Sample size question

[Byrd Sellers, Johnita]

Hello All,

 

I have a quick question about sample size calculations.  I’m currently working on a study that deals with couples therapy.  There will be several couples assigned to two different groups.  I would like to compare the two groups in hopes that the intervention group is better.  I wanted to know how best to perform a sample size for couples.  In this study, I will get feedback from both persons.  Since this is the case, I’m thinking that I would just include each person as an individual and perform the calculation as normal.  But I run into the problem of “paired” data within each group.  This is where I get confused.  Any assistance that you can give regarding this matter is greatly appreciated.

 

Thank you in advance,

Johnita

[Maguin, Eugene]

Johnita,

The standard frame for this type of question is what sample size do I need
to obtain a power of P (typically, .80 in social science research) given an
effect size of ES from this analysis method and an alpha level of A
(typically, .05). Intertwined in the sample size question is an analysis
question because  different analysis methods yield different powers given
the same data.

So it seems like your design is a two group. Although you don't say anything
about number of assessments, let's assume pre and post. Ignoring 'couple'
and treating each person as an individual means that you are assuming that
wives and husbands correlate 0.00. Probably not reasonable. Assuming no (or
ignoring) missing data, the simplest design would be a one between (group),
two within (time and couple). I'd be interested in the group by time
interaction (and hoping the group by time by couple interaction was not
significant).

Given this analysis, there is at least one commerical program, PASS,
possibly, web-based programs, and a method using Manova that can compute
power/sample size. Others can provide recommendations on web-based programs.
I can't. There have been postings about how to use Manova for power
computations and there is an article describing how to do it. I  don't
recall the citation, although I think it was in something like Behavior,
Research Methods, and Instrumentation. Perhaps someone has the citation. I
also think that this method is presented in Ray Levesque's book pdf that is
downloadable from spss for free (your spss documentation directory will have
file about this book). I also have example setups that I could send to you.

If you decide to go the Manova route this is how it works. You use Matrix
data to define a correlation matrix with means and SDs for the design you
will analyze. Manova will read that matrix, analyze it, and compute power
values for the effects. The major issue is defining the correlation matrix.
Once you understand it, it's kind of easy to do but a little tedious; it's
also easy to make mistakes. Those things aside, the imporant point is that
it is an iterataive process: modify the matrix, run the analysis, modify the
matrix, etc.

If you have missing data, an alternative analysis method is Mixed because
cases with a missing husband posttest or wife pretest can be included. While
that procedure computes effect size and power values, the program will not
analyze matrix data input. There are power computation programs (free, I
think) for mixed/multilevel models as I have seen them mentioned on the SEM
and ML listservs. I don't recall their names; perhaps others do.

Gene Maguin


>>I have a quick question about sample size calculations.  I'm currently
working on a study that deals with couples therapy.  There will be several
couples assigned to two different groups.  I would like to compare the two
groups in hopes that the intervention group is better.  I wanted to know how
best to perform a sample size for couples.  In this study, I will get
feedback from both persons.  Since this is the case, I'm thinking that I
would just include each person as an individual and perform the calculation
as normal.  But I run into the problem of "paired" data within each group.
This is where I get confused.  Any assistance that you can give regarding
this matter is greatly appreciated.

Thank you in advance,

Johnita

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

[Bruce Weaver]

Gene Maguin wrote

Johnita,

The standard frame for this type of question is what sample size do I need
to obtain a power of P (typically, .80 in social science research) given an
effect size of ES from this analysis method and an alpha level of A
(typically, .05). Intertwined in the sample size question is an analysis
question because  different analysis methods yield different powers given
the same data.

So it seems like your design is a two group. Although you don't say anything
about number of assessments, let's assume pre and post. Ignoring 'couple'
and treating each person as an individual means that you are assuming that
wives and husbands correlate 0.00. Probably not reasonable. Assuming no (or
ignoring) missing data, the simplest design would be a one between (group),
two within (time and couple). I'd be interested in the group by time
interaction (and hoping the group by time by couple interaction was not
significant).

Given this analysis, there is at least one commerical program, PASS,
possibly, web-based programs, and a method using Manova that can compute
power/sample size. Others can provide recommendations on web-based programs.
I can't. There have been postings about how to use Manova for power
computations and there is an article describing how to do it. I  don't
recall the citation, although I think it was in something like Behavior,
Research Methods, and Instrumentation. Perhaps someone has the citation. I
also think that this method is presented in Ray Levesque's book pdf that is
downloadable from spss for free (your spss documentation directory will have
file about this book). I also have example setups that I could send to you.

If you decide to go the Manova route this is how it works. You use Matrix
data to define a correlation matrix with means and SDs for the design you
will analyze. Manova will read that matrix, analyze it, and compute power
values for the effects. The major issue is defining the correlation matrix.
Once you understand it, it's kind of easy to do but a little tedious; it's
also easy to make mistakes. Those things aside, the imporant point is that
it is an iterataive process: modify the matrix, run the analysis, modify the
matrix, etc.

If you have missing data, an alternative analysis method is Mixed because
cases with a missing husband posttest or wife pretest can be included. While
that procedure computes effect size and power values, the program will not
analyze matrix data input. There are power computation programs (free, I
think) for mixed/multilevel models as I have seen them mentioned on the SEM
and ML listservs. I don't recall their names; perhaps others do.

Gene Maguin

The MANOVA method Gene mentions is illustrated here:

  http://www.angelfire.com/wv/bwhomedir/spss/power_analysis.txt

Example 3 is the one that matches the OP's situation best.  

When this method was discussed in the SPSS newsgroup some time ago, Rich Ulrich (a frequent poster there) had some concerns, as noted here:

  http://www.angelfire.com/wv/bwhomedir/spss/power_analysis_part2.txt

But AFAIK, the MANOVA method gives results similar to those you get from PASS for the same designs.  

[Byrd Sellers, Johnita]

Great, thanks!  This information is very helpful.  Suppose that I find
that I need a sample size of 60 people in each group.  Would this be 30
couples (to equal a total of 60 subjects) or a sample size of 60 couples
(a total of 120 subjects)?  Thanks in advance for any help.

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Bruce Weaver
Sent: Tuesday, May 04, 2010 1:06 PM
To: [hidden email]
Subject: Re: Sample size question

Gene Maguin wrote:
>
> Johnita,
>
> The standard frame for this type of question is what sample size do I
need
> to obtain a power of P (typically, .80 in social science research)
given
> an
> effect size of ES from this analysis method and an alpha level of A
> (typically, .05). Intertwined in the sample size question is an
analysis
> question because  different analysis methods yield different powers
given
> the same data.
>
> So it seems like your design is a two group. Although you don't say
> anything
> about number of assessments, let's assume pre and post. Ignoring
'couple'
> and treating each person as an individual means that you are assuming
that
> wives and husbands correlate 0.00. Probably not reasonable. Assuming
no
> (or
> ignoring) missing data, the simplest design would be a one between
> (group),
> two within (time and couple). I'd be interested in the group by time
> interaction (and hoping the group by time by couple interaction was
not
> significant).
>
> Given this analysis, there is at least one commerical program, PASS,
> possibly, web-based programs, and a method using Manova that can
compute
> power/sample size. Others can provide recommendations on web-based
> programs.
> I can't. There have been postings about how to use Manova for power
> computations and there is an article describing how to do it. I  don't
> recall the citation, although I think it was in something like
Behavior,
> Research Methods, and Instrumentation. Perhaps someone has the
citation. I
> also think that this method is presented in Ray Levesque's book pdf
that
> is
> downloadable from spss for free (your spss documentation directory
will
> have
> file about this book). I also have example setups that I could send to
> you.
>
> If you decide to go the Manova route this is how it works. You use
Matrix
> data to define a correlation matrix with means and SDs for the design
you
> will analyze. Manova will read that matrix, analyze it, and compute
power
> values for the effects. The major issue is defining the correlation
> matrix.
> Once you understand it, it's kind of easy to do but a little tedious;
it's
> also easy to make mistakes. Those things aside, the imporant point is
that
> it is an iterataive process: modify the matrix, run the analysis,
modify
> the
> matrix, etc.
>
> If you have missing data, an alternative analysis method is Mixed
because
> cases with a missing husband posttest or wife pretest can be included.
> While
> that procedure computes effect size and power values, the program will
not
> analyze matrix data input. There are power computation programs (free,
I
> think) for mixed/multilevel models as I have seen them mentioned on
the
> SEM
> and ML listservs. I don't recall their names; perhaps others do.
>
> Gene Maguin
>
>

The MANOVA method Gene mentions is illustrated here:

  http://www.angelfire.com/wv/bwhomedir/spss/power_analysis.txt

Example 3 is the one that matches the OP's situation best.

When this method was discussed in the SPSS newsgroup some time ago, Rich
Ulrich (a frequent poster there) had some concerns, as noted here:

  http://www.angelfire.com/wv/bwhomedir/spss/power_analysis_part2.txt

But AFAIK, the MANOVA method gives results similar to those you get from
PASS for the same designs.



-----
--
Bruce Weaver
[hidden email]
http://sites.google.com/a/lakeheadu.ca/bweaver/
"When all else fails, RTFM."

NOTE:  My Hotmail account is not monitored regularly.
To send me an e-mail, please use the address shown above.
--
View this message in context:
http://old.nabble.com/Sample-size-question-tp28447959p28450423.html
Sent from the SPSSX Discussion mailing list archive at 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

[Maguin, Eugene]

Hi Johnita,

Couples. Because you are recruiting couples to take part in the
intervention.

Gene Maguin

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Byrd Sellers, Johnita
Sent: Monday, May 10, 2010 10:48 AM
To: [hidden email]
Subject: FW: Sample size question

Great, thanks!  This information is very helpful.  Suppose that I find
that I need a sample size of 60 people in each group.  Would this be 30
couples (to equal a total of 60 subjects) or a sample size of 60 couples
(a total of 120 subjects)?  Thanks in advance for any help.

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Bruce Weaver
Sent: Tuesday, May 04, 2010 1:06 PM
To: [hidden email]
Subject: Re: Sample size question

Gene Maguin wrote:
>
> Johnita,
>
> The standard frame for this type of question is what sample size do I
need
> to obtain a power of P (typically, .80 in social science research)
given
> an
> effect size of ES from this analysis method and an alpha level of A
> (typically, .05). Intertwined in the sample size question is an
analysis
> question because  different analysis methods yield different powers
given
> the same data.
>
> So it seems like your design is a two group. Although you don't say
> anything
> about number of assessments, let's assume pre and post. Ignoring
'couple'
> and treating each person as an individual means that you are assuming
that
> wives and husbands correlate 0.00. Probably not reasonable. Assuming
no
> (or
> ignoring) missing data, the simplest design would be a one between
> (group),
> two within (time and couple). I'd be interested in the group by time
> interaction (and hoping the group by time by couple interaction was
not
> significant).
>
> Given this analysis, there is at least one commerical program, PASS,
> possibly, web-based programs, and a method using Manova that can
compute
> power/sample size. Others can provide recommendations on web-based
> programs.
> I can't. There have been postings about how to use Manova for power
> computations and there is an article describing how to do it. I  don't
> recall the citation, although I think it was in something like
Behavior,
> Research Methods, and Instrumentation. Perhaps someone has the
citation. I
> also think that this method is presented in Ray Levesque's book pdf
that
> is
> downloadable from spss for free (your spss documentation directory
will
> have
> file about this book). I also have example setups that I could send to
> you.
>
> If you decide to go the Manova route this is how it works. You use
Matrix
> data to define a correlation matrix with means and SDs for the design
you
> will analyze. Manova will read that matrix, analyze it, and compute
power
> values for the effects. The major issue is defining the correlation
> matrix.
> Once you understand it, it's kind of easy to do but a little tedious;
it's
> also easy to make mistakes. Those things aside, the imporant point is
that
> it is an iterataive process: modify the matrix, run the analysis,
modify
> the
> matrix, etc.
>
> If you have missing data, an alternative analysis method is Mixed
because
> cases with a missing husband posttest or wife pretest can be included.
> While
> that procedure computes effect size and power values, the program will
not
> analyze matrix data input. There are power computation programs (free,
I
> think) for mixed/multilevel models as I have seen them mentioned on
the
> SEM
> and ML listservs. I don't recall their names; perhaps others do.
>
> Gene Maguin
>
>

The MANOVA method Gene mentions is illustrated here:

  http://www.angelfire.com/wv/bwhomedir/spss/power_analysis.txt

Example 3 is the one that matches the OP's situation best.

When this method was discussed in the SPSS newsgroup some time ago, Rich
Ulrich (a frequent poster there) had some concerns, as noted here:

  http://www.angelfire.com/wv/bwhomedir/spss/power_analysis_part2.txt

But AFAIK, the MANOVA method gives results similar to those you get from
PASS for the same designs.



-----
--
Bruce Weaver
[hidden email]
http://sites.google.com/a/lakeheadu.ca/bweaver/
"When all else fails, RTFM."

NOTE:  My Hotmail account is not monitored regularly.
To send me an e-mail, please use the address shown above.
--
View this message in context:
http://old.nabble.com/Sample-size-question-tp28447959p28450423.html
Sent from the SPSSX Discussion mailing list archive at 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

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