Re: best test for this analyses?

Posted by Rich Ulrich on Mar 20, 2014; 1:32am
URL: http://spssx-discussion.165.s1.nabble.com/best-test-for-this-analyses-tp5724913p5724941.html

"99 participants" says that there are 3 sets of 33.

First: Do you need to abandon ordinary ANOVA?

Bruce Weaver provided the proper quotation after I botched the
reference a few years ago -
    The version of that quote I know is another George Box classic.
  "To make the preliminary test on variances is rather like
  putting to sea in a rowing boat to find out whether
  conditions are sufficiently calm for an ocean liner to
  leave port!"   Box G. E. P. (1953) Non-normality and tests on variances.
  Biometrika 40, 318�35. "

Second:  The weaker assumption for a paired t-test, which would test
whether 99 scorers are higher on biology than English, is that the
*difference* is normal.  If you set the design up as ANOVA, the
assumption is that the residuals are normal. 

If you actually do have opposite skew in the two tests... that seems
to be a rather unfortunate indication that, whatever assurances you
may have received, the "matching" of the tests was done "by sample",
not by scoring at different levels; and (I'm pretty sure) the match would
be reliable only for the particular level of skill of the matching sample.

I, therefore, might want to decide that the two scales are not parallel
enough to analyze together, except for by overall level.  After the paired-t
comparison, I would look at the separate linear trends for English and bio.

--
Rich Ulrich

> Date: Sat, 15 Mar 2014 10:14:58 -0400

> From: [hidden email]
> Subject: best test for this analyses?
> To: [hidden email]
>
> I would be greatful for advice on analysing this data in SPSS
>
> A study was done of comparing test results (% correct responses) on two
> subjects biology and english. The tests are supposedly adjusted so as to be
> of equal difficulty (approx).
>
> The study was done in 3 years of a school at the same time.
>
> So we have 33 subjects in year 1 tested on biology and english (the same
> participants were tested twice, once in each subject - i.e. repeated
> measures) from year 1, same for year 2, same for year 3 (total 99
> participants)
>
> SO the research question is whether there are differences on the basis of
> subject (i.e. do scores on biology differ significantly from those on
> english) and whether performance improves over the years (i.e. is year 2
> better than year 1, is year 3 better than year 2).
>
>
> I was thinking ANOVA with multiple comparison tests. Problem is that
> normality tests show not normal distribution (KS And SW)
>
> so then thinking kruskall wallis but the assumption of the variables having
> the same shaped distribution is not met (one has a positive skew the other
> negative).
>
> I ran the kruskall wallis test anyway and it showed significant differences
> by year for one of the subjects (english)but not for the other (biology).
>
> Even of I could use KW this still leaves the issue of how to do multiple
> comparison test with non parametric data, to find out which years differ
> from which. Reading suggests this is obscure process with non parametric
> data?
>
> Then there is the other research question - hoe to test for significant
> differences on basis of subject?
>
> Think I need to give this more thought bit any guidance much appreciated
>

...