Paired t-test and Repeated measures ANOVA

[thejasvi tv]

Dear Experts,

 

One of my doctors is carrying out a study on Radiographic bone loss. He is measuring the values (in mm) at 4 different time intervals. I am planning to carry out a paired t-test in order to statistically evaluate the gain/loss. But another statistician is insisting on using a repeated measures ANOVA. I am really not sure as to whether repeated measures ANOVA can be used as we are collecting data at different time intervals.

Request you people to guide me in this regard.

 

Thanks and regards

Thejasvi

[SR Millis]

I think that you need to tell us more about the design
of the study, sample size, and aims/hypotheses--before
anyone can advise as to a statistical analysis plan.

SR Millis

Scott R Millis, PhD, MEd, ABPP (CN,CL,RP), CStat
Professor & Director of Research
Dept of Physical Medicine & Rehabilitation
Wayne State University School of Medicine
261 Mack Blvd
Detroit, MI 48201
Email: smi...@med.wayne.edu
Tel: 313-993-8085
Fax: 313-966-7682

[Bruce Weaver]

If the variability in the time intervals is not great, repeated
measures ANOVA will still give a reasonable model, I think. If there
is too much variability in the intervals, take a look at individual
growth curve analysis.

--
Bruce Weaver
bwe...@lakeheadu.ca
www.angelfire.com/wv/bwhomedir
"When all else fails, RTFM."

[Adrian Sayers]

I think even with irregular time intervals and random effects
regression model would be appropriate. This preserves the growth
trajectories which can be very important in small data sets. I think
RIGLS estimates are the most appropriate in small samples. In large
samples i dont think there is much difference between RIGLS and IGLS.

MLwin is free to down load for academics and will do the job nicely.
Stata i think can do random intercepts but thats it and its
RIGLS(REML) estimate is not quite right. SAS and PROC MIXED will also
do the job and the REML estimate works.

Adrian

[thejasvi tv]

Thank you all for the expert advice.....the study contains 10 samples and measurements are taken on two sides of a teeth and the time intervals are Baseline, 3 months, 6 months, 9 months and 12 months.

 

Pls give your comments/additional comments and advice as to how the analysis can be approached.

 

Thanking you again

 

--
Warm regards,
Thejasvi.

[Bruce Weaver]

On Jan 7, 12:18 pm, "thejasvi tv" <thejasvi...@gmail.com> wrote:
> Thank you all for the expert advice.....the study contains 10 samples and measurements are taken on two sides of a teeth and the time intervals are Baseline, 3 months, 6 months, 9 months and 12 months.

I don't see any variability in the time periods. Are those figures
approximate? If yes, how much variability is there around the 3, 6,
9, and 12 month points? Do all subjects have all time points?

[Ramzi]

Note that you not only have correlations between measurements within a
subject, but you also have correlation between the two measurements on
each subject within a time point. Off the top of my head, one way to
handle this would be to use a random tooth within subject intercept
(that would model the correlation between the two measurements on a
tooth) along with modeling the serial correlation within subject. If
you are using SAS PROC MIXED, be careful with how you model the serial
correlation. Which "TYPE"s you can use depend on whether or not the
time points are equally spaced and on whether or not the subjects have
the same times of follow-up.

FYI - A great book on this topic is Fitzmaurice, Laird, and Ware
(2004) 'Applied Longitudinal Analysis'.

Ramzi

On Jan 7, 12:18 pm, "thejasvi tv" <thejasvi...@gmail.com> wrote:
> Thank you all for the expert advice.....the study contains 10 samples and
> measurements are taken on two sides of a teeth and the time intervals are
> Baseline, 3 months, 6 months, 9 months and 12 months.
>
> Pls give your comments/additional comments and advice as to how the analysis
> can be approached.
>
> Thanking you again
>

> On 1/7/08, Adrian Sayers <adriansay...@gmail.com> wrote:
>
>
>
>
>
> > I think even with irregular time intervals and random effects
> > regression model would be appropriate. This preserves the growth
> > trajectories which can be very important in small data sets. I think
> > RIGLS estimates are the most appropriate in small samples. In large
> > samples i dont think there is much difference between RIGLS and IGLS.
>
> > MLwin is free to down load for academics and will do the job nicely.
> > Stata i think can do random intercepts but thats it and its
> > RIGLS(REML) estimate is not quite right. SAS and PROC MIXED will also
> > do the job and the REML estimate works.
>
> > Adrian
>

> > On 06/01/2008, Bruce Weaver <bwea...@lakeheadu.ca> wrote:
>
> > > On Jan 6, 12:31 pm, "thejasvi tv" <thejasvi...@gmail.com> wrote:
> > > > Dear Experts,
>
> > > > One of my doctors is carrying out a study on Radiographic bone loss.
> > He is
> > > > measuring the values (in mm) at 4 different time intervals. I am
> > planning to
> > > > carry out a paired t-test in order to statistically evaluate the
> > gain/loss.
> > > > But another statistician is insisting on using a repeated measures
> > ANOVA. I
> > > > am really not sure as to whether repeated measures ANOVA can be used
> > as we
> > > > are collecting data at different time intervals.
> > > > Request you people to guide me in this regard.
>
> > > > Thanks and regards
> > > > Thejasvi
>
> > > If the variability in the time intervals is not great, repeated
> > > measures ANOVA will still give a reasonable model, I think. If there
> > > is too much variability in the intervals, take a look at individual
> > > growth curve analysis.
>
> > > --
> > > Bruce Weaver

> > > bwea...@lakeheadu.ca

[Adrian Sayers]

There is an interesting section in "missing data in Clinical Studies"
by Molenberghs and Kenward about how the choosing the covariance
structure has implications on the power of the analysis.

Its Chapter 5.6 Comparative power under different Covariance structures.

Its in the context of missing data, which you could assume to apply to
this problem if you consider that every time you collected data you
should have collected data on all subjects, but you didn't, you could
assume the observations which you didnt collect data on to be missing.

This chapters worth a read as your sample is small and a more
parsimonious covariance matrix could be more powerful.

Adrian