rm ANOVA in SPSS.

[Asger Lund]

Hi all.
I am trying to get a grip on all my data from my recent clinical
study. I have a few questions regarding rm ANOVA and the use of this.

I have conducted a trial with 10 patients. each patients have been
examined on 6 separate days (A,B,C,D,E and F) 60 days Total. during
these days the patients recieved 6 different hormones.

I have gathered bloodsamples every 5minutes (the first hour) every
15minuts (the secound hour) and every 30min (the last hour).

This gives me some fine curves of plasma glucose for each of the 6
days. I would like to know If these curve excursions are different or
not.

So is it posible to do this using rm ANOVA?. And if so, could I do a
massive rm ANOVA for all six days, and with post-hoc tests see to
witch timepoints my curves (days) differ?

-Is there a problem in using rm ANOVA when the timepoints are not
equally spread? (meaning I did not draw blood every 5min the hole
period) (but I did draw blood in the same frekvens at all the days I
am comparing) (I will only be looking at the time*treatment effect)?

Hope someone have some thoughts on this.

-thanks
Asger, Copenhagen

[Jeremy Miles]

I think this would be much easier as a multilevel model in SPSS.

Have a look at Andy Field's SPSS book - the 3rd edition has a new
chapter on multilevel (mixed) models in SPSS.

Jeremy

2009/12/17 Asger Lund <lund....@gmail.com>:

> --
> 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

--
Jeremy Miles
Psychology Research Methods Wiki: www.researchmethodsinpsychology.com

[Adrian Sayers]

I think you might be better looking at the mlwin software from
cmm.bristol.ac.uk. The software is specifically designed to do this,
and the models fit quickly, and also as you have small numbers you can
make sure your se's are correct by using RIGLS, instead of IGLS.

also if you have asymptotes, you can fit fractional polynomials as
well. For a good intro to this have a look a patrick roystons papers.

bw
ADrian

2009/12/18 Jeremy Miles <jeremy...@gmail.com>:

[BXC (Bendix Carstensen)]

My advice would be to fit a smooth parametric curve for each hormone, a function of the time, using splines or fractional polynomials. You would then need to take the serial correlation into account, presumably specifying the correlation between measurements as some function of the time between them. Failing to do so may give you seriously biased results. And that is where the ANOVA would fail.

On top of this you would presumably have to have a fixed or random effect of person, and maybe also random effects of time and time^2 (by person) to account for the variation between persons.

In formula terms:

y_pht = f_h(t) + a_p + b_p t + c_p t^2 + e_pth

Where f is some parametrically specified function of t, different for each hormone, and with (a_b, b_p, c_p) 3-dim normal with mean 0 and e_pth normal with corr(e_pth, e_psh) = g(t-s) whith g some decresing function.

Once that is sown up, you have fixed effects of time for each of the hormones, and estimates of your variance components.

You can then plot the estimates of the curves f_h with confidence intervals, and in order to compare them, the corresponding differences between the curves. These are all just linear functions of the parameters in the splines or fractional polynomials, albeit not simple functions.

This sort of model can in principle be fitted by any of the major statistcal packages (although I would not bet too much on SPSS due to my ignorance). The complicated thing is to extract the estimated mean curves and get sensible plots of them and their differences, and particularly their confidence intervals. You would invariably need some kind of matrix manipulation for this (or write 50 "estimate" statements in SAS proc mixed). So by that token you would be best off using R, but on the other hand the mixed models fitter in R, lme - which is very fast and efficient, has an almost impenetrable syntax.

Best regrads,
Bendix Carstensen
_______________________________________________

Bendix Carstensen
Senior Statistician
Steno Diabetes Center
Niels Steensens Vej 2-4
DK-2820 Gentofte
Denmark
+45 44 43 87 38 (direct)
+45 30 75 87 38 (mobile)
b...@steno.dk http://www.biostat.ku.dk/~bxc
www.steno.dk

[mdim...@gmail.com]

I agree with Adrian, I find it easy to fit multilevel models in MLWin as it gives you background equations and its ability to use RIGLS. Its becoming a specialised software on this subject although you can replicate same random effects models in stata or sas

Regards

Munya
Sent using BlackBerry® from Orange

[Kylie Lange]

Hi Asger,

In addition to all the suggestions so far I would suggest starting off by
summarising the glucose curves with some meaningful parameters, such as AUC,
peak, time to peak, average over a certain time period etc, depending on the
nature of the excursions. You might find a summary approach like this useful
first, before delving into the more complex questions of time-specific
differences between treatments.

Kylie.

Asger Lund <lund....@gmail.com> Dec 17 12:40PM -0800