| Author |
Message |
< ASReml ~ Fixed v random effects in GLMM
|
| andrew |
 Posted: Tue Feb 28, 2012 2:25 pm |
 |
|
|
Joined: 11 Aug 2011
Posts: 8
Location: Denmark, WA
|
I'm trying to use GLMM with binomial distribution to analyse the trait seedcount (which is actually a proportion, not a count).
My data are:
femclone !A maleclone1 !A ramet_age !I flowerload !/100 health !A sizeclass !A #4 size classes
First I fitted all significant effects as fixed effects:
!filter sizeclass !select 4 #restrict data to the proportion of seed in largest size class
seedcount !binomial !AOD ~ mu ramet_age health flowerload maleclone1 femclone
This resulted in the following AOD Table:
Analysis of Deviance Table for seedcount
Source of Variation df Deviance Derived F
ramet_age 3 3.21 11.527
health 4 2.12 5.711
flowerload 1 1.82 19.610
maleclone1 6 19.82 35.536
femclone 18 8.25 4.929
Deviance from GLM fit 122 11.34
Variance heterogeneity factor [Deviance/DF] 0.09
... which showed that femclone and maleclone1 were both significant effects.
Next, I tried fitting femclone and maleclone1 as random effects, while holding the other genetic effect as a fixed effect along with the other fixed effects listed above. In each case, the genetic effect that I fitted as random had no associated variance, even though both effects were significant when fitted as fixed. When I fit both femclone and maleclone1 as random, maleclone1 has a large variance but there is none at femclone (see attachment).
I haven't decided for myself whether these effects are better modeled as fixed or random (There is little meaningful genetic structure here, so not interested in heritability, repeatability calculalations etc)... but I do want to understand what is happening with the random and fixed parts of the model. What is going on here?
Thanks in advance, Andrew. |
|
|
| Back to top |
 
|
| andrew |
| Posted: Tue Feb 28, 2012 2:29 pm |
|
|
|
Joined: 11 Aug 2011
Posts: 8
Location: Denmark, WA
|
Sorry - I couldn't get the .asr file to attach. The pertinent results are copied below:
Univariate analysis of seedcount
Summary of 155 records retained of 620 read
Notice: !FILTER sizeclass !SELECT 4 is applied after any data transformation
Model term Size #miss #zero MinNon0 Mean MaxNon0 StndDevn
1 femclone 19 0 0 1 7.9355 19
2 maleclone1 7 0 0 1 3.0000 7
3 ramet_age 4 0 0 1 1.8968 4
4 paclo 2 0 0 1 1.6903 2
5 flowerload 0 0 0.1200 5.751 37.95 7.669
6 health 5 0 0 1 2.4581 5
7 sizeclass 4 0 0 4 4.0000 4
8 seedcount Variate 0 0 0.5000E-02 0.4725 0.9910 0.2601
9 total 0 0 1.000 1.000 1.000 0.000
10 mu 1
Forming 37 equations: 11 dense.
Initial updates will be shrunk by factor 0.010
Restarting iteration from previous solution
Distribution and link: Binomial; Logit Mu=P=1/(1+exp(-XB))
V=Mu(1-Mu)/N
Warning: The LogL value is unsuitable for comparing GLM models
Notice: 2 singularities detected in design matrix.
1 LogL=-140.534 S2= 1.0000 146 df : 1 components restrained
2 LogL=-140.520 S2= 1.0000 146 df Dev/DF= 0.1442
3 LogL=-140.520 S2= 1.0000 146 df Dev/DF= 0.1442
4 LogL=-140.520 S2= 1.0000 146 df Dev/DF= 0.1442
5 LogL=-140.520 S2= 1.0000 146 df Dev/DF= 0.1442
Final parameter values 0.69985 0.74212E-07 1.0000
Deviance from GLM fit 146 21.05
Variance heterogeneity factor [Deviance/DF] 0.14
- - - Results from analysis of seedcount - - -
Notice: While convergence of the LogL value indicates that the model
has stabilized, its value CANNOT be used to formally test differences
between Generalized Linear (Mixed) Models.
Source Model terms Gamma Component Comp/SE % C
maleclone1 7 7 0.699852 0.699852 1.07 0 P
femclone 19 19 0.742117E-07 0.742117E-07 0.00 0 B
Variance 155 146 1.00000 1.00000 0.00 0 F |
|
|
| Back to top |
|
| Arthur |
| Posted: Sat Mar 03, 2012 1:33 am |
|
|
|
Joined: 05 Aug 2008
Posts: 277
Location: Orange, NSW
|
Dear Andrew,
I can't see that you are highlighting a problem with ASReml, but rather
are seeking help with your analysis. However, this really is not
an ideal place to review your analysis unless you want to upload your data.
My concern is that you are analysing proportions without properly
accounting for the cell sizes (which must be around 100-200 given the
range of proportions (.005 to .991). Consequently, you have a very low
heterogeneity factor.
Given the average is 0.47 for the trait, the counts are evidently large,
I would begin with just doing a Normal analysis of the data.
I would be very suspiscious of any binomial analysis that did not appear similar to
the normal analysis.
You have not exlained the relationship between femclone and maleclone.
They are evidently not nested, but are they fully frossed?
It is possible that
it really is just maleclone that is significant.
My next step would be to look at the mean values in the femclone x maleclone
table, and compare them with adjusted means (from PREDICT famclone maleclone) |
_________________
Arthur Gilmour
Retired Principal Research Scientist (Biometrics) |
|
| Back to top |
 
|
|
|
![[vsni.co.uk]](/common/images/address.gif){kind=small}
Contact us | Site map
News Feed
