ANCOVA or mixed models
2,722 views
Skip to first unread message
Patrick
Jan 20, 2012, 6:54:38 PM1/20/12
to MedStats
Hi all,
My current understanding of the literature is that ANCOVA is the
prefered method of analysis in an RCT where you have follow-up data at
one post treatment time point as well as a baseline measurement of the
outcome.
I have a situation where I have two post treatment time points and
want to test the hypothesis that there is no difference between groups
post treatment. My current approach to this analysis is also an ANCOVA
type model within a GLMM framework to adjust for the correlated data
(i.e. two observations per person). My model includes the predictors
of treatment group, baseline value of the outcome and time (2 time
points). Does anyone know of any literature on this or does anyone
have an opinion of this approach?
Thanks,
Patrick
My current understanding of the literature is that ANCOVA is the
prefered method of analysis in an RCT where you have follow-up data at
one post treatment time point as well as a baseline measurement of the
outcome.
I have a situation where I have two post treatment time points and
want to test the hypothesis that there is no difference between groups
post treatment. My current approach to this analysis is also an ANCOVA
type model within a GLMM framework to adjust for the correlated data
(i.e. two observations per person). My model includes the predictors
of treatment group, baseline value of the outcome and time (2 time
points). Does anyone know of any literature on this or does anyone
have an opinion of this approach?
Thanks,
Patrick
John Sorkin
Jan 20, 2012, 7:43:01 PM1/20/12
to MedStats
Patrick
I believe you have three observations per patient, not three, i.e. baseline, post Rx 1 and post Rx2.
Repeated measures ANCOVA is one option, but as typically done this requires sphericity. Using ANCOVA with an appropriately selected variance-covariance structure (e.g. unstructured, exchangable, AR(1), etc.), which sounds like what you are thinking about using is a fine approach. It allows subjects to contribute to the analysis even if they are missing data at one of the three time points, under the assumption that the pattern of missing data is MCAR (missing completely at random). You can also use random effects ANOVA in which you let each subject have his, or her, own intercept or intercept and slope.
John
I believe you have three observations per patient, not three, i.e. baseline, post Rx 1 and post Rx2.
Repeated measures ANCOVA is one option, but as typically done this requires sphericity. Using ANCOVA with an appropriately selected variance-covariance structure (e.g. unstructured, exchangable, AR(1), etc.), which sounds like what you are thinking about using is a fine approach. It allows subjects to contribute to the analysis even if they are missing data at one of the three time points, under the assumption that the pattern of missing data is MCAR (missing completely at random). You can also use random effects ANOVA in which you let each subject have his, or her, own intercept or intercept and slope.
John
John David Sorkin M.D., Ph.D.
Chief, Biostatistics and Informatics
University of Maryland School of Medicine Division of Gerontology
Baltimore VA Medical Center
10 North Greene Street
GRECC (BT/18/GR)
Baltimore, MD 21201-1524
(Phone) 410-605-7119
(Fax) 410-605-7913 (Please call phone number above prior to faxing)
>>> Patrick <patrick....@gmail.com> 1/20/2012 6:54 PM >>>
Hi all,
Thanks,
Patrick
--
To post a new thread to MedStats, send email to MedS...@googlegroups.com .
MedStats' home page is http://groups.google.com/group/MedStats .
Rules: http://groups.google.com/group/MedStats/web/medstats-rules
Confidentiality Statement:
This email message, including any attachments, is for the sole use of the intended recipient(s) and may contain confidential and privileged information. Any unauthorized use, disclosure or distribution is prohibited. If you are not the intended recipient, please contact the sender by reply email and destroy all copies of the original message.
Bruce Weaver
Feb 7, 2012, 3:05:38 PM2/7/12
to MedStats
I apologize for the tardiness of my reply, but this thread only
arrived in my inbox today.
Re sphericity, it is defined as homogeneity of variance of all of the
pair-wise difference scores for the levels of a repeated measures
factor. When there are only two levels (which is the case here), then
there is only one pair-wise difference score, and so it is impossible
to violate the assumption of sphericity.
For readers who are interested, Thom Baguley has a very nice note on
sphericity, which can be seen here:
http://homepages.gold.ac.uk/aphome/spheric.html
Cheers!
Bruce
p.s. - Anyone using SPSS for repeated measures models that include
covariates might want to look at this:
http://journals.cambridge.org/download.php?file=%2FINS%2FINS13_05%2FS1355617707071147a.pdf&code=1161664e8b41c165ff41ae708a1ef5e1
On Jan 20, 7:43 pm, "John Sorkin" <jsor...@grecc.umaryland.edu> wrote:
> Patrick
> I believe you have three observations per patient, not three, i.e. baseline, post Rx 1 and post Rx2.
> Repeated measures ANCOVA is one option, but as typically done this requires sphericity. Using ANCOVA with an appropriately selected variance-covariance structure (e.g. unstructured, exchangable, AR(1), etc.), which sounds like what you are thinking about using is a fine approach. It allows subjects to contribute to the analysis even if they are missing data at one of the three time points, under the assumption that the pattern of missing data is MCAR (missing completely at random). You can also use random effects ANOVA in which you let each subject have his, or her, own intercept or intercept and slope.
> John
>
> John David Sorkin M.D., Ph.D.
> Chief, Biostatistics and Informatics
> University of Maryland School of Medicine Division of Gerontology
> Baltimore VA Medical Center
> 10 North Greene Street
> GRECC (BT/18/GR)
> Baltimore, MD 21201-1524
> (Phone) 410-605-7119
> (Fax) 410-605-7913 (Please call phone number above prior to faxing)
>
> >>> Patrick <patrick.mceld...@gmail.com> 1/20/2012 6:54 PM >>>
arrived in my inbox today.
Re sphericity, it is defined as homogeneity of variance of all of the
pair-wise difference scores for the levels of a repeated measures
factor. When there are only two levels (which is the case here), then
there is only one pair-wise difference score, and so it is impossible
to violate the assumption of sphericity.
For readers who are interested, Thom Baguley has a very nice note on
sphericity, which can be seen here:
http://homepages.gold.ac.uk/aphome/spheric.html
Cheers!
Bruce
p.s. - Anyone using SPSS for repeated measures models that include
covariates might want to look at this:
http://journals.cambridge.org/download.php?file=%2FINS%2FINS13_05%2FS1355617707071147a.pdf&code=1161664e8b41c165ff41ae708a1ef5e1
On Jan 20, 7:43 pm, "John Sorkin" <jsor...@grecc.umaryland.edu> wrote:
> Patrick
> I believe you have three observations per patient, not three, i.e. baseline, post Rx 1 and post Rx2.
> Repeated measures ANCOVA is one option, but as typically done this requires sphericity. Using ANCOVA with an appropriately selected variance-covariance structure (e.g. unstructured, exchangable, AR(1), etc.), which sounds like what you are thinking about using is a fine approach. It allows subjects to contribute to the analysis even if they are missing data at one of the three time points, under the assumption that the pattern of missing data is MCAR (missing completely at random). You can also use random effects ANOVA in which you let each subject have his, or her, own intercept or intercept and slope.
> John
>
> John David Sorkin M.D., Ph.D.
> Chief, Biostatistics and Informatics
> University of Maryland School of Medicine Division of Gerontology
> Baltimore VA Medical Center
> 10 North Greene Street
> GRECC (BT/18/GR)
> Baltimore, MD 21201-1524
> (Phone) 410-605-7119
> (Fax) 410-605-7913 (Please call phone number above prior to faxing)
>
Bruce Weaver
Feb 8, 2012, 2:53:49 PM2/8/12
to MedStats
Apparently the second link I gave below wasn't working for some
people, so here's full info on that item.
Cheers,
Bruce
LETTER TO THE EDITOR
Use of covariates in randomized controlled trials
GERARD J.P. VAN BREUKELEN [1] and KOENE R.A. VAN DIJK [2,3]
1 Department of Methodology and Statistics, Maastricht University, The
Netherlands
2 Department of Neurocognition, Maastricht, The Netherlands
3 Department of Psychiatry and Neuropsychology, Maastricht, The
Netherlands
(Received March 9, 2007; Final Revision April 5, 2007; Accepted April
5, 2007 )
Journal of the International Neuropsychological Society (2007), 13,
903–904.
Copyright © 2007 INS. Published by Cambridge University Press. Printed
in the USA.
DOI: 10.10170S1355617707071147
On Feb 7, 3:05 pm, Bruce Weaver <bwea...@lakeheadu.ca> wrote:
> I apologize for the tardiness of my reply, but this thread only
> arrived in my inbox today.
>
> Re sphericity, it is defined as homogeneity of variance of all of the
> pair-wise difference scores for the levels of a repeated measures
> factor. When there are only two levels (which is the case here), then
> there is only one pair-wise difference score, and so it is impossible
> to violate the assumption of sphericity.
>
> For readers who are interested, Thom Baguley has a very nice note on
> sphericity, which can be seen here:
>
> http://homepages.gold.ac.uk/aphome/spheric.html
>
> Cheers!
> Bruce
>
> p.s. - Anyone using SPSS for repeated measures models that include
> covariates might want to look at this:
>
> http://journals.cambridge.org/download.php?file=%2FINS%2FINS13_05%2FS...
people, so here's full info on that item.
Cheers,
Bruce
LETTER TO THE EDITOR
Use of covariates in randomized controlled trials
GERARD J.P. VAN BREUKELEN [1] and KOENE R.A. VAN DIJK [2,3]
1 Department of Methodology and Statistics, Maastricht University, The
Netherlands
2 Department of Neurocognition, Maastricht, The Netherlands
3 Department of Psychiatry and Neuropsychology, Maastricht, The
Netherlands
(Received March 9, 2007; Final Revision April 5, 2007; Accepted April
5, 2007 )
Journal of the International Neuropsychological Society (2007), 13,
903–904.
Copyright © 2007 INS. Published by Cambridge University Press. Printed
in the USA.
DOI: 10.10170S1355617707071147
On Feb 7, 3:05 pm, Bruce Weaver <bwea...@lakeheadu.ca> wrote:
> I apologize for the tardiness of my reply, but this thread only
> arrived in my inbox today.
>
> Re sphericity, it is defined as homogeneity of variance of all of the
> pair-wise difference scores for the levels of a repeated measures
> factor. When there are only two levels (which is the case here), then
> there is only one pair-wise difference score, and so it is impossible
> to violate the assumption of sphericity.
>
> For readers who are interested, Thom Baguley has a very nice note on
> sphericity, which can be seen here:
>
> http://homepages.gold.ac.uk/aphome/spheric.html
>
> Cheers!
> Bruce
>
> p.s. - Anyone using SPSS for repeated measures models that include
> covariates might want to look at this:
>
Frank Harrell
Feb 10, 2012, 11:58:53 AM2/10/12
to MedStats
It is important to treat baseline as baseline, and as others have
said, to properly model correlation structure. This is easiest to do
with generalized least squares. If the two follow-up times are rigid,
you can just used an unstructured covariance matrix. If they really
vary by many days away from the target follow-up dates, then a
continuous time AR1 or similarly flexible structure is recommended.
Frank
said, to properly model correlation structure. This is easiest to do
with generalized least squares. If the two follow-up times are rigid,
you can just used an unstructured covariance matrix. If they really
vary by many days away from the target follow-up dates, then a
continuous time AR1 or similarly flexible structure is recommended.
Frank
Reply all
Reply to author
Forward
0 new messages