[R-sig-ME] Testing if a variance component is zero?

Spencer Graves spencer.graves at pdf.com
Wed Dec 10 23:49:34 CET 2008

Hello All: 

      'anova' in 'lme4' to test a difference a zero variance component 
uses 'ML', at least in the following example: 

fm1 <- lmer( score ~ Machine+(1|Worker), data = Machines )
fm2 <- lmer( score ~ Machine+(1|Worker/Machine), data = Machines )
anova( fm1, fm2 )
   fm1: score ~ Machine + (1 | Worker)
   fm2: score ~ Machine + (1 | Worker/Machine)
       Df     AIC     BIC  logLik  Chisq Chi Df Pr(>Chisq)   
   fm1  5  303.75  313.70 -146.88                            
   fm2  6  237.47  249.40 -112.73 68.289      1  < 2.2e-16 ***

fm1m <- lmer( score ~ Machine+(1|Worker), data = Machines, REML=FALSE )
   'log Lik.' -146.8516 (df=5)

      This surprised me, because I would have expected REML to be used 
in this case, at least according to my reading of Pinheiro and Bates 
(2000, Figure 2.4), which got very good agreement between simulation 
results and cutting in half the nominal REML-based p-values. 
      Has more recent work suggested that Pinheiro and Bates (2000, 
Figure 2.4) is too specialized to worry about -- and this kind of 
'anova' has largely been obsoleted by "mcmcsamp"? 

      Spencer Graves
R version 2.8.0 (2008-10-20)

LC_COLLATE=English_United States.1252;LC_CTYPE=English_United 
States.1252;LC_NUMERIC=C;LC_TIME=English_United States.1252

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base    

other attached packages:
[1] MEMSS_0.3-4        lme4_0.999375-27   Matrix_0.999375-16 

loaded via a namespace (and not attached):
[1] grid_2.8.0

More information about the R-sig-mixed-models mailing list