Author 
Message 
← ASReml ~ Problem with fitting common environment effect in models

g5614200126 
Posted: Sat Jan 06, 2018 11:56 am 


Joined: 10 Oct 2017
Posts: 3

Dear ASReml users,
I’m trying to estimate variance components of survival trait using threshold model (for binary response) and linear model (for day of death). At first, my models consisted of Weight covariate (Weight), fixed effect of tank (Tank) and two random effects of animal’s additive genetic effect (RenAnimal), the result seem fine. But, when I tried to add common fullsib effect (Family) in the model.
ASReml showed errors which were 1. “LogLnot converged” for threshold model and 2. “Code Bfixed at boundary” on common fullsib line for linear model.
1. For threshold model, I tried to use !CONTINUE and ran the job several times but I still can’t fixed this. Is there a way to solve this problem, “LogLnot converged”?
# as file for threshold model
RenAnimal !P
RenDam !P
RenSire !P
Family !A
Weight
Tank !A
Batch 2
Survival
DayofDeath
D:\ASREML\ASREML_Binary.ped !MAKE !SKIP 1 !REPEAT
D:\ASREML\ASREML_Binary.dat !SKIP 1 !MAXIT 500 !ASUV !CONTINUE
Survival !BIN !PROBIT ~ mu Weight Tank !r RenAnimal Family
2. For Linear model, I got very small component of common fullsib effect (0.452250E09). Is it alright to interpret that I can ignore and exclude this effect from the model because of very small value? Or I need to solve code B first?
#as file for linear model
RenAnimal !P
RenDam !P
RenSire !P
Family !A
Weight
Tank !A
Batch 2
Survival
DayofDeath
D:\ASREML\ASREML_DD.ped !MAKE !SKIP 1 !REPEAT
D:\ASREML\ASREML_DD.dat !SKIP 1 !MAXIT 500 !ASUV !CONTINUE
DayofDeath ~ mu Weight Tank !r RenAnimal Family
For the description of my data, I have data of 3,544 fishes from 128 fullsib families with 35 parternal halfsib families, each family consist of 30 individuals. I assume that there might be common fullsib effect since each family was reared separately before tagging.
I have attached .asr files for threshold model “ASREML_Binary” and linear model “ASREML_DD”.
Any suggestion would be helpful and I’m really appreciate.
Sincerely,
Sila 
Last edited by g5614200126 on Tue Jan 09, 2018 12:42 pm; edited 1 time in total
Description: 

Download 
Filename: 
ASREML_DD.asr 
Filesize: 
3.75 KB 
Downloaded: 
247 Time(s) 
Description: 

Download 
Filename: 
ASREML_Binary.asr 
Filesize: 
44.1 KB 
Downloaded: 
255 Time(s) 


Back to top 

Arthur 
Posted: Sun Jan 07, 2018 11:44 pm 


Joined: 05 Aug 2008
Posts: 471
Location: Orange, NSW

Dear Sita,
First I note
WARNING: 16 missing values were detected in the design variables
Missing values are treated as zeros
Consider deleting the records in which they occur
To my mind it is nonsensical to use weight as a covariated with values of Zero
for missing weights. But you must choose what to do.
You could delete these records, or substitue a more likely value.
A more likely value is 0.077 if Weight is not related to survival.
If all those with missing weight did not survive, maybe the average weight of those that did
not survive might be more sensible.
2)
An animal model is generally not suited to GLMM modelling. Many simulation studies have shown
that the estimated value is not well related to the value used to simulate the data.
But the results are very dependent on the data structure.
3) In your case, it would appear that you have 35 sires and 128 dams (nested in sires)
so I would recommend fitting Sire and Dam=family rather than Animal and Family.
This may be easier to interpret.
4) The normal analysis failed to converge; it appears to oscillate between two local maxima.
I suggest a Sire+Family model will work better in this case as well.
The two models of analysis should be consistent, but in your case they are not. 
_________________ Arthur Gilmour
Retired Principal Research Scientist (Biometrics) 

Back to top 

g5614200126 
Posted: Tue Jan 09, 2018 12:50 pm 


Joined: 10 Oct 2017
Posts: 3

Dear Dr. Gilmour,
Thank you very much for your kind response. I really appreciate for your help.
I have followed your instruction, I replaced missing value with average weight and used Sire+Family model instead of animal model for both threshold and linear model. The analysis went well as you suggested.
#Threshold model
Survival !BIN !PROBIT ~ mu Weight Tank !r Sire Family
#Linear model
DayofDeath ~ mu Weight Tank !r Sire Family
However, there are still something that I don’t understand and need your advice.
1. In point 3), you also suggested me to fit Sire and Dam=family. So, I tried to overlay Sire and Dam and ignored pedigree. But, I couldn’t run the program. It showed “Error reading model factor list”. I suspected that the problem must be the functions I used (bold). What should I do to fit Sire and Dam=family?
The model are:
Survival or DayofDeath ~ mu Weight Tank Batch !r RenSire ide(RenSire and(RenDam))
2. Could you please explain why sire model work better in this case? Is siredam model also valid? Someone told me that they used siredam model for survival trait.
3. My data contains pedigree of only experiment fishes and its parents (one generation). Is it sufficient to use sire model?
Any suggestion would be really appreciate.
Best regards,
Sila 


Back to top 

Arthur 
Posted: Wed Jan 10, 2018 9:54 pm 


Joined: 05 Aug 2008
Posts: 471
Location: Orange, NSW

Dear Sita,
When I said 'fit Sire and Dam=family.' I meant
!r Sire Family
not
!r RenSire ide(RenSire and(RenDam))
The syntax is not correct in the last line.
For a normal trait, you should compare the models
DayofDeath ~ mu Weight Tank !r Sire Family
DayofDeath ~ mu Weight Tank !r RenAnim Family
DayofDeath ~ mu Weight Tank !r RenAnim
DayofDeath ~ mu Weight Tank !r RenSire and(RenDam)
DayofDeath ~ mu Weight Tank !r RenAnim ide(RenSire) and(RenDam) # This may be singular
I understand you have no depth of pedigree, and that dams are nested in sires,
so the first two models will have the same LogL but quite different components.
For a binary trait, models 2, 3 and 5 fail because the observations 0,1 on the probit/logit scale
map to +/ infinity and you are fitting an effect and a residual to every point. While the model will often run,
the results tend to be arbitrary. However, with Sire + Dam model, the sire and dam means tend not to be 0,1 but
some intermediate value, which maps to a real number on the probit/logit scale and so the effect can be estimated,
and the residual (on the working scale) is also estimable.
If you understand these 5 models on the normal scale, it should help you see why
Sire + Family is reasonable on the underlying scale for binomial data 
_________________ Arthur Gilmour
Retired Principal Research Scientist (Biometrics) 

Back to top 


All times are GMT

You cannot post new topics in this forum You cannot reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum You can attach files in this forum You can download files in this forum

