Analysis Advice - Self-Esteem Issue Updated

[zstatman]

Have 4 elements or levels/variables that represent "self esteem," 2 levels
of race and a scale variable for body weight. My objective is to determine
if there is a difference in esteem by race as body weight increases.

Bit more clarity ...

The question is: "Is a difference in esteem by race as body weight
increases"

Esteem is a set of 4 scale variables created from various Likert scale
questions, race is a 3-level nominal & weight is scale but also have it
categorized into a 4-level nominal

I am looking into a Repeated Measures ANOVA setup like

GLM
body_cat wt_cat phy_cat oth_cat WITH race
/WSFACTOR = Esteem 4 Simple(1)
/METHOD = SSTYPE(3)
/PRINT = DESCRIPTIVE
/CRITERIA = ALPHA(.05)
/WSDESIGN = Esteem
/DESIGN = race .


Tks
W

[Maguin, Eugene]

Will,

I have to take a break from what I'm working on.

I think others are better qualified but here is my two cents worth of
comments. If I understand you correctly you are looking for an interaction
between race and weight across your four esteem measures. I would treat
weight as a continuous covariate rather than as a categorical main effect on
the theory that categorizing weight decreases it's predictive ability.
However, in doing so I assume that the relationship between esteem and
weight is linear within each race category. But that can be checked ahead of
time as a preliminary. My main concern would be in using a repeated measures
structure rather than a multivariate structure because a repeated measures
structure imposes specific assumptions about covariance matrix structure in
addition to between group homogeneity. That's something to investigate. I'd
think that a more general method for analyzing this would be to use GLM
multivariate anova (i.e., manova).

I think the ideal method for working this problem would actually be an
structural equation model, which could be done if you have Amos because you
can test assumptions in a structured way, a way that I am not sure that you
can do even in the Mixed procedure.

Gene Maguin

[zstatman]

Thanks Gene, good points

W

-----Original Message-----
From: Gene Maguin [mailto:[hidden email]]
Sent: Friday, March 30, 2007 3:31 PM
To: 'Will Bailey [Statman]'; [hidden email]
Subject: RE: Analysis Advice - Self-Esteem Issue Updated

Will,

I have to take a break from what I'm working on.

I think others are better qualified but here is my two cents worth of
comments. If I understand you correctly you are looking for an interaction
between race and weight across your four esteem measures. I would treat
weight as a continuous covariate rather than as a categorical main effect on
the theory that categorizing weight decreases it's predictive ability.
However, in doing so I assume that the relationship between esteem and
weight is linear within each race category. But that can be checked ahead of
time as a preliminary. My main concern would be in using a repeated measures
structure rather than a multivariate structure because a repeated measures
structure imposes specific assumptions about covariance matrix structure in
addition to between group homogeneity. That's something to investigate. I'd
think that a more general method for analyzing this would be to use GLM
multivariate anova (i.e., manova).

I think the ideal method for working this problem would actually be an
structural equation model, which could be done if you have Amos because you
can test assumptions in a structured way, a way that I am not sure that you
can do even in the Mixed procedure.

Gene Maguin

[Richard Ristow]

At 03:31 PM 3/30/2007, Gene Maguin wrote:

>Here is my two cents worth of comments.

This is a one-cent addition - the budget is tight.

>I assume that the relationship between esteem and weight is linear
>within each race category.

That's a very tenuous assumption. This is a human subjective response,
and subjective responses to continuous quantities are most often
non-linear.

Most often there's a 'roll-off' - subjective response increases less
than in proportion to the increase in the quantity.

Two choices:

. Taking the response as logarithmic often works well. I believe that
has good experimental verification in some cases where it's
quantifiable, such as response to the intensity of a light stimulus;
and it's probably about right for the subjective effects of wealth and
income. So, it's a candidate to try.

. If you have enough data (as you may), and you don't want to defend
the logarithmic assumption, try just adding a quadratic term in weight.
The dynamic range in weights - ratio between largest and smallest
commonly observed values - may be too small to distinguish a quadratic
from an logarithmic response, anyway. Follow the usual precautions for
inserting a quadratic term in a model. Notably, control the correlation
with the linear term; this may be done by centering the quadratic term
around somewhere near the midpoint of the observed range. (You can also
use a formally orthogonal polynomial, but that's probably more trouble
than it's worth.)

If you have gender for your subjects, you can pretty well count on
their being a weight-by-gender effect. That's not for any subtle
psychological reason, but because (probably) the perception is not of
weight as such, but of weight relative to 'normal'; and,
physiologically, the 'normal' weight ranges are different for men and
women.

-Good luck,
  Richard

[Linda Case]

Greetings -

I have a set of data that include multiple measures on each of 15 subjects.
The study is 35 days long and on each day, subjects visited a food site as
many times as they liked. Each visit resulted in generation of a "time
stamp".  So, when these data are formatted in the long format, each subject
has a varying number of repeated measures, depending on how many visits
occurred in a given day.  I took the time stamp variable and converted it to
two variables, one for date and the second for time on that date.  In
addition, I would like to create a new variable that counts the number of
visits per day, assigning a sequential number; 1, 2, 3, etc. So, on each new
date, the subject's number would start again at 1 with the first visit to
the site.  Does anyone have any thoughts about how to do this (or syntax
that would be of help?)

Thanks very much in advance!

Linda Case

Linda P. Case
AutumnGold Consulting
[phone]
www.autumngoldconsulting.com
[hidden email] or [hidden email]

[Hal 9000]

I have a pretty good idea how to approach this problem (aggregate
/subcommand) but if you posted sample data (real sanitized data is
preferable) + the outcome you desire, you'll get more usable answers.
-Gary


On 3/30/07, Linda Case <[hidden email]> wrote:

>
> Greetings -
>
> I have a set of data that include multiple measures on each of 15
> subjects.
> The study is 35 days long and on each day, subjects visited a food site as
> many times as they liked. Each visit resulted in generation of a "time
> stamp".  So, when these data are formatted in the long format, each
> subject
> has a varying number of repeated measures, depending on how many visits
> occurred in a given day.  I took the time stamp variable and converted it
> to
> two variables, one for date and the second for time on that date.  In
> addition, I would like to create a new variable that counts the number of
> visits per day, assigning a sequential number; 1, 2, 3, etc. So, on each
> new
> date, the subject's number would start again at 1 with the first visit to
> the site.  Does anyone have any thoughts about how to do this (or syntax
> that would be of help?)
>
> Thanks very much in advance!
>
> Linda Case
>
> Linda P. Case
> AutumnGold Consulting
> [phone]
> www.autumngoldconsulting.com
> [hidden email] or [hidden email]
>

[Richard Ristow]

At 04:52 PM 3/30/2007, Linda Case wrote:

>I have a set of data [in which] on each day, subjects visited a food
>site as many times as they liked. Each visit [received a new record
>and] a "time stamp".  I took the time stamp variable and converted it
>to two variables, one for date and the second for time on that
>date.  In addition, I would like to create a new variable that counts
>the number of visits per day, assigning a sequential number; 1, 2, 3,
>etc.

Assuming that it is 'long', i.e. one record per visit; that your data
includes variables Subject, VistDate, and VistTime as you describe
them; and that VistDate is an SPSS date variable, and is a true date
(time portion is zero); then, something like this, though untested.

It assumes that those variables are never missing in the data.

NUMERIC   VistNumb (F3).
VAR LABEL VistNumb 'Visit sequential number, within day'.
LEAVE     VistNumb.

SORT CASES BY Subject VistDate VistTime.

DO IF   MISSING(LAG(Subject)).
.  COMPUTE #NewDay = 1.
ELSE IF MISSING(LAG(VistDate)).
.  COMPUTE #NewDay = 1.
ELSE IF Subject  NE LAG(Subject).
.  COMPUTE #NewDay = 1.
ELSE IF VistDate NE LAG(VistDate).
.  COMPUTE #NewDay = 1.
ELSE.
.  COMPUTE #NewDay = 0.
END IF.

IF #NewDay VistNumb = 0.
COMPUTE    VistNumb = VistNumb + 1.
...................................
That "DO IF" construct is to identify a subject's first visit on a day.
Here's a simpler way, with the drawback that "@NewDay" is a normal
variable and remains in (cluttering) the file.

NUMERIC   VistNumb (F3).
VAR LABEL VistNumb 'Visit sequential number, within day'.
LEAVE     VistNumb.

SORT CASES BY Subject VistDate VistTime.

ADD FILES
    /FILE=*
    /BY Subject VistDate
    /FIRST=@NewDay.

IF @NewDay VistNumb = 0.
COMPUTE    VistNumb = VistNumb + 1.

[Dennis Deck]

While you could create a sequence number, the more direct way is to use
AGGREGATE to create a record for each date and compute the count for
that day.

But note that you must first create a new date that drops the time
portion of the original datetime variable else you will not get the
desired result:

COMPUTE NewDate = XDATE.DATE(OldDate) .

Dennis Deck, PhD
RMC Research Corporation
[hidden email]


On 3/30/07, Linda Case <[hidden email]> wrote:
>
> Greetings -
>
> I have a set of data that include multiple measures on each of 15
> subjects.
> The study is 35 days long and on each day, subjects visited a food
site as
> many times as they liked. Each visit resulted in generation of a "time
> stamp".  So, when these data are formatted in the long format, each
> subject
> has a varying number of repeated measures, depending on how many
visits
> occurred in a given day.  I took the time stamp variable and converted
it
> to
> two variables, one for date and the second for time on that date.  In
> addition, I would like to create a new variable that counts the number
of
> visits per day, assigning a sequential number; 1, 2, 3, etc. So, on
each
> new
> date, the subject's number would start again at 1 with the first visit
to
> the site.  Does anyone have any thoughts about how to do this (or
syntax

> that would be of help?)
>
> Thanks very much in advance!
>
> Linda Case
>
> Linda P. Case
> AutumnGold Consulting
> [phone]
> www.autumngoldconsulting.com
> [hidden email] or [hidden email]
>