[R-lang] Re: Investigating random slope variance

Titus von der Malsburg malsburg@posteo.de
Thu Apr 3 11:31:10 PDT 2014


Hi Roger!

On 2014-04-03 Thu 18:32, Levy, Roger <rlevy@ucsd.edu> wrote:
> My interpretation would be as follows: in aggregate, there is ample
> evidence in your dataset that there is variation across items in the
> effect of condition, but you don’t have enough data on any individual
> item to conclude securely that *that particular item’s sensitivity* is
> significantly different from the group average in one direction or the
> other.

Hm, that's what I was afraid of.

>> - What's the proper way to find out which regions were significantly
>>  slowed down and which were speeded up by the manipulation?
>
> Sorry, did you mean regions or items? I’m assuming below that you
> meant items...

Regions and items are the same thing here because the participants read
only one text and the regions within the text act like the items in a
more typical reading experiment.

> It looks like you are using treatment coding for item, which renders
> the interpretation of your coefficients for this model a bit different
> than those in your dot plot.  How do things look when you use sum
> coding for item?

Good catch!  I intended to use contr.sum but due to a typo I ended up
using the default contrast.  The summary for the model using the sum
contrast is at the bottom of this mail.  Obviously, the results look
different from the results that I got when using the treatment contrast
but they are still inconsistent with the dotplot.  In the dotplot, I
find a positive slope for item 25 and a borderline significant negative
slope for item 36.  In the model using item as a fixed effect, I find
significant negative effects for items 11 and 36 and positive effects
for items 35 and 23.

It seems that the fixed-effects model is more consistent with the
descriptive stats:

  > with(d, tapply(trt, list(item, cond), mean, na.rm=T))
             A         B
  1  1165.3636 1128.5652
  2   992.5455 1144.6087
  3   602.1818  583.0909
  4   613.9048  719.3913
  5   599.8182  646.8696
  6   406.9048  489.2174
  7   620.0000  589.0435
  8   644.5000  763.8696
  9   576.3182  631.8696
  10  596.3182  600.7826
  11  806.8182  660.3913 * signf. in fixef-mode
  12  442.9524  552.4783
  13 1084.0000 1008.9130
  14  994.4091  878.1739
  15  898.4545  797.3913
  16 1037.9545 1113.6087
  17 1186.4545 1162.0435
  18  608.6818  786.2174
  19  582.6818  647.2727
  20  617.4545  618.2609
  21  434.7727  642.8095
  22 1179.8182 1031.2609
  23  528.2727  721.2609 * signf. in fixef-mode
  24  571.5455  600.5909
  25  319.6190  386.0435 * signf. in dotplot
  26  851.6364  713.3913
  27 1528.5909 1486.6957
  28  720.3182  603.8261
  29  726.9091  773.9565
  30  381.8095  452.8636
  31  846.6818  976.2273
  32  634.2273  878.5652
  33  740.1818  748.4348
  34  713.7727  879.3913
  35  720.8182 1052.8696 * signf. in fixef-mode
  36 1216.5909  921.2174 * signf. in fixef-mode and dotplot
  37  594.8636  588.9565
  38  459.5909  624.8261
  39  690.1818  885.2727
  40  449.6818  628.0870

> One last thing — I would recommend that you double-check all your
> analyses using lme4.0.  People have been reporting odd and
> contradictory results with the newest version of lme4, especially when
> using the default optimizer.

I reran the models using the bobyqa and optimx (method="nlminb") and got the same results.

  Titus


Summary of model using sum-coded items as fixed effect:

  Linear mixed model fit by maximum likelihood  ['lmerMod']
  Formula: log(trt) ~ item * cond + (1 | subj)
     Data: d
  Control: lmerControl(optimizer = "bobyqa")
  
       AIC      BIC   logLik deviance df.resid 
    2311.5   2761.5  -1073.7   2147.5     1705 
  
  Scaled residuals: 
      Min      1Q  Median      3Q     Max 
  -3.8821 -0.6375 -0.0172  0.6402  3.5181 
  
  Random effects:
   Groups   Name        Variance Std.Dev.
   subj     (Intercept) 0.07732  0.2781  
   Residual             0.18107  0.4255  
  Number of obs: 1787, groups: subj, 45
  
  Fixed effects:
                 Estimate Std. Error t value
  (Intercept)    6.446746   0.042668  151.09
  item1          0.481033   0.062657    7.68
  item2          0.381353   0.062657    6.09
  item3          0.214172   0.063347   -3.38
  item4          0.065494   0.063414   -1.03
  item5          0.164766   0.062657   -2.63
  item6          0.489277   0.063414   -7.72
  item7          0.103389   0.062657   -1.65
  item8          0.007107   0.062657    0.11
  item9          0.151968   0.062657   -2.43
  item10         0.187189   0.062657   -2.99
  item11         0.022083   0.062657    0.35
  item12         0.342824   0.063414   -5.41
  item13         0.421248   0.062657    6.72
  item14         0.265824   0.062657    4.24
  item15         0.095553   0.062657    1.53
  item16         0.428998   0.062657    6.85
  item17         0.508100   0.062657    8.11
  item18         0.104577   0.062657   -1.67
  item19         0.189842   0.063348   -3.00
  item20         0.176217   0.062657   -2.81
  item21         0.346756   0.064095   -5.41
  item22         0.464940   0.062657    7.42
  item23         0.211371   0.062657   -3.37
  item24         0.239727   0.063348   -3.78
  item25         0.702424   0.063414  -11.08
  item26         0.074705   0.062657    1.19
  item27         0.778773   0.062657   12.43
  item28         0.065349   0.062657   -1.04
  item29         0.076321   0.062657    1.22
  item30         0.529148   0.064096   -8.26
  item31         0.253587   0.063347    4.00
  item32         0.028338   0.062657   -0.45
  item33         0.062202   0.062657    0.99
  item34         0.110623   0.062657    1.77
  item35         0.205651   0.062657    3.28
  item36         0.411933   0.062657    6.57
  item37         0.237561   0.062657   -3.79
  item38         0.293674   0.062657   -4.69
  item39         0.104801   0.063348    1.65
  condB-A         0.062645   0.085337    0.73
  item1:condB-A  -0.099294   0.125314   -0.79
  item2:condB-A   0.116658   0.125314    0.93
  item3:condB-A  -0.124696   0.126694   -0.98
  item4:condB-A   0.169267   0.126827    1.33
  item5:condB-A   0.010062   0.125314    0.08
  item6:condB-A   0.029288   0.126827    0.23
  item7:condB-A  -0.159953   0.125314   -1.28
  item8:condB-A   0.082437   0.125314    0.66
  item9:condB-A  -0.088934   0.125314   -0.71
  item10:condB-A -0.031815   0.125314   -0.25
  item11:condB-A -0.315189   0.125314   -2.52
  item12:condB-A  0.196759   0.126827    1.55
  item13:condB-A -0.138070   0.125314   -1.10
  item14:condB-A -0.193227   0.125314   -1.54
  item15:condB-A -0.188834   0.125314   -1.51
  item16:condB-A -0.023788   0.125314   -0.19
  item17:condB-A -0.052790   0.125314   -0.42
  item18:condB-A  0.037642   0.125314    0.30
  item19:condB-A -0.011895   0.126697   -0.09
  item20:condB-A -0.135061   0.125314   -1.08
  item21:condB-A  0.211311   0.128191    1.65
  item22:condB-A -0.184232   0.125314   -1.47
  item23:condB-A  0.319074   0.125314    2.55
  item24:condB-A -0.018217   0.126697   -0.14
  item25:condB-A  0.216689   0.126827    1.71
  item26:condB-A -0.192187   0.125314   -1.53
  item27:condB-A -0.073881   0.125314   -0.59
  item28:condB-A -0.147357   0.125314   -1.18
  item29:condB-A -0.055930   0.125314   -0.45
  item30:condB-A  0.091677   0.128193    0.72
  item31:condB-A  0.055254   0.126694    0.44
  item32:condB-A  0.170249   0.125314    1.36
  item33:condB-A -0.021665   0.125314   -0.17
  item34:condB-A  0.075640   0.125314    0.60
  item35:condB-A  0.261821   0.125314    2.09
  item36:condB-A -0.316347   0.125314   -2.52
  item37:condB-A  0.001981   0.125314    0.02
  item38:condB-A  0.157791   0.125314    1.26
  item39:condB-A  0.105193   0.126697    0.83
  
  Correlation matrix not shown by default, as p = 80 > 20.
  Use print(...., correlation=TRUE)  or
	 vcov(....)	 if you need it



More information about the ling-r-lang-L mailing list