with regard to their change over
time. That has to be done directly, not by means of 4 separate analyses.

BTW, the Wilcoxon signed rank test requires interval scaled data. For paired ordinal data,
the sign test can be used.

You did not describe the topic of your research, your research question(s), whether
allocation of subjects to groups was carried out by randomization, the sample sizes,
or how you measured your variables. Therefore it is difficult to determine what to do.
Maybe you could perform a ordinal logistic regression with the baseline value as
covariate, or an analysis using generalized estimating equation (GEE), or maybe
something else.

With kind regards

Karabiner

 

[Helena]

Hello!

Thank you so much for your answer. I will try and clarify my research in this post. Oh an why does the Wilcoxon matched pair signed rank test require interval data. I have learned that it is a non-parametric test that requires at least ordinal data.

Concerning my research... I want to find out if there are differences in the intention to perfom physicial activity, the attitude towards it and the perceived behavioural control (theory of planned behaviour) after randomly being chosen to see on of four advertisements. I have three "interventional" advertisements and one neutrally formulated advertisement and I would like to see if there are differences in the effectiveness of the advertisements.
I have data about participants' intention, attitude towards physical activity, perceived behavioural control before seeing the advertisements and after seeing one of the advertisements.

I measured every variable with multiple items using a 7-point Likert scale. However, I cannot sum up the scores since the data are ordinal but I don't know if it is wise to test for differences for every single item.

I have a sample size of n=164.

I hope that helps and that someone of you will be able to help me.

Btw, I have also thought about conducting an ordinal regression analysis.

Kind regards,
Helena

 

[Karabiner]

why does the Wilcoxon matched pair signed rank test require interval data. I have learned that it is a non-parametric test that requires at least ordinal data.

It is a nonparametric test which requires intervals scaled data. The basic idea of
the test is to compare the magnitude of negative differences with positive differences,
which does not make sense with ordinal data.

I measured every variable with multiple items using a 7-point Likert scale.
However, I cannot sum up the scores since the data are ordinal but I don't know if it is wise to test for differences for every single item.

"Likert scale" is not the name of the response format of single Likert-type items,
but the name of an instrument which consists of several Likert-type items,
which are summed up. If you have validated Likert scales, consisting of
several items measuring the same construct, then it would be common
to sum up the item scores and consider the result as interval scaled. In
that case, a mixed analysis of variance with 1 repeated-measures factor
(pre-post) and 1 repeated-measures factor (4 groups) would be an option.

With kind regards

Karabiner

 

[Lukan27]

One thing is to use a test, but you can always show some descriptive statistics about the matter. It's pretty easy to juxtapose some histograms or whatever for the groups, and let the readers judge. Sadly it's not that common in academia to do this.

 

[Luca P.]

It is a nonparametric test which requires intervals scaled data. The basic idea of
the test is to compare the magnitude of negative differences with positive differences,
which does not make sense with ordinal data.


"Likert scale" is not the name of the response format of single Likert-type items,
but the name of an instrument which consists of several Likert-type items,
which are summed up. If you have validated Likert scales, consisting of
several items measuring the same construct, then it would be common
to sum up the item scores and consider the result as interval scaled. In
that case, a mixed analysis of variance with 1 repeated-measures factor
(pre-post) and 1 repeated-measures factor (4 groups) would be an option.

With kind regards

Karabiner

Thank you so much!

I could not find any validated Likert-scales, so I mostly used items that were used in previous studies. However, I do not assume them to be interval scale. Unfortunately, there are not that many statistical methods that I know of that can analyse several dependent ordinal variables all at once. Does that mean that I would need to analyse every item on its own?

Kind regards

 

[Luca P.]

One thing is to use a test, but you can always show some descriptive statistics about the matter. It's pretty easy to juxtapose some histograms or whatever for the groups, and let the readers judge. Sadly it's not that common in academia to do this.

Thank you! I will definitely include some descriptive statistics.

 

[Karabiner]

Does that mean that I would need to analyse every item on its own?

So, Helena has turned into Luca P.?

The description of your measurements and of your study is insufficient, so one would have to guess.
But it sounds as if it doesn't make much sense. You should find a way to aggregate those of your
measurements which are assumed to measure the same construct.

With kind regards

Karabiner

 

[Luca P.]

Thank you Karabiner!

Unfortunately, I cannot aggregate, i.e sum up the measurements since they are ordinal scale and sum scores are not appropriate here...

 

[Karabiner]

Of course one can aggregate ordinal data...

 

[Luca P.]

Oooooh I think I got it now. You mean I could calculate the median of all the variables measuring the same construct, right? I was so focused on building sum scores that I completely forgot that I do not need to define the median per item but for all the items measuring one variable.

Thank you