[R-sig-ME] Interpreting lmer() interactions with Helmert contrasts
Becky Gilbert
beckyannegilbert at gmail.com
Fri Aug 21 14:46:49 CEST 2015
Hi Paul,
Thanks very much for the suggestion! I tried using lsmeans() to get the
pairwise comparisons as you suggested, and the results are below.
I'm a little confused by the results because the pairwise comparison tests
all show p > .05, but the WordType x Time interaction was significant when
tested via model comparisons...? I think this might be due to the Tukey
adjustment for multiple comparisons, but I'm not sure. Specifically the
contrast for the two levels of Time at WordType = 2 looks like it might
have been significant before the multiple comparisons correction, thus
accounting for the significance of the interaction term in model
comparisons. Any thoughts?
Thanks again!
Becky
$lsmeans
WordType = 0:
Time lsmean SE df lower.CL upper.CL
-1 2.880592 0.02209390 21.58 2.834721 2.926464
1 2.887315 0.02144245 22.13 2.842860 2.931769
WordType = 1:
Time lsmean SE df lower.CL upper.CL
-1 2.856211 0.02156603 19.78 2.811193 2.901229
1 2.888640 0.02089339 20.17 2.845080 2.932200
WordType = 2:
Time lsmean SE df lower.CL upper.CL
-1 2.852485 0.02181905 20.72 2.807072 2.897898
1 2.893827 0.02113775 21.12 2.849883 2.937770
Confidence level used: 0.95
$contrasts
WordType = 0:
contrast estimate SE df t.ratio p.value
-1 - 1 -0.00672255 0.02078469 19.31 -0.323 0.7498
WordType = 1:
contrast estimate SE df t.ratio p.value
-1 - 1 -0.03242907 0.02097452 20.02 -1.546 0.1377
WordType = 2:
contrast estimate SE df t.ratio p.value
-1 - 1 -0.04134141 0.02146707 21.93 -1.926 0.0672
____________________________________________
Dr Becky Gilbert
On 21 August 2015 at 12:19, paul debes <paul.debes at utu.fi> wrote:
> Hi Becky,
>
> Maybe you are interested in pairwise comparisons? The "lsmeans" package
> comes in handy.
>
> Try something like this:
>
> library("pbkrtest") # gives you KW-adjusted denDF for tests, but must be
> installed
> library("lsmeans")
>
> Model.lmer.means = lsmeans(Model, spec = pairwise ~ WordType|Time)
> Model.lmer.means = summary(Model.lmer.means)
> Model.lmer.means
>
> Maybe you want the contrast conditional on WordType, not Time? Swap it to:
> "spec = pairwise ~ Time|WordType"
>
> Best,
> Paul
>
>
> On Fri, 21 Aug 2015 14:04:07 +0300, Becky Gilbert <
> beckyannegilbert at gmail.com> wrote:
>
> Dear list,
>>
>> I'm wondering if someone could help me interpret an interaction between
>> two
>> factors, when one of the factors uses Helmert contrasts?
>>
>> I ran a linear mixed effects model (lmer) with reaction times as the DV, 2
>> fixed factors: Time (2 levels) and Word Type (3 levels), and 2 random
>> factors: Subjects and Items. I used Helmert contrasts for the Word Type
>> factor:
>> - Contrast 1 = level 1 (Untrained) vs levels 2 & 3 (Trained-related and
>> Trained-unrelated)
>> - Contrast 2 = level 2 vs. level 3 (Trained-related vs Trained-unrelated)
>> The data, contrasts, model, summary and model comparisons are listed at
>> the
>> end of the message.
>>
>> Model comparisons with anova() showed a significant interaction between
>> Time and Word Type. However, I don't know how to get the statistics for
>> the interactions between Time and each Word Type contrast.
>>
>> Based on the t-values for coefficients in the model summary, it looks like
>> the significant Word Type x Time interaction is driven by the interaction
>> with the 1st contrast for Word Type (t = 2.61). However I don't think
>> that
>> the statistics for the fixed effects coefficients are exactly what I'm
>> looking forward (they are sequential tests, right?). And if these are the
>> appropriate statistics, I'm aware of the problems with trying to get
>> p-values from these estimates. So is there a way to do likelihood ratio
>> tests for each Word Type contrast, or some other way of interpreting the
>> Word Type x Time interaction?
>>
>> Data structure:
>>
>>> str(rtData)
>>>
>> 'data.frame': 1244 obs. of 11 variables:
>> $ Subject : Factor w/ 16 levels "AB","AS","AW",..: 1 1 1 1 1 1 1 1 1
>> 1 ...
>> $ Item : Factor w/ 48 levels "ANT","BANDAGE",..: 3 4 6 12 13 14 22
>> 29 30 34 ...
>> $ Response : int 960 1255 651 1043 671 643 743 695 965 589 ...
>> $ Time : Factor w/ 2 levels "-1","1": 1 1 1 1 1 1 1 1 1 1 ...
>> $ WordType : Factor w/ 3 levels "0","1","2": 1 1 1 1 1 1 1 1 1 1 ...
>> $ logRT : num 2.98 3.1 2.81 3.02 2.83 ...
>>
>> contrasts(rtData$Time)
>>>
>> [,1]
>> -1 0.5
>> 1 -0.5
>>
>> contrasts(rtData$WordType)
>>>
>> [,1] [,2]
>> 0 0.6666667 0.0
>> 1 -0.3333333 -0.5
>> 2 -0.3333333 0.5
>>
>> Model:
>> lmer(logRT ~ 1 + WordType + Time + WordType:Time +
>> (1 + Time|Subject) +
>> (1|Item),
>> data = rtData)
>>
>> REML criterion at convergence: -2061.2
>> Scaled residuals:
>> Min 1Q Median 3Q Max
>> -2.7228 -0.6588 -0.0872 0.5712 3.7790
>> Random effects:
>> Groups Name Variance Std.Dev. Corr
>> Item (Intercept) 0.000933 0.03054
>> Subject (Intercept) 0.004590 0.06775
>> Time1 0.005591 0.07478 0.05
>> Residual 0.009575 0.09785
>> Number of obs: 1244, groups: Target, 46; Subject, 16
>> Fixed effects:
>> Estimate Std. Error t value
>> (Intercept) 2.8765116 0.0177527 162.03
>> WordType1 0.0111628 0.0110852 1.01
>> WordType2 0.0007306 0.0071519 0.10
>> Time1 -0.0268310 0.0195248 -1.37
>> WordType1:Time1 0.0301627 0.0115349 2.61
>> WordType2:Time1 -0.0089123 0.0141624 -0.63
>>
>> Model comparisons with anova() for main effects and interaction:
>>
>> -full model vs no Word Type x Time interaction
>> Df AIC BIC logLik deviance Chisq
>> Chi Df Pr(>Chisq)
>> rtModelNoInteraction 9 -2077.5 -2031.3 1047.7 -2095.5
>>
>> rtModelFull 11 -2080.5 -2024.1 1051.2 -2102.5 7.0388 2
>> 0.02962 *
>>
>> -full model vs model without Time and interaction
>> Df AIC BIC logLik deviance Chisq Chi Df
>> Pr(>Chisq)
>> rtModelNoTime 8 -2077.8 -2036.7 1046.9 -2093.8
>> rtModelFull 11 -2080.5 -2024.1 1051.2 -2102.5 8.7424 3
>> 0.03292 *
>>
>> -full model vs model without Word Type and interaction
>> Df AIC BIC logLik deviance Chisq Chi Df
>> Pr(>Chisq)
>> rtModelNoWT 7 -2080.4 -2044.5 1047.2 -2094.4
>> rtModelFull 11 -2080.5 -2024.1 1051.2 -2102.5 8.0875 4
>> 0.08842
>> .
>>
>> Thanks in advance for any advice!
>> Becky
>> ____________________________________________
>>
>> Dr Becky Gilbert
>>
>> [[alternative HTML version deleted]]
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>
>
>
> --
> Paul Debes
> DFG Research Fellow
> University of Turku
> Department of Biology
> Itäinen Pitkäkatu 4
> 20520 Turku
> Finland
>
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models
mailing list