two-way repeated measures MANOVA; variation on Friedman test

[Esther]

I am working on an analysis with the following characteristics:
• There is one independent variable, group (control v. treatment, about 15
respondents in each group)
• There are seven dependent variables, each measured at 3 different points
in time (T1, T2, T3)
• Of the dependent variables (each measured, as I said, at 3 points in time,
though there were a few dropouts):
 One (KJ1-KJ3) is a single, normally distributed variable, which can be
studied by two-way repeated measures ANOVA
 Two (LT1-LT3 and SA1-SA3) are normally distributed, and go together, I'd
like to use a two-way (doubly multivariate?) repeated measures MANOVA, but I
don't know how to do so in SPSS.
 Four (VAS11-VAS13, VAS21-VAS23, VAS31-VAS33, VAS41-VAS43) are non-normally
distributed, I'd like to study each one separately using a non-parametric
version of a two-way repeated measures ANOVA.  Am looking for such a test.
Friedman does not work for multiple groups, and most of the others I've seen
don't work for repeated measures.
• Although I'm a bit interested in the change over time, the main point of
the analysis is to compare control v. treatment groups and, if relevant, the
interaction between group and the various dependent variables.
The below file is a layout of what the relevant data look like (actual
values, except of group, left out):

AnalysisQ1table.xlsx
<http://spssx-discussion.1045642.n5.nabble.com/file/t341677/AnalysisQ1table.xlsx> 



--
Sent from: http://spssx-discussion.1045642.n5.nabble.com/

=====================
To manage your subscription to SPSSX-L, send a message to
[e-mail] (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

[Rich Ulrich]

N=30 is not very many to support a multivariate analysis,

regression or whatever, unless there is are "large" effects.

Aren't some of these more important, more central, than

the others? "Too many tests" is a relevant warning.

Therefore: I think you ought to think a lot about reducing

the number of "main hypotheses" (one per variable) or

figure how to combine them by creating composite variables.

If you have a predicted difference between groups, across

time, for some subset of the variables, the simple composite

would be an average across time for each variable, followed

by a weighted average of, say, the z-scores of those scores,

so that each of the several get equal weight.

As to the "non-normal" variables: Taking the rank-transforms

for doing a "non-parametric analysis" is, frankly, taking

transformations that run into trouble in multi-variable

situations. Where do those variables come from? What do

they measure? - Very often, there is power transformation

(taking the log or reciprocal) that will create near-normality

much better than a rank-transformation does, and will allow

you to use the transformed scores in the regular ANOVAs.

Also, in the interest of reducing multiplicity of tests, and in

controlling which degrees of freedom are used up by the

analyses: Do you expect linear change?  change mostly from

Period 1 to 2?  You can create one or several variables which

directly measure the change, like taking the linear component,

or taking (End-First).  You might do better to analyze with dummy

variables to represent the time effects, rather than work with

defining a limited number of interactions in the ANOVA.

--

Rich Ulrich

---

From: SPSSX(r) Discussion <[hidden email]> on behalf of Esther <[hidden email]>
Sent: Tuesday, October 29, 2019 7:26 PM
To: [hidden email] <[hidden email]>
Subject: MANOVA

 

I am working on an analysis with the following characteristics:
•       There is one independent variable, group (control v. treatment, about 15
respondents in each group)
•       There are seven dependent variables, each measured at 3 different points
in time (T1, T2, T3)
•       Of the dependent variables (each measured, as I said, at 3 points in time,
though there were a few dropouts):
       One (KJ1-KJ3) is a single, normally distributed variable, which can be
studied by two-way repeated measures ANOVA
       Two (LT1-LT3 and SA1-SA3) are normally distributed, and go together, I'd
like to use a two-way (doubly multivariate?) repeated measures MANOVA, but I
don't know how to do so in SPSS.
       Four (VAS11-VAS13, VAS21-VAS23, VAS31-VAS33, VAS41-VAS43) are non-normally
distributed, I'd like to study each one separately using a non-parametric
version of a two-way repeated measures ANOVA.  Am looking for such a test.
Friedman does not work for multiple groups, and most of the others I've seen
don't work for repeated measures.
•       Although I'm a bit interested in the change over time, the main point of
the analysis is to compare control v. treatment groups and, if relevant, the
interaction between group and the various dependent variables.
The below file is a layout of what the relevant data look like (actual
values, except of group, left out):

AnalysisQ1table.xlsx
<http://spssx-discussion.1045642.n5.nabble.com/file/t341677/AnalysisQ1table.xlsx> 



--
Sent from: http://spssx-discussion.1045642.n5.nabble.com/

=====================
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

[Art Kendall]

Does it make sense to create  one or two summative score for your DV's?

It would help in suggesting approaches to analysis if you provided the
meanings of your variables and the values they can take. (I.e., the variable
labels, value labels, and missing value labels.)

In very many instances whether the variables themselves are non-normally
distributed is not of as much as whether is is plausible to assume that
residuals from the model are not severely non-normal.

What specifically are your concerns about normality?



-----
Art Kendall
Social Research Consultants
--
Sent from: http://spssx-discussion.1045642.n5.nabble.com/

=====================
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