[R-sig-ME] MCMCglmm interpretation with contrast coding and multinomial IVs and DV

Adriana Guevara Rukoz adriana.guevara.rukoz at ens.fr
Tue Apr 26 18:19:45 CEST 2016


Dear all,

I would like to analyze some data from an identification task, in which 
subjects hear a V1h[V2]pV1 stimulus (e.g. ah[i]pa), where [V2] 
designates the coarticulation within the /hp/ cluster (V1 != V2 when the 
stimuli have been spliced). Subjects are asked what vowel they perceived 
between the consonants of the cluster (vowel epenthesis).

In particular, I'm interested in the relative contributions of V1 and V2 
on the choice of each response vowel (e.g., do subjects choose /i/ more 
often than other vowels when V1=/i/? Even more or less so when 
COART=[i]? Which contribution is more important between V1 and V2 for 
each vowel?)

First, here is a recap of my variables (baseline = u ; as that is the 
"default" epenthetic vowel):

** RV: *
- RESP (categorical, 5 levels): {u, o, i, e, a}
** IV: *
- V1 (categorical, 5 levels): {u, o, i, e, a} - sum-contrast coded
- V2, henceforth named COART for easier reading, (categorical, 5 
levels): {u, o, i, e, a} - sum-contrast coded
** Random effects: *subject id (SUBJ), item id (ITEM).

Which gives me the following model:

k <- length(levels(h.epenth$RESP))
I <- diag(k-1)
J <- matrix(rep(1, (k-1)^2), c(k-1, k-1))

prior3 = list(R = list(fix=1, V=(1/k) * (I + J), n = k),
               G = list(G1=list(V = diag(k-1), n = k),
                        G1=list(V = diag(k-1), n = k)))

mh1 <- MCMCglmm(RESP ~  - 1 + trait + V1:trait + COART:trait,
                random = ~us(trait):SUBJ + us(trait):ITEM,
                rcov = ~ us(trait):units,
                prior = prior3,
                burnin = nburnin,
                nitt = nnitt,
                thin = nthin,
                family="categorical",
                data=epenth)

I'm currently running the model on pilot data and with 20000 iterations 
only, so the actual values of the estimated coefficients are not 
important, but here is the output from summary(mh1), for location effects:

Location effects: RESP ~ -1 + trait + V1:trait + COART:trait

                    post.mean  l-95% CI  u-95% CI eff.samp pMCMC
traitRESP.o         -3.59168  -6.51885  -0.40005   340.00 0.03529 *
traitRESP.i         -6.97617 -10.20436  -3.63351   340.00 < 0.003 **
traitRESP.e         -7.58867 -10.16525  -4.44600    17.06 < 0.003 **
traitRESP.a         -6.32136  -8.75020  -3.92369   170.22 < 0.003 **
----
traitRESP.o:V1o      0.96720  -1.91595   4.68645    57.00 0.57059
traitRESP.i:V1o      0.34601  -3.13529   3.65788   275.27 0.81176
traitRESP.e:V1o      0.01064  -2.31169   2.48395   179.78 0.94706
traitRESP.a:V1o     -0.08993  -2.52314   2.14315   134.33 0.97059
traitRESP.o:V1i     -0.57546  -3.82799   2.97239    18.85 0.81765
traitRESP.i:V1i      5.34557   2.23693   8.74984   398.06 0.01176 *
traitRESP.e:V1i      2.43926   0.20358   5.15729    68.81 0.07059 .
traitRESP.a:V1i      1.65012  -1.21129   3.84189   177.83 0.17647
traitRESP.o:V1e     -1.16602  -4.16428   2.01458    86.38 0.46471
traitRESP.i:V1e      4.68686   1.38014   7.48422   202.13 < 0.003 **
traitRESP.e:V1e      5.03065   2.33411   7.25455    17.88 < 0.003 **
traitRESP.a:V1e      0.55748  -2.02347   3.63707     4.68 0.74706
traitRESP.o:V1a     -0.49801  -4.05488   2.82815   340.00 0.78824
traitRESP.i:V1a      1.20223  -2.48158   4.27902   252.94 0.53529
traitRESP.e:V1a      1.41483  -1.13591   3.77316   254.96 0.28235
traitRESP.a:V1a      3.04426   0.49008   5.59015   340.00 0.01765 *
----
traitRESP.o:COARTo  -0.30663  -3.86220   2.94796   287.64 0.88824
traitRESP.i:COARTo   0.92669  -2.28283   4.60833   340.00 0.56471
traitRESP.e:COARTo  -0.05533  -2.61372   2.40955   140.88 0.96471
traitRESP.a:COARTo   0.13161  -2.46333   2.22642   340.00 0.92353
traitRESP.o:COARTi  -1.37804  -5.35041   1.66590    51.21 0.42353
traitRESP.i:COARTi   5.54896   1.53047   8.69056   340.00 0.00588 **
traitRESP.e:COARTi   2.37792  -0.33337   4.73054   340.00 0.08824 .
traitRESP.a:COARTi   0.59577  -1.90546   2.99576   175.13 0.61176
traitRESP.o:COARTe  -1.90106  -5.67091   1.52474    33.76 0.33529
traitRESP.i:COARTe   5.57507   2.15612   8.66226   340.00 0.00588 **
traitRESP.e:COARTe   2.44821  -0.28758   4.69727   340.00 0.05882 .
traitRESP.a:COARTe   0.53713  -1.72573   2.66476   340.00 0.64118
traitRESP.o:COARTa  -1.16952  -4.53311   2.54438    45.44 0.52353
traitRESP.i:COARTa   0.96946  -3.23285   4.17157   206.69 0.61765
traitRESP.e:COARTa   0.76110  -1.74846   3.68853   340.00 0.54118
traitRESP.a:COARTa   1.98751  -0.73148   4.65783    29.15 0.15294

1) Knowing that I sum-contrast coded the independent variables V1 and 
COART, is it correct for me to interpret
                                post.mean l-95% CI  u-95% CI eff.samp   
pMCMC
/traitRESP.i:COARTe   5.57507 2.15612   8.66226   340.00 0.00588 **
/as: "there is a significant increase in "i" responses relative to "u" 
responses when the coarticulation in the cluster (i.e. COART) is [e], 
for an 'average' V1" (as opposed to only comparing the change in  "i" vs 
"u"  responses when going from uh[u]pu to uh[e]pu, since /u/ is the 
baseline for V1 and COART)

2) Following up on this, would contrast coding the response variable 
RESP (if there is any sense in doing that, model-wise) allow me to 
change the statement above to:  "there is a significant increase in "i" 
responses relative to any other response vowel when the coarticulation 
in the cluster (i.e. COART) is [e], for an 'average' V1", or does the 
trait always take "u" as a baseline?

3) Also, I guess I am a confused about how to interpret the significance 
codes. Basically, for
                    post.mean  l-95% CI  u-95% CI eff.samp   pMCMC
traitRESP.o         -3.59168  -6.51885  -0.40005   340.00 0.03529 *
traitRESP.i         -6.97617 -10.20436  -3.63351   340.00 < 0.003 **
traitRESP.e         -7.58867 -10.16525  -4.44600    17.06 < 0.003 **
traitRESP.a         -6.32136  -8.75020  -3.92369   170.22 < 0.003 **
I understand this as: subjects responded "u" significantly more often 
than "o", "i", "e" and "a", in general (because of contrast coding). Is 
this correct?
If so, are these significance tests "independent" from the significance 
testing for V1:trait and for COART:trait?

I hope that my explanations are not too unclear...

Thank you in advance for your help!

Best,
Adriana

-- 
Adriana Guevara Rukoz

PhD Student
Laboratoire de Sciences Cognitives et Psycholinguistique
École Normale Supérieure
29 rue d’Ulm
75005 Paris, France


	[[alternative HTML version deleted]]



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