T-test vs nonparametric for likert scale (very small sample size)


New Member
Hi everyone,
I'm in the process of completing my masters thesis, and while I did my share of psych stats in my undergrad it's been YEARS and I'm a lil rusty! Any advice or help anyone could provide would be greatly appreciated, and you may even get a mention in my thesis acknowledgement section :)

The study:
Mixed method, feasibility design study that is looking at comparing modes of music therapy delivery (two groups: traditional music instruments versus "app" instruments on portable technology) at reducing social isolation and improving self-esteem for older adults. The traditional instrument group (TMI) has n = 3, and the portable tech (PED) has an n = 2.

IV: mode of music therapy delivery (2 levels - TMI and PED)
DV: Social isolation; self-esteem (both self-rated; measured at TI & T2)

The measures:
Participants in both conditions completed pre- and post- measures of isolation and self-esteem. These are likert scales with several items that are summed to equal a total score.
I want to see if there are any differences between the two groups (TMI versus PED at both pre and post), and also whether there are any differences for each group at pre versus post (e.g TMI-T1 versus TMI-T2 etc).

I'm am expecting that there will not be any differences between TMI and PED (so portable technology is as effective as traditional instruments); and that scores of self esteem and social isolation will improve from T1 to T2 for both groups (although honestly this is pretty unlikely given the very short program duration etc).

My question:
What is the best way to compare my groups in SPSS? I know likert items typically violate the normality assumptions of t-tests (and the population itself has a skewed distribution for both isolation and self-esteem, with the majority of about 40% having high levels of support and self-esteem). I've attempted to run t-tests, kruskal Wallis, Mann Whitney U etc but I really need some guidance as to what the best method of analysis would be!

Thanks so much for taking the time to read it all, its a bit embarrassing how quickly I have forgotten it all, so be gentle talkstats peeps!



Less is more. Stay pure. Stay poor.
Wilcoxon would be the better option, but your very small sample size will put null conclusions into question. I believe their may be an exact form of the wilcoxon, if so use that approach.