[R-sig-ME] lmer() vs. lme() gave different variance component estimates

[Peter Dalgaard]

On 09/18/2010 11:24 AM, Jarrod Hadfield wrote:
> Fedora Core 13 seems fine:
> 
> Linear mixed model fit by REML
> Formula: score ~ trt + (1 | id/eye)
>     Data: dat
>     AIC   BIC logLik deviance REMLdev
>   425.2 474.2 -201.6    412.7   403.2
> Random effects:
>   Groups   Name        Variance Std.Dev.
>   eye:id   (Intercept) 3.59532  1.89613
>   id       (Intercept) 3.51024  1.87356
>   Residual             0.01875  0.13693
> Number of obs: 640, groups: eye:id, 160; id, 80

Not in 32bit:

> lmer(score~trt+(1|id/eye),dat)
Linear mixed model fit by REML
Formula: score ~ trt + (1 | id/eye)
   Data: dat
   AIC   BIC logLik deviance REMLdev
 446.7 495.8 -212.4    430.9   424.7
Random effects:
 Groups   Name        Variance   Std.Dev.
 eye:id   (Intercept) 6.9208e+00 2.6307e+00
 id       (Intercept) 4.0996e-12 2.0248e-06
 Residual             1.8750e-02 1.3693e-01
Number of obs: 640, groups: eye:id, 160; id, 80

Diddling the start values can change the result, though. Could you all try

(a) setting verbose=T
(b) try start=list(matrix(10),matrix(10))
(c) same thing, but with .81 and .57


In the (c) case, I get

> l <- lmer(score~trt+(1|id/eye),dat,
start=list(matrix(.81),matrix(.57)), verbose=T)
  0:     2226.5319: 0.810000 0.570000
  1:     1304.7507:  2.42844  1.75398
  2:     776.66839:  6.30805  0.00000
  3:     657.56008:  7.73875  0.00000
  4:     514.89378:  10.7783  0.00000
  5:     460.61371:  13.2310  0.00000
  6:     434.88972:  15.6779 1.52026e-08
  7:     426.68822:  17.5387 6.57868e-08
  8:     424.89428:  18.6843 1.41900e-07
  9:     424.71823:  19.1258 2.47916e-07
 10:     424.71356:  19.2075 3.75766e-07
 11:     424.71354:  19.2122 5.37851e-07
 12:     424.71354:  19.2122 7.55161e-07

which suggests an algorithm error where the parameters are getting stuck
on the boundary. These values match the automatically generated starting
values only to two digits, so I'm somewhat baffled that you apparently
get something completely different in 64 bits!

-- 
Peter Dalgaard
Center for Statistics, Copenhagen Business School
Phone: (+45)38153501
Email: pd.mes at cbs.dk  Priv: PDalgd at gmail.com