Testing for Normal Distributed Data

[Krupa Sheth]

I am running the analysis on a sample size of 32 of three groups for a test-retest reliability. I first want to find out whether the data is normally distributed.

I have two ideas – one is using through Analyze>Explore but got the tests as being all significant (ie data not being normally distributed). I looked at mainly the Shapiro-Wilk tests as the sample size Non Parametric Tests>1-sample KS test I got varied results.

The issue is that first method using the explore function I understand uses an additional correction which is more conservative and explains why I am more likely to have significant results. Is there any way to use a less conservative approach for my groups??

Thanks!!

[Anonymous]

the distirbutional assumptions of most analyses are not necessarily on the data itself but on conditional distribution of the data given the model/technique you are using (i.e. on the residuals)…

 

… so what exactly are you trying to do? it would help us out a lot if you told us about if you told us about your data and what is it that you’re doing…

[Krupa Sheth]

My data basically contains three groups of disability who were given the same assessment twice within a maximum of 7 months between the first and second assessment. I am trying to establish reliability on the assessment and using the three groups disability groups. The main aim is to if the assessment has test-retest reliability.

I need to know whether my data is normally distributed before running any analysis. The reliability analysis I am trying to run would either be correlations or intraclass coefficients.So far looking at the Kolomogrov-Smiroff test has given all significant values despite the data being heavily skewed. The other option would be using P-P or Q plots.

Hope this helps.

[Anonymous]

i see… well, for the issue of correlations you can also think about the reliability of your instrument to see whether the scale that you use ranks them the same way (or similarly) at fime 1 and time 2. in such case you can use Spearman’s rho which is the non-parametric/more robust version of the Pearson correlation coefficient.

for the intraclass correlation you’ll need to think a little bit more. in their classic paper, Shrout & Fleiss (1979) show that there are i think 3 or 4 different kinds of intra-class correlations. i think people have come up with more and there’re about 6 or 7 of them. they’re all based on ANOVA-like variance decompositions depending on what you choose to be a fixed or a random factor. in this case the situation is a little bit easier because, once again, the distirbutional assumptions of ANOVA are also on the residuals and not on the variables themselves so you can test the residuals of the ANOVA you’re supposed to use and see whether they are normally distributed or  not.

[Author]

Posts