Adjusted significance for multiple comparisons following Friedman test

[Anton-24]

Hello~

Following a Friedman test, tests for multiple comparisons were requested.  The output includes two columns labelled "Sig." and "Adj. Sig.".  Am I correct to assume that "Sig." is simply the unadjusted significance and "Adj. Sig." is the adjusted significance using the Dunn-Bonferroni method?

The scale I'm using is ordinal, but has only three values (1, 2 & 3), so there a many tied scores.  I understand that tied scores are transformed into average ranks, but I haven't found any source that warns against using Friedman's test with a limited range scale / many tied ranks.  Any thoughts?

Thanks.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Art Kendall]

Please describe your data more fully.

What questions are you using the data to answer?

Do you have the variable(s) that was/were coarsened into 3 groups of cases?

[Anton-24]

Thanks for your interest, Art.

It's a fairly simple design.  Respondents rated themselves using a three-point ordinal scale (High, Medium, Low) at three different time periods.  The scale is what it is;  There are not less-coarsened values.

The researcher is interested in assessing change between time periods.

I did confirm that the post-hoc tests following the Friedman test indeed use Dunn's (1964) procedure.

I have read some concerns about having too many tied ranks with some non-parametric tests, but wasn't sure if this is an issue with Friedman's test.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Art Kendall]

I do not have my files and books where I am right now, but IIRC, Friedman is for use on ranks with few ties, i.e., when there are (almost) as many values of the variables as there are cases.

At last in my mind there is a big difference between ordinal variables that are ranks (i.e., there are (almost) as many values as there are cases) and variables that have only a few values that can be ranked, but that is not the same as ranking cases.

It is possible that you want to use the CATEGORIES module, and see if CATREG gives you substantively different results if you treat the variables, as ordinal or whether the intervals between values are not very very discrepant from being equal.  Some form of multiple correspondence analysis may be what you are looking for.

Do you have only 1 group of respondents/patients/subjects?

Perhaps others on the list would like to suggest ways to deal with 3 repeated ordinal variables with 3 levels.

[Rich Ulrich]

WJ Conover showed that, in general, that the commonly-used adjustments for ties
for computing error terms for large samples in various nonparametric tests do not
work very accurately.  In fact, using a rank-transform followed by the ordinary ANOVA
often gave more accurate p-values in his simulations.  That was a side-result in his classic
discussion of the nature of rank-order testing.  Here is a link to the abstract -

http://amstat.tandfonline.com/doi/abs/10.1080/00031305.1981.10479327#.U_d232MfgXQ

The typical "nonparametric test" is almost identical to doing the rank-transformation,
followed by an ordinary ANOVA.  This raises the proper question:  Is ranking always a reasonable
transformation (especially when assumptions - including "no ties" - are not met)?  Does ranking
result in a better or worse fit to what would be the face-valid notion of "equal intervals"?  or does the
original scoring (such as, the values of 1,2,3) make as much (or better) sense?

--
Rich Ulrich

> Date: Fri, 22 Aug 2014 11:39:59 -0400

> From: [hidden email]
> Subject: Re: Adjusted significance for multiple comparisons following Friedman test
> To: [hidden email]
>
> Thanks for your interest, Art.
>
> It's a fairly simple design. Respondents rated themselves using a three-point ordinal scale (High, Medium, Low) at three different time periods. The scale is what it is; There are not less-coarsened values.
>
> The researcher is interested in assessing change between time periods.
>
> I did confirm that the post-hoc tests following the Friedman test indeed use Dunn's (1964) procedure.
>
> I have read some concerns about having too many tied ranks with some non-parametric tests, but wasn't sure if this is an issue with Friedman's test.
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD