From: Bruce Southey <southey_at_UIUC.EDU>

Date: Fri, 27 Oct 2006 08:42:23 -0500

Date: Fri, 27 Oct 2006 08:42:23 -0500

Hi,

Basically two main issues need to addressed: Poisson assumption and relationship between size and time. A lesser issue is the number of missing values as you indicated that just over half of the observations were missing (631 missing values out of 1053).

1) What is happening with 'size' across time?

You are assuming size has a constant effect over time. Usually most thing grow nonlinearly with time so you may need to allow size to vary with time or use a different approach (multivariate or a ratio).

2) What is the distribution of these missing values?

Really you should, at least initially, only consider individuals with as close as possible to having all 13 measurements. Given that you really need at least 3 points to fit a straight line, actually more to test it, you should remove any individual with only 2 or even 3 measurements.

3) What is the reason why not all measurements were recorded?

The missing values should be missing at random otherwise you need to address the non-randomness. Perhaps you need to remove later measurements.

4) Are the Poisson assumptions holding?

These are probably over-rated but could be an issue. Just because the data is a count it does not mean that a Poisson distribution is appropriate. Really the question is how appropriate is the Poisson assumption - see question 6.

5) Why do you use !DISP?

This implies that the Poisson distribution is not appropriate as it would not be fitted in the final model. I would be very concerned if the data exhibits underdispersion and if the data exhibits overdispersion then a negative binomial should be used if the Poisson model is appropriate.

6) What is the range of values for 'call'?

The normal distribution becomes very similar to the Poisson once the mean is over 3. The comment in your first code indicates a maximum value of 59 so I would be looking at assuming normality using some sort of transformation like square root or log if required to stabilize variance or address skewness.

Regards

Bruce

---- Original message ----

*>Date: Fri, 27 Oct 2006 16:23:56 +1000
*

*>From: Felix Zajitschek - BEES - UNSW <felix.zajitschek_at_STUDENT.UNSW.EDU.AU>
*

*>Subject: Re: unbalanced repeated measures
*

*>To: ASREML-L_at_AGRIC.NSW.GOV.AU
*

*>
*

*> Link: File-List
*

*>
*

*> Dear ASRemlers,
*

*>
*

*>
*

*>
*

*> After having posted a question of how to best analyze my longitudinal
*

*> data with missing values, Arthur suggested I should use a random
*

*> regression approach with fitting cubic splines to the longitudinal
*

*> data of individuals. The working code is below.
*

*>
*

*> The result suggests that age at measurement (=callage) is not
*

*> significant, but size of the animal (=size) is highly significant.
*

*>
*

*> In order to fully understand this, I tried to visualize it, but when I
*

*> plot eg. the residuals out of the model without size, against size, I
*

*> can't see any sort of pattern.
*

*>
*

*>
*

*>
*

*> My first question is whether I'm missing something fundamentally.
*

*>
*

*> I'm also not quite sure whether I need to fit cubic splines, as my
*

*> ultimate question is only to see whether my dv (=call) shows a
*

*> decrease with increasing age.
*

*>
*

*>
*

*>
*

*> As before, I'm thankful for any comments,
*

*>
*

*>
*

*>
*

*> Cheers,
*

*>
*

*> Felix
*

*>
*

*>
*

*>
*

*>
*

*>
*

*>
*

*>
*

*>
*

*>
*

*> Working code (measured variable is call=calling effort, at 13 ages
*

*> (callage) throughout an individuals life):
*

*>
*

*> Random regression analysis
*

*>
*

*> ind !I
*

*>
*

*> eclosed
*

*>
*

*> size
*

*>
*

*> temp
*

*>
*

*> callage
*

*>
*

*> call
*

*>
*

*> males-readyFORasremlreduced.txt !SKIP1 !DISPLAY 9 !MAXIT 200
*

*>
*

*> call !POISSON !DISP ~ mu callage size !r spl(callage) fac(callage)
*

*> ind callage.ind !f mv
*

*>
*

*> 1 2 1
*

*>
*

*> 81
*

*>
*

*> 13 0 AR1 0.1 !GP
*

*>
*

*> ind 2
*

*>
*

*> ind
*

*>
*

*> 2 0 US !GP
*

*>
*

*> .1 0.01 .1
*

*>
*

*> ind
*

*>
*

*>
*

*>
*

*> ______________________________________________________________
*

*>
*

*> ::Felix Zajitschek
*

*> ::School of Biological, Earth and Environmental Sciences (BEES)
*

*> ::University of New South Wales (UNSW) NSW 2052 - Australia
*

*> ::Labtel +61 (0)2 9385 2124
*

*> ::Fax +61 (0)2 9385 1558
*

*> ::eMail felix.zajitschek_at_student.unsw.edu.au
*

*>
*

*> ::www.bees.unsw.edu.au/school/researchstudents/zajitschekfelix.html
*

*>
*

*> ::Lab Blog
*

*>
*

*>
*

Received on Wed Oct 27 2006 - 08:42:23 EST

This webpage is part of the ASReml-l discussion list archives 2004-2010. More information on ASReml can be found at the VSN website. This discussion list is now deprecated - please use the VSN forum for discussion on ASReml. (These online archives were generated using the hypermail package.)