Re: df with random effects
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: df with random effects

> Hello:
> In the example Ex11a, we have 3 lines and 9 sires,(actually the sires are 
> in the lines). There is 65 observations and 74 animals in total. The addition 
> an animal effect or a sire effects does not change the degrees of freedom of 
> error (62), which are 65-2-1. The test for lines will be with 2, 62 df with 
> number of other random effects?.
> If we use Mixed of SAS with a sire model we got 56 df for the denominator
> for testing
> lines, which comes from deducting 2 df for lines, 9-3 = 6 df for sires/lines
> and 1 df for mu. All other things are identical.
> Could someone explain the difference in df's?.

For testing for line, one can argue for 6, 56, 62 or some intermediate value
as the appropriate error degrees of

The value of 56  is obtained by treating sires as fixed in ASREML.
But if sirtes are then regarded as random, nested within lines, lines
should be tested against the 'sire' variance rather than against the
'residual' variance.

In general, it is not easy to work out the proper denominator
degrees of freedom for any test of fixed effects in a mixed model.

The EX11A example is sufficient to demonstrate the problem.

If we fit  AD ~ mu Line sire  we get the analysis of variance

  SOurce    df     MS      F
  Line       2    2227    16.81 [against Error], 6.8 against Sire
  SIre       6     327.2   2.47
  Error     56     132.5
  If we equate the sire MS to its expection, we get a sire variance component 
  of about   (327.2 - 132.5)/7 = 27.8
  From the mixed model fitting AD ~ mu Line !r sire

  SOurce    df     MS      F
  Line       2      ?    6.43   [ which agrees with the test against Sire above]
  Error     62     132.4
   The variance component for sires in this model is 27.2; similar to the ANOVA 
   estimate.  It is similar but not exact because the data is not fully
   balanced and 7 is only approximately the average progeny per sire [65/9].

This is most obvious in a Split plot analysis where some components
would be tested against an 'Error A' and others against an 'Error B'.

I believe Kenward and ROger (1997) discuss this problem [Biometrics 
53: 983-997].

I thought I had a discussion of this in the manual but apparantly not.

Essentially, only in a few well defined cases can we work out
precisely what the error degrees of freedom should be.  However,
we see in the example that ASREML does give the appropriate F statistic
[even if we do not know its proper distribution].

Some programs just quote a Wald statistic distributed as Chi-square but this 
assumes the error variance is known [infinite df].  I therefore prefer
the F test where I can usually get some indication [ie use my knowledge
of the structure of the data to quess] of what the error
df should be.

I hope this helps.
> Thank you
> Hugo

Arthur Gilmour PhD                    email:
Senior Research Scientist (Biometrics)                 fax: <61> 2 6391 3899
NSW Agriculture                                             <61> 2 6391 3922
Orange Agricultural Institute               telephone work: <61> 2 6391 3815
Forest Rd, ORANGE, 2800, AUSTRALIA                    home: <61> 2 6362 0046

ASREML is currently free by anonymous ftp from pub/aar on
    Point your web browser at 
    in the IACR-Rothamsted information system 

To join the asreml discussion list, send the message  

The address for messages to the list is

                        <> <> <> <> <> <> <>
"Seek first the kingdom of God and His righteousness"  Jesus; Matthew 6: 33