Skip to content
Prev 7843 / 20628 Next

Conflicting p-values from pvals.fnc

Hi all,

This example has gotten me pretty confused about how the "weights" argument works for lmer.  I'd always assumed that setting a weight of k for an observation would cause lmer to act as if it had seen k replicates of that observation (i.e., the contribution of the observation to the likelihood would be taken to the k-th power).  After reading this query I found the following post that they are "precision weights not sampling weights":

http://tolstoy.newcastle.edu.au/R/e17/help/12/01/2099.html

I'm not sure what that means -- was my interpretation one of "sampling weights"?

Regardless, I'm noticing what seems to me to be inconsistent behavior in how setting the weights argument affects the t statistic in lmer() output and how the output of pvals.fnc() is affected.  I'm including an example below with a fixed effect of "x" and a random intercept of "a": the higher one sets the weights, the larger the t statistic for x becomes (which is what I'd originally expected given my assumptions about the weights semantics), but the broader the posterior on x becomes in the MCMC output.  This doesn't seem right, does it?  (I also don't understand why the estimate of the random-intercept variance but not the residual variance reported in the lmer output changes.)
Linear mixed model fit by REML 
Formula: y ~ x + (1 | a) 
   Data: dat 
   AIC   BIC logLik deviance REMLdev
 32.32 32.64 -12.16    29.97   24.32
Random effects:
 Groups   Name        Variance Std.Dev.
 a        (Intercept) 44.09718 6.64057 
 Residual              0.87763 0.93682 
Number of obs: 8, groups: a, 2

Fixed effects:
            Estimate Std. Error t value
(Intercept)   5.0792     4.7186   1.076
x            10.1045     0.6624  15.254

Correlation of Fixed Effects:
  (Intr)
x -0.070
$fixed
            Estimate MCMCmean HPD95lower HPD95upper  pMCMC Pr(>|t|)
(Intercept)    5.079    5.037     -1.553      11.32 0.1004   0.3231
x             10.104   10.110      4.917      15.28 0.0038   0.0000

$random
    Groups        Name Std.Dev. MCMCmedian MCMCmean HPD95lower HPD95upper
1        a (Intercept)   6.6406     2.6325   3.3750     0.0000     7.6330
2 Residual               0.9368     2.8649   3.2042     0.8809     6.0746
Linear mixed model fit by REML 
Formula: y ~ x + (1 | a) 
   Data: dat 
   AIC   BIC logLik deviance REMLdev
 69.17 69.48 -30.58    66.81   61.17
Random effects:
 Groups   Name        Variance Std.Dev.
 a        (Intercept) 0.44097  0.66406 
 Residual             0.87763  0.93682 
Number of obs: 8, groups: a, 2

Fixed effects:
            Estimate Std. Error t value
(Intercept)  5.07921    0.47185   10.76
x           10.10449    0.06624  152.54

Correlation of Fixed Effects:
  (Intr)
x -0.070
$fixed
            Estimate MCMCmean HPD95lower HPD95upper  pMCMC Pr(>|t|)
(Intercept)    5.079     5.38    -23.816      38.45 0.3434        0
x             10.104    10.11      4.115      16.06 0.0068        0

$random
    Groups        Name Std.Dev. MCMCmedian MCMCmean HPD95lower HPD95upper
1        a (Intercept)   0.6641     5.8216  15.7201     0.0000    61.7422
2 Residual               0.9368    23.8808  30.8847     5.2868    71.7694
Linear mixed model fit by REML 
Formula: y ~ x + (1 | a) 
   Data: dat 
    AIC    BIC logLik deviance REMLdev
 -4.516 -4.199  6.258   -6.868  -12.52
Random effects:
 Groups   Name        Variance   Std.Dev.
 a        (Intercept) 4409.71731 66.40570
 Residual                0.87763  0.93682
Number of obs: 8, groups: a, 2

Fixed effects:
            Estimate Std. Error t value
(Intercept)    5.079     47.187   0.108
x             10.104      6.624   1.525

Correlation of Fixed Effects:
  (Intr)
x -0.070
$fixed
            Estimate MCMCmean HPD95lower HPD95upper pMCMC Pr(>|t|)
(Intercept)    5.079    5.053    -0.8904      11.28 0.085   0.9178
x             10.104   10.107     5.3060      15.21 0.002   0.1780

$random
    Groups        Name Std.Dev. MCMCmedian MCMCmean HPD95lower HPD95upper
1        a (Intercept)  66.4057     2.6379   3.1098     1.1580     6.5406
2 Residual               0.9368     0.2828   0.3121     0.1214     0.5694


I'd be glad to be enlightened...!

Best

Roger
On Mar 26, 2012, at 10:35 AM PDT, Geoff Brookshire wrote:

            
--

Roger Levy                      Email: rlevy at ucsd.edu
Assistant Professor             Phone: 858-534-7219
Department of Linguistics       Fax:   858-534-4789
UC San Diego                    Web:   http://idiom.ucsd.edu/~rlevy