From: <damian.collins_at_INDUSTRY.NSW.GOV.AU>

Date: Fri, 29 Jan 2010 17:19:15 +1100

Dear Craig,

(NB. Arthur is away at present)

I think in the first model you mean "SOME MISSING MUM VALUES AND !MVINCLUDE SPECIFIED" ?

If so, then I think I know where the problem is.

From the V3 user guide http://www.vsni.co.uk/downloads/asreml/release3/UserGuide.pdf , p113 (p139)

"Where the factor level is zero (or missing and the !MVINCLUDE qualifier is specified), no level is assigned to the factor for that record. These effectively defines an extra level (class) in the factor which becomes a reference level."

In your first model, using !mvinclude, ASReml effectively imputes one extra "mum" (class) for *all* the missing values.

In your second model, you have imputed your own unique dummy codes, a separate one for each missing value.

So the first model fits one additional level of mum, the second model fits m additional levels of mum (where m here is the number of missing values). I think this explains the difference between the two outputs.

Hope this helps,

Damian

Damian Collins, Biometrician, I&I NSW

damian.collins@industry.nsw.gov.au

ph: 02 4640 6451

-----ASReml users discussion group <ASREML-L@DPI.NSW.GOV.AU> wrote: -----

Received on Fri Jan 29 2010 - 17:19:15 EST

Date: Fri, 29 Jan 2010 17:19:15 +1100

Dear Craig,

(NB. Arthur is away at present)

I think in the first model you mean "SOME MISSING MUM VALUES AND !MVINCLUDE SPECIFIED" ?

If so, then I think I know where the problem is.

From the V3 user guide http://www.vsni.co.uk/downloads/asreml/release3/UserGuide.pdf , p113 (p139)

"Where the factor level is zero (or missing and the !MVINCLUDE qualifier is specified), no level is assigned to the factor for that record. These effectively defines an extra level (class) in the factor which becomes a reference level."

In your first model, using !mvinclude, ASReml effectively imputes one extra "mum" (class) for *all* the missing values.

In your second model, you have imputed your own unique dummy codes, a separate one for each missing value.

So the first model fits one additional level of mum, the second model fits m additional levels of mum (where m here is the number of missing values). I think this explains the difference between the two outputs.

Hope this helps,

Damian

Damian Collins, Biometrician, I&I NSW

damian.collins@industry.nsw.gov.au

ph: 02 4640 6451

To: ASREML-L@DPI.NSW.GOV.AU

From: craig walling <asremlforum@VSNI.CO.UK>

Sent by: ASReml users discussion group <ASREML-L@DPI.NSW.GOV.AU>

Date: 01/29/2010 03:29AM

Subject: using !MVINCLUDE with missing random effects

Dear Forum,

I was wondering if someone could help me to understand what the !MVINCLUDE function is doing when I have missing effects in one of my random effect structures. I am attempting to run a model looking at between sex genetic correlations for a trait that combines survival and reproduction within a year. Females and males are both measured in multiple years, so this is a repeat measures model. The model is specified as follows:

femalesr malesr ~ Trait Trait.age !r Trait.id Trait.ide(id) Trait.byr Trait.mum

age is the age of the individual as a factor, id is a pedigree linked term, ide(id) accounts for the repeat measures per individual, byr is the year of birth of each individual and mum is the mother of each individual. The only missing data is in the mum term, where males sometimes have unknown mothers, but females never do. I thought that when specifying !MVINCLUDE, this would assign no level to mum when it was missing. In order to check this, I dummy coded unique mothers for all males with unknown mothers (each unique male gets the same unique mother code each year it is measured) and then re-ran the model, assuming that this would give identical results given that there is no information in these dummy mothers. However, despite giving similar outputs, the 2 models do not give identical outputs and the differences are enough to make me think I don't understand what !MVREMOVE is doing. The results from the 2 models are:

MODEL WITH SOME MISSING MUM VALUES AND !MVREMOVE SPECIFIED

13 LogL= 2112.07 S2= 1.0000 5726 df

Source Model terms Gamma Component Comp/SE % C

Residual UnStructured 1 1 0.978555E-01 0.978555E-01 42.35 0 P

Residual UnStructured 2 1 0.00000 0.00000 0.00 0 F

Residual UnStructured 2 2 0.486247 0.486247 26.55 0 P

Trait.id UnStructured 1 1 0.131282E-02 0.131282E-02 1.31 0 P

Trait.id UnStructured 2 1 -0.695530E-03 -0.695530E-03 -0.20 0 P

Trait.id UnStructured 2 2 0.117350E-01 0.117350E-01 0.60 0 P

Trait.ide(id) UnStructured 1 1 0.536386E-06 0.536386E-06 0.00 0 B

Trait.ide(id) UnStructured 2 1 0.00000 0.00000 0.00 0 F

Trait.ide(id) UnStructured 2 2 0.135719 0.135719 4.86 0 P

Trait.mum UnStructured 1 1 0.825148E-03 0.825148E-03 1.02 0 U

Trait.mum UnStructured 2 1 0.104853E-02 0.104853E-02 0.36 0 U

Trait.mum UnStructured 2 2 -0.127838E-01 -0.127838E-01 -1.07 0 U

Trait.byr UnStructured 1 1 0.216229E-03 0.216229E-03 0.69 0 U

Trait.byr UnStructured 2 1 0.114162E-02 0.114162E-02 1.17 0 U

Trait.byr UnStructured 2 2 0.103753E-02 0.103753E-02 0.18 0 U

Warning: Code B - fixed at a boundary (!GP) F - fixed by user

? - liable to change from P to B P - positive definite

C - Constrained by user (!VCC) U - unbounded

S - Singular Information matrix

S means there is no information in the data for this parameter.

Very small components with Comp/SE ratios of zero sometimes indicate poor

scaling. Consider rescaling the design matrix in such cases.

Covariance/Variance/Correlation Matrix UnStructured Residual

0.9786E-01 0.000

0.000 0.4862

Covariance/Variance/Correlation Matrix UnStructured Trait.id

0.1313E-02 -0.1772

-0.6955E-03 0.1173E-01

Covariance/Variance/Correlation Matrix UnStructured Trait.ide(id)

0.5364E-06 0.000

0.000 0.1357

Covariance/Variance/Correlation Matrix UnStructured Trait.mum

0.8251E-03 0.3228

0.1049E-02 -0.1278E-01

Covariance/Variance/Correlation Matrix UnStructured Trait.byr

0.2162E-03 2.410

0.1142E-02 0.1038E-02

Analysis of Variance NumDF F_inc

31 Trait 2 12331.88

32 Trait.age 21 24.73

36 Trait.byr 74 effects fitted ( 6 are zero)

37 Trait.mum 1230 effects fitted ( 466 are zero)

33 Trait.id 8102 effects fitted ( 1110 are zero)

35 Trait.ide(id) 8102 effects fitted ( 7073 are zero)

SLOPES FOR LOG(ABS(RES)) on LOG(PV) for Section 1

2.25 2.20

63 possible outliers: see .res file

Finished: 28 Jan 2010 15:15:01.413 LogL Converged

MODEL WITH MISSING MUMS VALUES REPLACED WITH UNIQUE DUMMY CODES:

13 LogL= 2111.55 S2= 1.0000 5726 df

Source Model terms Gamma Component Comp/SE % C

Residual UnStructured 1 1 0.978627E-01 0.978627E-01 42.35 0 P

Residual UnStructured 2 1 0.00000 0.00000 0.00 0 F

Residual UnStructured 2 2 0.486746 0.486746 26.55 0 P

Trait.id UnStructured 1 1 0.130749E-02 0.130749E-02 1.30 0 P

Trait.id UnStructured 2 1 -0.117752E-02 -0.117752E-02 -0.33 0 P

Trait.id UnStructured 2 2 0.126129E-01 0.126129E-01 0.62 0 P

Trait.ide(id) UnStructured 1 1 0.121802E-06 0.121802E-06 0.00 0 B

Trait.ide(id) UnStructured 2 1 0.00000 0.00000 0.00 0 F

Trait.ide(id) UnStructured 2 2 0.113950 0.113950 3.78 0 P

Trait.nwmum UnStructured 1 1 0.835708E-03 0.835708E-03 1.03 0 U

Trait.nwmum UnStructured 2 1 0.150108E-02 0.150108E-02 0.45 0 U

Trait.nwmum UnStructured 2 2 0.994648E-02 0.994648E-02 0.53 0 U

Trait.byr UnStructured 1 1 0.202625E-03 0.202625E-03 0.66 0 U

Trait.byr UnStructured 2 1 0.120962E-02 0.120962E-02 1.22 0 U

Trait.byr UnStructured 2 2 0.207076E-02 0.207076E-02 0.34 0 U

Warning: Code B - fixed at a boundary (!GP) F - fixed by user

? - liable to change from P to B P - positive definite

C - Constrained by user (!VCC) U - unbounded

S - Singular Information matrix

S means there is no information in the data for this parameter.

Very small components with Comp/SE ratios of zero sometimes indicate poor

scaling. Consider rescaling the design matrix in such cases.

Covariance/Variance/Correlation Matrix UnStructured Residual

0.9786E-01 0.000

0.000 0.4867

Covariance/Variance/Correlation Matrix UnStructured Trait.id

0.1307E-02 -0.2900

-0.1178E-02 0.1261E-01

Covariance/Variance/Correlation Matrix UnStructured Trait.ide(id)

0.1218E-06 0.000

0.000 0.1139

Covariance/Variance/Correlation Matrix UnStructured Trait.nwmum

0.8357E-03 0.5206

0.1501E-02 0.9946E-02

Covariance/Variance/Correlation Matrix UnStructured Trait.byr

0.2026E-03 1.867

0.1210E-02 0.2071E-02

Analysis of Variance NumDF F_inc

31 Trait 2 12449.92

32 Trait.age 21 24.70

36 Trait.byr 74 effects fitted ( 6 are zero)

37 Trait.nwmum 1700 effects fitted ( 708 are zero)

33 Trait.id 8102 effects fitted ( 1110 are zero)

35 Trait.ide(id) 8102 effects fitted ( 7073 are zero)

SLOPES FOR LOG(ABS(RES)) on LOG(PV) for Section 1

2.23 2.20

63 possible outliers: see .res file

Finished: 28 Jan 2010 15:48:51.563 LogL Converged

I realise that this model is not running particularly well given that the ide(id) term for females is bound at 0 and the mum term for males is negative in the first model, but can this be the reason the models aren't identical? Could anyone give me some advice on why these 2 models would differ?

Many thanks!

Craig

PS, appologies if I have not provided enough information here, please just ask if you need more to answer this question!

-------------------- m2f --------------------

Sent using Mail2Forum (http://www.mail2forum.com).

Read this topic online here:

http://www.vsni.co.uk/forum/viewtopic.php?p=1302#1302

-------------------- m2f --------------------

Received on Fri Jan 29 2010 - 17:19:15 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.)