Hello R-help subscribers,
I am analyzing a data set using a mixed logit model, and I have recently
discovered some curious behavior. I am hoping you all can help.
I first ran the following model in December 2012.
lmer(Response.binary ~ ItemType.c * Block + (1 | Subject) + (1 | Word),
data=lexdec, family="binomial")
I then took a break from the data for the holidays. I returned to the data
yesterday and discovered that running the exact same model on the exact
same data set yields different output. The overall patterns are the same,
but the coefficients, variance estimates, and model fits (AIC, BIC) differ.
The model outputs from the old and current attempt are appended below.
I have triple checked the code and the data set to ensure that what I'm
working with now is the same as in December. Having found no differences, I
can only suspect that some function has changed. During my hiatus from
these data, I updated plyr and its dependencies (and maybe some other
packages). But to my understanding, these updates mostly concerned
documentation, not algorithms. Any ideas then about why the model outputs
differ?
Thank you for your help!
Kodi
*
OLD MODEL OUTPUT*:
Generalized linear mixed model fit by the Laplace approximation
Formula: Response.binary ~ ItemType.c * Block + (1 + ItemType.c + Block |
Subject) + (1 | Word)
Data: lexdec
AIC BIC logLik deviance
4788 4957 -2370 4740
Random effects:
Groups Name Variance Std.Dev. Corr
Word (Intercept) 1.66447 1.29014
Subject (Intercept) 0.50865 0.71320
ItemType.cFV-L 0.89270 0.94483 0.261
ItemType.cFV-R 1.26385 1.12421 0.210 0.978
ItemType.cFV-B 1.33556 1.15566 0.143 0.916 0.979
Blockpost 0.93878 0.96891 -0.349 -0.093 0.037 0.163
Number of obs: 8500, groups: Word, 298; Subject, 17
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.8233 0.2688 14.225 < 2e-16 ***
ItemType.cFV-L -6.0547 0.3749 -16.149 < 2e-16 ***
ItemType.cFV-R -6.8649 0.4130 -16.621 < 2e-16 ***
ItemType.cFV-B -7.3542 0.4285 -17.164 < 2e-16 ***
Blockpost 0.9754 0.3238 3.013 0.00259 **
ItemType.cFV-L:Blockpost 0.5921 0.2725 2.173 0.02980 *
ItemType.cFV-R:Blockpost 0.5835 0.2926 1.994 0.04612 *
ItemType.cFV-B:Blockpost -0.2718 0.3083 -0.882 0.37793
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
*NEW MODEL OUTPUT:
*Generalized linear mixed model fit by the Laplace approximation
Formula: Response.binary ~ ItemType.c * Block + (1 + ItemType.c + Block |
Subject) + (1 | Word)
Data: lexdec
AIC BIC logLik deviance
4791 4961 -2372 4743
Random effects:
Groups Name Variance Std.Dev. Corr
Word (Intercept) 1.57837 1.25633
Subject (Intercept) 0.58476 0.76470
ItemType.cFV-L 0.92922 0.96396 -0.105
ItemType.cFV-R 1.36398 1.16790 -0.241 0.990
ItemType.cFV-B 1.59667 1.26360 -0.323 0.956 0.978
Blockpost 1.03413 1.01692 -0.511 0.198 0.264 0.406
Number of obs: 8500, groups: Word, 298; Subject, 17
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.3659 0.2704 12.448 < 2e-16 ***
ItemType.cFV-L -5.6366 0.3710 -15.193 < 2e-16 ***
ItemType.cFV-R -6.3466 0.4128 -15.376 < 2e-16 ***
ItemType.cFV-B -6.6767 0.4372 -15.271 < 2e-16 ***
Blockpost 1.1657 0.3288 3.546 0.000391 ***
ItemType.cFV-L:Blockpost 0.5119 0.2681 1.909 0.056243 .
ItemType.cFV-R:Blockpost 0.4834 0.2859 1.691 0.090886 .
ItemType.cFV-B:Blockpost -0.4394 0.3002 -1.463 0.143336
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 *
*
--
Kodi Weatherholtz
Ph.D. Student
Department of Linguistics
The Ohio State University
[[alternative HTML version deleted]]
