Fw: Sample size for repeated measure
Locked
 
|
A problem that Gene and Bruce seem to be ignoring is the following:
Chunki wrote: "Cases will assumed to be dropped out over the time points." If I am interpreting this correctly, although one would like to treat this as 2x5 "mixed" design (i.e., a design with a between-subject and a within-subject factor) and do a repeated measures ANOVA, if there are drop outs over time, one will not have complete data at all time points. If one uses only cases with complete data, then modifying the syntax that Gene provides could work. But what percentage of cases are dropouts? If the percentage is small, then using complete cases could be okay but if the percentage is large, then the repeated measures ANOVa is probably not the right analysis to do and it is likely that it will have low power. A multilevel analysis with repeated mesures nested within subjects would seem to be the more appropriate analysis because it would use the data from all subjects. I believe that there are ways/programs that will allow one to calculate the power of such designs but one would probably have to ask about this on the multilevel mailing list. -Mike Palij New York University [hidden email] ----- Original Message ----- From: "Gene Maguin" <[hidden email]> To: <[hidden email]> Sent: Wednesday, November 03, 2010 4:58 PM Subject: Re: Sample size for repeated measure > Chunki Fong, > > This is the article you need to look at. The message is from Bruce Weaver > and posted back in 2003 (Bruce, I didn't realize you were on the list back > then.) > > Hello group, > Just the other day, I came across an article in Behavior Research > Methods, Instruments & Computers (2001, Vol 33, 479-484) that showed how > to perform power analysis (and sample size estimation) for various > ANOVA/ANCOVA type models using Matrix Data and the MANOVA procedure. If > anyone is interested, here is a syntax file I wrote that performs the > analyses for the 3 examples given in the paper: > > Here is a specific example. > > * ---------------------------------------------------------- . > * Example 3: Mixed design (between-within) ANOVA . > * ---------------------------------------------------------- . > > * Dependent variable: Ratings of depression . > * Between-subjects factor: Group (3 levels). > * Within-subjects factor: Time (DV measured on 3 occasions). > > * The effect of most interest is the Group x Time interaction. > * How many subjects do we need to have 80% power for this term?. > > * Read in data in matrix format . > > matrix data > variables = group rowtype_ dv1 dv2 dv3 > /factor = group > /format = lower nodiagonal. > begin data > 1 mean 0.00 0.21 0.42 > 1 n 100 100 100 > 2 mean 0.00 0.21 0.42 > 2 n 100 100 100 > 3 mean 0.00 0.00 0.00 > 3 n 100 100 100 > . sd 1.0 1.0 1.0 > . corr 0.4 > . corr 0.4 0.4 > end data. > > manova dv1 dv2 dv3 by group(1,3)/wsfactors dv(3)/method = unique > /error = within+residual/matrix = in(*)/power t (.05) F (.05) > /print signif (mult averf)/noprint param(estim). > > > Once you understand the structure you can easily modify this example for > your needs. I don't know what advice to offer for your dichotomous outcome > because the example will give a correct computation for a dichotomous DV. > There isn't anything else in spss. Perhaps there is something on the web for > repeated measures for dichotomous DVs. Perhaps others will comment. > > Gene Maguin > > > > ________________________________ > > From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of > Chunki Fong > Sent: Wednesday, November 03, 2010 3:04 PM > To: [hidden email] > Subject: Sample size for repeated measure > >>>>I like to determine a sample size for the two study groups for the > following design: > Two groups comparing two primary measures (one continuous and > onedichotomous) over 5 time points over 1 year (baseline, 3m, 6m, 9m, > and12m). Cases will assumed to be dropped out over the time points. > Maysomeone give me some guidance in getting the minimun required samplesize > (power analysis). Thank you. > > ===================== > 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 |
Re: Fw: Sample size for repeated measure
Administrator
|
Good catch, Mike. I did overlook the drop-outs. And I agree that using a multilevel model (i.e., MIXED) would be a nice way to handle that.
Simulation is always one way to obtain a sample size estimate--and I know that PASS now includes simulation for some things. So the OP may want to look into that. Gene, that old post of mine you dug up, was it to this list, or to the comp.soft-sys.stat.spss newsgroup? I was thinking of suggesting that too, but thought we needed more information first. Cheers, Bruce
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
|
|
In reply to this post by Mike
Thank you Gene, Bruce, and Mike for the comments.
If I understand correctly, the power analysis for a repeated measure can be done by Bruce's manova code posted by Gene. However, since attrition is taking into account, as Mike had pointed out, this extra restriction over-estimate the power, which would means a higher sample size is required than the resulting analysis for repeating measure. Mixed model for multilevel modeling is a better option. As I was reading the and doing some more search, I came across this article. Although this is an SPSS list, this article is about SAS PROC MIXED and appears to be relevant to this topic which may of interest to some. http://www.stat.ncsu.edu/people/arellano/courses/st524/Fall08/Homeworks/homewotk2/Stroup_MixedModelspower_sas.pdf In a message dated 11/3/2010 6:21:48 PM Eastern Daylight Time, [hidden email] writes: Good catch, Mike. I did overlook the drop-outs. And I agree that using a |
Re: Fw: Sample size for repeated measure
|
|
In reply to this post by Bruce Weaver
Hi Bruce,
I went back and looked at the message and although Ray Levesque posted it to this list on Jan 15, 2003, it looks like he got from another place because there is this line before your salutation: '*** Copy of 2002/12/12 post to spss newsgroup ****.' So perhaps it did come from the comp.soft... List. Regardless, I certainly glad you posted it and Ray forwarded it because it is very useful. It's certainly not all-inclusively useful but in the absence of access to a power analysis program such as PASS, it's very helpful. And, it actually has a broader range because any design that can be analyzed can be specified to develop power values. Gene -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Bruce Weaver Sent: Wednesday, November 03, 2010 6:20 PM To: [hidden email] Subject: Re: Fw: Sample size for repeated measure Good catch, Mike. I did overlook the drop-outs. And I agree that using a multilevel model (i.e., MIXED) would be a nice way to handle that. Simulation is always one way to obtain a sample size estimate--and I know that PASS now includes simulation for some things. So the OP may want to look into that. Gene, that old post of mine you dug up, was it to this list, or to the comp.soft-sys.stat.spss newsgroup? I was thinking of suggesting that too, but thought we needed more information first. Cheers, Bruce Mike Palij wrote: > > A problem that Gene and Bruce seem to be ignoring is the following: > Chunki wrote: > "Cases will assumed to be dropped out over the time points." > > If I am interpreting this correctly, although one would like to > treat this as 2x5 "mixed" design (i.e., a design with a between-subject > and a within-subject factor) and do a repeated measures ANOVA, > if there are drop outs over time, one will not have complete data > at all time points. If one uses only cases with complete data, then > modifying the syntax that Gene provides could work. But what > percentage of cases are dropouts? If the percentage is small, then > using complete cases could be okay but if the percentage is large, > then the repeated measures ANOVa is probably not the right > analysis to do and it is likely that it will have low power. > > A multilevel analysis with repeated mesures nested within subjects > would seem to be the more appropriate analysis because it would > use the data from all subjects. I believe that there are ways/programs > that will allow one to calculate the power of such designs but one > would probably have to ask about this on the multilevel mailing list. > > -Mike Palij > New York University > [hidden email] > > > ----- Original Message ----- > From: "Gene Maguin" <[hidden email]> > To: <[hidden email]> > Sent: Wednesday, November 03, 2010 4:58 PM > Subject: Re: Sample size for repeated measure > > >> Chunki Fong, >> >> This is the article you need to look at. The message is from Bruce Weaver >> and posted back in 2003 (Bruce, I didn't realize you were on the list >> back >> then.) >> >> Hello group, >> Just the other day, I came across an article in Behavior Research >> Methods, Instruments & Computers (2001, Vol 33, 479-484) that showed how >> to perform power analysis (and sample size estimation) for various >> ANOVA/ANCOVA type models using Matrix Data and the MANOVA procedure. If >> anyone is interested, here is a syntax file I wrote that performs the >> analyses for the 3 examples given in the paper: >> >> Here is a specific example. >> >> * ---------------------------------------------------------- . >> * Example 3: Mixed design (between-within) ANOVA . >> * ---------------------------------------------------------- . >> >> * Dependent variable: Ratings of depression . >> * Between-subjects factor: Group (3 levels). >> * Within-subjects factor: Time (DV measured on 3 occasions). >> >> * The effect of most interest is the Group x Time interaction. >> * How many subjects do we need to have 80% power for this term?. >> >> * Read in data in matrix format . >> >> matrix data >> variables = group rowtype_ dv1 dv2 dv3 >> /factor = group >> /format = lower nodiagonal. >> begin data >> 1 mean 0.00 0.21 0.42 >> 1 n 100 100 100 >> 2 mean 0.00 0.21 0.42 >> 2 n 100 100 100 >> 3 mean 0.00 0.00 0.00 >> 3 n 100 100 100 >> . sd 1.0 1.0 1.0 >> . corr 0.4 >> . corr 0.4 0.4 >> end data. >> >> manova dv1 dv2 dv3 by group(1,3)/wsfactors dv(3)/method = unique >> /error = within+residual/matrix = in(*)/power t (.05) F (.05) >> /print signif (mult averf)/noprint param(estim). >> >> >> Once you understand the structure you can easily modify this example for >> your needs. I don't know what advice to offer for your dichotomous >> outcome >> because the example will give a correct computation for a dichotomous DV. >> There isn't anything else in spss. Perhaps there is something on the web >> for >> repeated measures for dichotomous DVs. Perhaps others will comment. >> >> Gene Maguin >> >> >> >> ________________________________ >> >> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of >> Chunki Fong >> Sent: Wednesday, November 03, 2010 3:04 PM >> To: [hidden email] >> Subject: Sample size for repeated measure >> >>>>>I like to determine a sample size for the two study groups for the >> following design: >> Two groups comparing two primary measures (one continuous and >> onedichotomous) over 5 time points over 1 year (baseline, 3m, 6m, 9m, >> and12m). Cases will assumed to be dropped out over the time points. >> Maysomeone give me some guidance in getting the minimun required >> samplesize >> (power analysis). Thank you. >> >> ===================== >> 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 > > ----- -- 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://spssx-discussion.1045642.n5.nabble.com/Fw-Sample-size-for-repeated-me asure-tp3249253p3249263.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 |
|
|
In reply to this post by C. Fong
Don Hedeker and Robert Gibbons of U of Illinois-Chicago have
worked on
these designs and provide additional background on how to
analyze this
type of design/situation; The first link below is to an
1999 article by
Hedeker, Gibbons & Waternaux in the Journal of Educational
and Behavioral
Statistics (on the Sagepub website - abstract available,
subscription is required
for text). The second link is to an article by Roy et al
that was published
in Biometrics 2006 (full text is available, it is technical
and does not provide
computer code):
-Mike Palij
New York University
|
Re: Fw: Sample size for repeated measure
|
|
In reply to this post by C. Fong
Chunki,
That article is a really excellent discovery. Thank you for finding it and posting the reference. The big question for spss users is whether we can do the computations in spss. Has anyone used spss to implement the procedure described? How well does it work? I'd be very interested to hear from you. Perhaps one of spss' statisticians could comment as well. Gene Maguin ________________________________ From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Chunki Fong Sent: Wednesday, November 03, 2010 10:57 PM To: [hidden email] Subject: Re: Fw: Sample size for repeated measure Thank you Gene, Bruce, and Mike for the comments. If I understand correctly, the power analysis for a repeated measure can be done by Bruce's manova code posted by Gene. However, since attrition is taking into account, as Mike had pointed out, this extra restriction over-estimate the power, which would means a higher sample size is required than the resulting analysis for repeating measure. Mixed model for multilevel modeling is a better option. As I was reading the and doing some more search, I came across this article. Although this is an SPSS list, this article is about SAS PROC MIXED and appears to be relevant to this topic which may of interest to some. http://www.stat.ncsu.edu/people/arellano/courses/st524/Fall08/Homeworks/home wotk2/Stroup_MixedModelspower_sas.pdf In a message dated 11/3/2010 6:21:48 PM Eastern Daylight Time, [hidden email] writes: Good catch, Mike. I did overlook the drop-outs. And I agree that using a multilevel model (i.e., MIXED) would be a nice way to handle that. Simulation is always one way to obtain a sample size estimate--and I know that PASS now includes simulation for some things. So the OP may want to look into that. Gene, that old post of mine you dug up, was it to this list, or to the comp.soft-sys.stat.spss newsgroup? I was thinking of suggesting that too, but thought we needed more information first. Cheers, Bruce Mike Palij wrote: > >A problem that Gene and Bruce seem to be ignoring is the following: >Chunki wrote: >"Cases will assumed to be dropped out over the time points." > >If I am interpreting this correctly, although one would like to >treat this as 2x5 "mixed" design (i.e., a design with a between-subject >and a within-subject factor) and do a repeated measures ANOVA, >if there are drop outs over time, one will not have complete data >at all time points. If one uses only cases with complete data, then >modifying the syntax that Gene provides could work. But what >percentage of cases are dropouts? If the percentage is small, then >using complete cases could be okay but if the percentage is large, >then the repeated measures ANOVa is probably not the right >analysis to do and it is likely that it will have low power. > >A multilevel analysis with repeated mesures nested within subjects >would seem to be the more appropriate analysis because it would >use the data from all subjects. I believe that there are ways/programs >that will allow one to calculate the power of such designs but one >would probably have to ask about this on the multilevel mailing list. > >-Mike Palij >New York University >[hidden email] > > >----- Original Message ----- >From: "Gene Maguin" <[hidden email]> >To: <[hidden email]> >Sent: Wednesday, November 03, 2010 4:58 PM >Subject: Re: Sample size for repeated measure > > >>Chunki Fong, >> >>This is the article you need to look at. The message is from Bruce Weaver >>and posted back in 2003 (Bruce, I didn't realize you were on the list >>back >>then.) >> >>Hello group, >> Just the other day, I came across an article in Behavior Research >>Methods, Instruments &Computers (2001, Vol 33, 479-484) that showed how >>to perform power analysis (and sample size estimation) for various >>ANOVA/ANCOVA type models using Matrix Data and the MANOVA procedure. If >>anyone is interested, here is a syntax file I wrote that performs the >>analyses for the 3 examples given in the paper: >> >>Here is a specific example. >> >>* ---------------------------------------------------------- . >>* Example 3: Mixed design (between-within) ANOVA . >>* ---------------------------------------------------------- . >> >>* Dependent variable: Ratings of depression . >>* Between-subjects factor: Group (3 levels). >>* Within-subjects factor: Time (DV measured on 3 occasions). >> >>* The effect of most interest is the Group x Time interaction. >>* How many subjects do we need to have 80% power for this term?. >> >>* Read in data in matrix format . >> >>matrix data >>variables = group rowtype_ dv1 dv2 dv3 >>/factor = group >>/format = lower nodiagonal. >>begin data >>1 mean 0.00 0.21 0.42 >>1 n 100 100 100 >>2 mean 0.00 0.21 0.42 >>2 n 100 100 100 >>3 mean 0.00 0.00 0.00 >>3 n 100 100 100 >>. sd 1.0 1.0 1.0 >>. corr 0.4 >>. corr 0.4 0.4 >>end data. >> >>manova dv1 dv2 dv3 by group(1,3)/wsfactors dv(3)/method = unique >>/error = within+residual/matrix = in(*)/power t (.05) F (.05) >>/print signif (mult averf)/noprint param(estim). >> >> >>Once you understand the structure you can easily modify this example for >>your needs. I don't know what advice to offer for your dichotomous >>outcome >>because the example will give a correct computation for a dichotomous DV. >>There isn't anything else in spss. Perhaps there is something on the web >>for >>repeated measures for dichotomous DVs. Perhaps others will comment. >> >>Gene Maguin >> >> >> >>________________________________ >> >>From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of >>Chunki Fong >>Sent: Wednesday, November 03, 2010 3:04 PM >>To: [hidden email] >>Subject: Sample size for repeated measure >> >>>>>I like to determine a sample size for the two study groups for the >>following design: >>Two groups comparing two primary measures (one continuous and >>onedichotomous) over 5 time points over 1 year (baseline, 3m, 6m, 9m, >>and12m). Cases will assumed to be dropped out over the time points. >>Maysomeone give me some guidance in getting the minimun required >>samplesize >>(power analysis). Thank you. >> >>===================== >>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 > > ----- -- 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://spssx-discussion.1045642.n5.nabble.com/Fw-Sample-size-for-repeated-me asure-tp3249253p3249263.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 |
|
|
I just came across this thread when looking for some information about power calculations for repeated measures designs. The article and code that Chunki posted is great. I searched for the original article and found another method of doing power calcs using a similar technique with nice explanations in this article.
http://digitalcommons.wayne.edu/cgi/viewcontent.cgi?article=1350&context=jmasm I've used sample power and some online power calculators but using SPSS syntax really gives me a better understanding of the implications of the different correlations and effect sizes. Carol |
Re: Sample size for repeated measure
|
|
A message posted by Bruce Weaver to a Usenet group in 2004
=====================
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
http://sci.tech-archive.net/Archive/sci.stat.edu/2004-11/0121.html points to an earlier post by me, where I complained about the ease of being misled by results from SPSS-MANOVA when it comes to power analyses. https://groups.google.com/forum/#!msg/comp.soft-sys.stat.spss/OhVYmU8oR6Y/VebFiYlt7OAJ That post to the SPSS group includes code that illustrated the problem, which would be most acute if you thought to use very-small-sample data as a pilot for power analysis. I scanned the article that Carol cites below; it starts out with intentional, designedly small Ns and modifies the matrix output in order to look at other numbers. Perhaps it is fine, but I don't have SPSS to test-drive it, so I'm not going to study it further. I guess my warning is: Don't try to use the SPSS MANOVA power results unless you have enough documentation to know that you are using them sensibly. That reference might or might not be correct enough and clear enough. -- Rich Ulrich > Date: Thu, 31 Jul 2014 09:11:33 -0700 > From: [hidden email] > Subject: Re: Fw: Sample size for repeated measure > To: [hidden email] > > I just came across this thread when looking for some information about power > calculations for repeated measures designs. The article and code that Chunki > posted is great. I searched for the original article and found another > method of doing power calcs using a similar technique with nice explanations > in this article. > > > http://digitalcommons.wayne.edu/cgi/viewcontent.cgi?article=1350&context=jmasm > > I've used sample power and some online power calculators but using SPSS > syntax really gives me a better understanding of the implications of the > different correlations and effect sizes. > > Carol > ... |
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |