Hi Andrew,
As replies to emails have been a bit sporadic lately I'll attempt to answer
your questions. There are more knowledgeable folk out there though...
2 is easiest I think and sort of leads into 3. I get the impression that
the answer to 1 is a bit more personal taste.
2) Depending on what exactly you want to do, yeah, its probably ok. What
youre doing is testing the parameter/estimate of the related and unrelated
levels separately against the nonword level. So youre not comparing the 3
means (i.e. the answer to 3 - it doesnt). You might want to read up a
little on contrasts perhaps as theres various options. Your professor
perhaps wants an F-Test and post-hoc (Tukey or whatever), in which case you
could use the LMERConvenienceFunctions:::pamer.fnc function (this provides
conservative and anticonservative p-values for the effect as a whole,
TarCond; there are other possibilities for this too - see the MixMod
package for instance) or fit a model with and without the IV and do a
likelihood ratio test which would provide you with something similar.
HTH
Alan
--------------------------------------------------
Email: aghaynes@gmail.com
Mobile: +41794385586
Skype: aghaynes
On 13 September 2012 03:24, Obermeier Andrew wrote:
> Hello Mixed Modelers,
>
> I came to this list with a question a couple months ago, and will admit it
> was a very humbling experience. There were a few kind responses, but my
> question was mostly ignored or I was told I needed to read basic statistics
> books. I take full blame for this, and know that I was not asking my
> question in the right way. Now, 2 months later, I am trying again.
>
> I am not a statistician but want to learn. Especially, I like learning
> statistics with R because it helps me to see the nuts and bolts of
> statistics much more than the big commercial user interface packages do,
> and I get a much better feel for how to look at data.
>
> I am writing a dissertation in second language acquisition, analyzing my
> data using lme4 for a mixed model with subjects and items set as random
> effects. My dependent variable is reaction time (RT). My independent
> variable is Target Condition (TarCond). TarCond is a three-level factor:
> nonword, related, unrelated.
>
> When I run the model, lmer provides the summary below. The Fixed effects
> section tells me that for the three factors of my independent variable I
> get the following mean reaction times (RT), and all of these have big t
> values:
> nonword (Intercept): 1239.64
> TarCondrelated: 1239.64 - 370.72 = 868.92
> TarCondunrelated: 1239.64 + 318.34 = 1557.98
>
> (lmer summary)
> -----------------------------------------------------------------
> Linear mixed model fit by REML
> Formula: RT ~ TarCond + (1 | Subject) + (1 | Item)
> Data: postest
> AIC BIC logLik deviance REMLdev
> 30949 30982 -15468 30966 30937
> Random effects:
> Groups Name Variance Std.Dev.
> Item (Intercept) 22371 149.57
> Subject (Intercept) 65149 255.24
> Residual 560665 748.78
> Number of obs: 1920, groups: Item, 96; Subject, 20
>
> Fixed effects:
> Estimate Std. Error t value
> (Intercept) 1239.64 65.61 18.894
> TarCondrelated -370.72 56.13 -6.605
> TarCondunrelated 318.34 56.13 5.672
>
> Correlation of Fixed Effects:
> (Intr) TrCndr
> TarCondrltd -0.285
> TarCndnrltd -0.285 0.333
> ----------------------------------------------------------------------
>
>
> Since the t values in the above model were all well above 2 or below -2, I
> read that this might mean I have a significant effect.
>
> I followed the analysis in Baayen, Davidson, and Bates' (2007) article
> titled "Mixed-effects modeling with crossed random effects for subjects and
> items." Though this article informs me that p-values calculated with the
> degrees of freedom used by pvals.fnc will be "anti-conservative", I
> proceeded to analyze the fixed effects in this model with Markov chain
> Monte Carlo sampling.
>
>
> -------------------------------------------------------------------------------------------------
> > mcmcpostest <- pvals.fnc(postest.lmer, nsim = 10000)
> > mcmcpostest$fixed
> Estimate MCMCmean HPD95lower HPD95upper pMCMC Pr(>|t|)
> (Intercept) 1239.6 1239.5 1108.9 1365.0 0.0001 0
> TarCondrelated -370.7 -371.2 -474.6 -263.8 0.0001 0
> TarCondunrelated 318.3 317.8 214.3 428.8 0.0001 0
>
> -------------------------------------------------------------------------------------------------
>
> My apologies for this long preamble, but my questions are as follows.
> (1) Is this analytical procedure appropriate?
> (2) Specifically, is it OK for me to test the significance of a 3-level
> factor in this way? (My advising professor tells me that the t value can
> only be used to compare 2 means.)
> (3) How does lme4 calculate t values for a factor that is greater than 2
> levels.
>
> Below I have attached my dataset in CSV format. If there is any other
> information I need to provide, please let me know and I will very gladly do
> so.
>
> My apologies if my questions are misguided or inappropriate. I am way out
> of my league here.
>
> I would be very grateful for any attention that anyone can offer to my
> questions.
>
> Sincerely,
>
> Andrew Obermeier
> Associate Professor, Kyoto University of Education
> PhD Candidate, Temple University Japan
>
>
>
>
>
>
> _______________________________________________
> R-sig-mixed-models@r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>
[[alternative HTML version deleted]]