From mlopez at iattc.org  Tue Sep 15 11:45:30 2009
From: mlopez at iattc.org (Milton Lopez)
Date: Tue, 15 Sep 2009 11:45:30 -0700
Subject: [R-lang] Moving from  Linux to Mac OS
Message-ID: <D1EEF18B0DF02445870ED9C6CEA595E901530C28@email1.lajolla.iattc.org>

One of our researchers is considering replacing her Red Hat Linux
workstation with a Mac Pro. One of her concerns is being able to build R
libraries on the Mac. As an R-illiterate sysadmin I suspect this would
no more difficult on the Mac, but we would like to make sure.

Also, some years ago I was told that R does not take advantage of
multiple processors. Has there been any improvement in that area? This
would have an impact on configuring the new system with more processors
vs. fewer but faster ones.

Thanks in advance for your comments.


Milton F. Lopez
Inter-American Tropical Tuna Commission (IATTC)
8604 La Jolla Shores Drive
La Jolla, CA 92037 - USA
mlopez at iattc.org
www.iattc.org




From ken.williams at thomsonreuters.com  Tue Sep 15 11:52:22 2009
From: ken.williams at thomsonreuters.com (Ken Williams)
Date: Tue, 15 Sep 2009 13:52:22 -0500
Subject: [R-lang] Moving from  Linux to Mac OS
In-Reply-To: <D1EEF18B0DF02445870ED9C6CEA595E901530C28@email1.lajolla.iattc.org>
Message-ID: <C6D54C96.246B3%ken.williams@thomsonreuters.com>

I use a Mac Pro as my primary workstation, and installing R packages is a
snap.  Just "R CMD INSTALL packagefoo.tar.gz" as usual, or use the stuff
inside the R gui.

Tell her that the R gui (R.app) is really excellent, btw.

Out of the box R doesn't do special stuff to use multiple processors, but
there are various existing ways to get it to, and Grand Central Dispatch
gives me optimism that even better multiproc support is around the corner.

 -Ken


On 9/15/09 1:45 PM, "Milton Lopez" <mlopez at iattc.org> wrote:

> One of our researchers is considering replacing her Red Hat Linux
> workstation with a Mac Pro. One of her concerns is being able to build R
> libraries on the Mac. As an R-illiterate sysadmin I suspect this would
> no more difficult on the Mac, but we would like to make sure.
> 
> Also, some years ago I was told that R does not take advantage of
> multiple processors. Has there been any improvement in that area? This
> would have an impact on configuring the new system with more processors
> vs. fewer but faster ones.
> 
> Thanks in advance for your comments.
> 
> 
> Milton F. Lopez
> Inter-American Tropical Tuna Commission (IATTC)
> 8604 La Jolla Shores Drive
> La Jolla, CA 92037 - USA
> mlopez at iattc.org
> www.iattc.org
> 
> 
> 
> _______________________________________________
> R-lang mailing list
> R-lang at ling.ucsd.edu
> http://pidgin.ucsd.edu/mailman/listinfo/r-lang

-- 
Ken Williams
Sr. Research Scientist
Thomson Reuters
Phone: 651-848-7712
ken.williams at thomsonreuters.com




From rachelbaker2010 at u.northwestern.edu  Tue Sep 15 21:16:54 2009
From: rachelbaker2010 at u.northwestern.edu (Rachel Baker)
Date: Tue, 15 Sep 2009 23:16:54 -0500
Subject: [R-lang] Contrast Coding in R Regressions
Message-ID: <b4200db30909152116x3cbd0bfaufccd737a5cd829b8@mail.gmail.com>

Hi,

I've recently started using R to do regressions, using the 'lmer' function.
I am currently re-running some analyses that originally had treatment
coding, so that they now have contrast coding.  My question is about how to
interpret contrast coded regression outputs.

One of my independent variables (nativeLanguage) has 3 levels: English,
Chinese, and Korean.  As this experiment was conducted in English,
participants in the English group were native speakers, and participants in
the other two groups were non-native speakers.  In my original
treatment-coded analysis, English was the reference level.  My output for
e.g. 'langCompare.lmer = lmer(duration~nativeLanguage+(1|Subject),
data=myData)' had lines like:

                                                 Estimate Std. Error t value
nativeLanguageChinese              0.025920   0.002384  10.872
nativeLanguageKorean                -0.004416   0.002091  -2.112

As I understood it, such lines gave information about the comparison between
Chinese and English, and between Korean and English, respectively.

I contrast coded this variable with the code:
'contrasts(myData$nativeLanguage) = c(-1, .5, .5)' (after ordering the
levels: English, Chinese, Korean).  This was in order to compare the native
(English) group to the non-native (Chinese and Korean) groups.  After this
contrast coding, my output had lines like:

                                        Estimate Std. Error t value
nativeLanguage1               0.10002   0.010113  11.242
nativeLanguage2
-0.00046   0.639887  1.388

I was wondering how to interpret this output.  My guess is that
nativeLanguage1 is the comparison between the native and non-native groups,
and native_language2 is the comparison between Chinese and Korean, but I
haven't been able to find any resources to confirm this.

I would greatly appreciate any advice on how to interpret regressions after
contrast coding, or pointers to appropriate resources on this topic!

Thanks very much,
Rachel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://pidgin.ucsd.edu/pipermail/r-lang/attachments/20090915/4dcd6094/attachment.htm>

From mlopez at iattc.org  Wed Sep 16 10:18:29 2009
From: mlopez at iattc.org (Milton Lopez)
Date: Wed, 16 Sep 2009 10:18:29 -0700
Subject: [R-lang] Moving from  Linux to Mac OS
In-Reply-To: <C6D54C96.246B3%ken.williams@thomsonreuters.com>
References: <D1EEF18B0DF02445870ED9C6CEA595E901530C28@email1.lajolla.iattc.org>
	<C6D54C96.246B3%ken.williams@thomsonreuters.com>
Message-ID: <D1EEF18B0DF02445870ED9C6CEA595E901530C7C@email1.lajolla.iattc.org>

Ken.

Thanks for the feedback. I am also curious to see how Apple's GCD will
find its way into R. FYI I got another reply from campus pointing to
http://www.revolution-computing.com/  I talked with a sales rep who told
me they will be looking closely at GCD.


Cheers


M.


-----Original Message-----
From: Ken Williams [mailto:ken.williams at thomsonreuters.com] 
Sent: Tuesday, September 15, 2009 11:52 AM
To: Milton Lopez; r-lang at ling.ucsd.edu
Cc: Cleridy Lennert
Subject: Re: [R-lang] Moving from Linux to Mac OS

I use a Mac Pro as my primary workstation, and installing R packages is
a
snap.  Just "R CMD INSTALL packagefoo.tar.gz" as usual, or use the stuff
inside the R gui.

Tell her that the R gui (R.app) is really excellent, btw.

Out of the box R doesn't do special stuff to use multiple processors,
but
there are various existing ways to get it to, and Grand Central Dispatch
gives me optimism that even better multiproc support is around the
corner.

 -Ken


On 9/15/09 1:45 PM, "Milton Lopez" <mlopez at iattc.org> wrote:

> One of our researchers is considering replacing her Red Hat Linux
> workstation with a Mac Pro. One of her concerns is being able to build
R
> libraries on the Mac. As an R-illiterate sysadmin I suspect this would
> no more difficult on the Mac, but we would like to make sure.
> 
> Also, some years ago I was told that R does not take advantage of
> multiple processors. Has there been any improvement in that area? This
> would have an impact on configuring the new system with more
processors
> vs. fewer but faster ones.
> 
> Thanks in advance for your comments.
> 
> 
> Milton F. Lopez
> Inter-American Tropical Tuna Commission (IATTC)
> 8604 La Jolla Shores Drive
> La Jolla, CA 92037 - USA
> mlopez at iattc.org
> www.iattc.org
> 
> 
> 
> _______________________________________________
> R-lang mailing list
> R-lang at ling.ucsd.edu
> http://pidgin.ucsd.edu/mailman/listinfo/r-lang

-- 
Ken Williams
Sr. Research Scientist
Thomson Reuters
Phone: 651-848-7712
ken.williams at thomsonreuters.com




From njsmith at cogsci.ucsd.edu  Wed Sep 16 15:08:45 2009
From: njsmith at cogsci.ucsd.edu (Nathaniel Smith)
Date: Wed, 16 Sep 2009 15:08:45 -0700
Subject: [R-lang] Moving from Linux to Mac OS
In-Reply-To: <D1EEF18B0DF02445870ED9C6CEA595E901530C28@email1.lajolla.iattc.org>
References: <Aco2NLXGvu2kiD2CS3OANzy0EZPHnw==>
	<D1EEF18B0DF02445870ED9C6CEA595E901530C28@email1.lajolla.iattc.org>
Message-ID: <961fa2b40909161508k7df643a4j959dde7e104384fb@mail.gmail.com>

On Tue, Sep 15, 2009 at 11:45 AM, Milton Lopez <mlopez at iattc.org> wrote:
> Also, some years ago I was told that R does not take advantage of
> multiple processors. Has there been any improvement in that area? This
> would have an impact on configuring the new system with more processors
> vs. fewer but faster ones.

That's correct, R cannot automatically take advantage of multiple
processors. There are more and more packages that will help you
more-or-less easily write code that can run in parallel over multiple
processors:
  http://cran.r-project.org/web/views/HighPerformanceComputing.html
But you still have to actually learn how to use the package, write
your code in a parallel way, etc., if you want to get the speedup.

-- Nathaniel


From rlevy at ling.ucsd.edu  Sat Sep 19 20:19:26 2009
From: rlevy at ling.ucsd.edu (Roger Levy)
Date: Sat, 19 Sep 2009 20:19:26 -0700
Subject: [R-lang] Contrast Coding in R Regressions
In-Reply-To: <b4200db30909152116x3cbd0bfaufccd737a5cd829b8@mail.gmail.com>
References: <b4200db30909152116x3cbd0bfaufccd737a5cd829b8@mail.gmail.com>
Message-ID: <2BC4A78A-C269-47FD-B531-FB3F826ED084@ling.ucsd.edu>


On Sep 15, 2009, at 9:16 PM, Rachel Baker wrote:

> Hi,
>
> I've recently started using R to do regressions, using the 'lmer'  
> function.  I am currently re-running some analyses that originally  
> had treatment coding, so that they now have contrast coding.  My  
> question is about how to interpret contrast coded regression outputs.
>
> One of my independent variables (nativeLanguage) has 3 levels:  
> English, Chinese, and Korean.  As this experiment was conducted in  
> English, participants in the English group were native speakers, and  
> participants in the other two groups were non-native speakers.  In  
> my original treatment-coded analysis, English was the reference  
> level.  My output for e.g. 'langCompare.lmer =  
> lmer(duration~nativeLanguage+(1|Subject), data=myData)' had lines  
> like:
>
>                                                  Estimate Std. Error  
> t value
> nativeLanguageChinese              0.025920   0.002384  10.872
> nativeLanguageKorean                -0.004416   0.002091  -2.112
>
> As I understood it, such lines gave information about the comparison  
> between Chinese and English, and between Korean and English,  
> respectively.
>
> I contrast coded this variable with the code: 'contrasts(myData 
> $nativeLanguage) = c(-1, .5, .5)' (after ordering the levels:  
> English, Chinese, Korean).  This was in order to compare the native  
> (English) group to the non-native (Chinese and Korean) groups.   
> After this contrast coding, my output had lines like:
>
>                                         Estimate Std. Error t value
> nativeLanguage1               0.10002   0.010113  11.242
> nativeLanguage2              -0.00046   0.639887  1.388
>
> I was wondering how to interpret this output.  My guess is that  
> nativeLanguage1 is the comparison between the native and non-native  
> groups, and native_language2 is the comparison between Chinese and  
> Korean, but I haven't been able to find any resources to confirm this.

Hi Rachel,

Your guess is correct, but the situation may be a little more  
complicated than you think.  First, you need to realize that you  
didn't specify a complete contrast.  Here's a little code snippet to  
illustrate:

 > m <- 20
 > n <- 3
 > lang <- factor(rep(c("English", "Chinese", "Korean"),m*n),  
levels=c("English", "Chinese", "Korean"))
 > old.contrasts <- contrasts(lang)
 > contrasts(lang) <- c(1,-.5,-.5)
 > new.contrasts <- contrasts(lang)

Now, let's take a look at the old and new contrast matrices:

 > old.contrasts
         Chinese Korean
English       0      0
Chinese       1      0
Korean        0      1
 > new.contrasts
         [,1]          [,2]
English  1.0 -5.551115e-17
Chinese -0.5 -7.071068e-01
Korean  -0.5  7.071068e-01

The value of old.contrasts derives from the fact that by default, R  
uses contr.treatment for unordered factors, with the first level of  
the factor being the baseline (which for you is English, so that the  
contrast matrix is all zeroes in the English row):

 > options()$contrasts
         unordered           ordered
"contr.treatment"      "contr.poly"

The value of contrasts(lang) reflects the fact that -- quoting from ? 
contrasts -- "If too few [entries for the contrast matrix] are  
supplied, a suitable contrast matrix is created by extending value  
after ensuring its columns are contrasts (orthogonal to the constant  
term) and not collinear."

Now let's generate some artificial data and look at how to interpret  
models fit using the old and new contrast matrices:

 > set.seed(3)
 > beta <- c(0,0.26,-0.004)
 > speaker <- rep(1:m,langs*n)
 > b <- rnorm(m,0,0.1)
 > y <- beta[lang] + b[speaker] + rnorm(3*m*n)
 > contrasts(lang) <- old.contrasts
 > print(m.old <- lmer(y ~ lang + (1 | speaker),REML=F))
[...]
Fixed effects:
             Estimate Std. Error t value
(Intercept) -0.01403    0.13236 -0.1060
langChinese  0.49860    0.18719  2.6636
langKorean  -0.11447    0.18719 -0.6115
[...]
 > contrasts(lang) <- new.contrasts
 > print(m.new <- lmer(y ~ lang + (1 | speaker),REML=F))
[...]
Fixed effects:
             Estimate Std. Error t value
(Intercept)  0.11402    0.07642   1.492
lang1       -0.12804    0.10807  -1.185
lang2       -0.43350    0.13236  -3.275
[...]

Ignoring speaker-specific effects, the predicted mean for a given  
language is the intercept plus the dot product of the language's  
contrast-matrix representation with the coefficients for the language  
factor.  Since the two models are equivalent, their predicted means  
should be the same for each language.  And they are:

 > ## compare old contrasts and new contrasts
 > ## English: old model
 > fixef(m.old)[1] + sum(old.contrasts["English",] * fixef(m.old)[2:3])
(Intercept)
-0.01402749
 > ## English: new model
 > fixef(m.new)[1] + sum(new.contrasts["English",] * fixef(m.new)[2:3])
(Intercept)
-0.01402749

The same will come out to be the case for the other two languages.

So -- to get back to your question: what do the nativeLanguage1 and  
nativeLanguage2 coefficients mean in your new model?  First, your  
contrast matrix has columns summing to 0, so the intercept can loosely  
be thought of as the predicted grand mean.  The coefficient for  
nativeLanguage1 is the difference between (a) the intercept and the  
English mean, and (b) twice the difference between the intercept and  
the average of the Chinese and Korean means.  The coefficient for  
nativeLanguage2 is the difference between Chinese and Korean divided  
by the square root of two.  So your guess was basically correct.  But  
it is important to recognize that these two coefficients operate on  
different scales, as reflected by the fact that the two columns of  
new.contrasts are vectors of different lengths.

> I would greatly appreciate any advice on how to interpret  
> regressions after contrast coding, or pointers to appropriate  
> resources on this topic!

So -- I wish I knew a really good reference on contrast coding.  There  
is some useful information in Chambers & Hastie 1991, Section 2.3.2,  
and in Venables & Ripley 2002, Section 6.2.  I think that Healy 2000  
("Matrices for Statistics") is a useful book that has some pertinent  
information.  But if anyone out there knows a great reference for  
contrast coding -- I'd love to hear it too!

Best

Roger

--

Roger Levy                      Email: rlevy at ling.ucsd.edu
Assistant Professor             Phone: 858-534-7219
Department of Linguistics       Fax:   858-534-4789
UC San Diego                    Web:   http://ling.ucsd.edu/~rlevy









From dderrick at interchange.ubc.ca  Wed Sep 23 23:42:25 2009
From: dderrick at interchange.ubc.ca (Donald Derrick)
Date: Wed, 23 Sep 2009 23:42:25 -0700
Subject: [R-lang] MCMCglmm with relatively small multinomial data set
Message-ID: <8122C193-959B-4E59-8120-5CBDA2DE31A6@interchange.ubc.ca>

Hello R-lang,

My name is Donald Derrick.

Along with my supervisor Bryan Gick, I am researching categorical  
kinematic variation in flaps and taps in English, and have identified  
four forms:

1) Alveolar taps - where the tongue come from below the alveolar  
ridge, hits it and returns to a lower position
2) Down flaps - where the tongue comes from above the alveolar ridge,  
hits it and continues downward
3) Up flaps - where the tongue comes from below the alveolar ridge,  
hits it and continues upward
4) Post-alveolar taps - where the tongue comes from above the alveolar  
ridge, goes straight forward, hits a spot at or above the alveolar  
ridge and comes back.

In words/phrases with only 1 flap, the independent variables are:

1) Subject variability (of course)
2) tongue speed (which I have measured)
3) tongue tip position before and after the flap

The tongue tip is always low with regular vowels, but it can be either  
high for rhotic vowels (tongue tip up bunched 'r' or retroflex 'r'),  
or low (tongue tip down bunched 'r')

This leaves the factors of

3a) tongue tip position immediately before the flap
3b) tongue tip position immediately after the flap


If I use more traditional balanced statistics (like a Friedman test)  
to compare rhotic and non-rhotic vowels, I cannot add in the details  
of type of rhotic vowel (up or down) because different subjects use  
different tongue positions and I no longer have balanced data.

Also, the Friedman test is ordinal, so I have to argue for a ranking  
of flaps.
[An easy ranking is tongue height favoring end position, where I get  
1) (low-low) alveolar tap, 2) (high-low) down flap, 3) (low-high) up- 
flap and 4) (high-high) post-alveolar-tap.  It's a bit bogus, but then  
so is every other ordinal ranking scheme I've ever seen!]


For both these reasons, I think I should be using a mixed effects  
linear model.


Unfortunately, It seems I cannot use the glmer() command here because  
my dependent variable is multinomial

SO, I am tring to use MCMCglmm, but I don't know if I am using it  
right, and also I cannot get convergence.  I need help.

My first question is, should I even bother?

I have small datasets by the standards of many of you, and since they  
took an entire year to build and need to graduate someday, I will  
never get bigger ones.  When I say small, I mean 18 subjects, and  
typically 12-24 phrases for any given category of analysis, as little  
as 216 phrases in each category, and as many as 432.

On the other hand, these are typical of my datasets in that the fixed  
predictors are hugely predictive.


SO, my next question is, can I even use MCMCglmm and hope to ever get  
convergence?

Now, if all is not hopeless, I will go on:

To make it easy to help, I've provided code I am using based on the  
HLP/Jaeger blog, and ask detailed questions throughout:

///////////

prior = NULL
flap.glmm = NULL
k <- length(levels(flap$firstflaptext))
I <- diag(k-1)
J <- matrix(rep(1, (k-1)^2), c(k-1, k-1))

# k = the number of levels for type of flap.  This is 4 - as described  
above.  I coded them "U", "D", "H", and "L"

    prior <- list(R = list(fix=1, V=0.5 * (I + J), n = 3),
    G = list(
    G1 = list(V = diag(3), n = 3),
    G2 = list(V = diag(4), n = 4)
    ))

# R = the standard setup for a "categorical" family test based on the  
HLP/Jaeger lab blog help

# fix=1 I believe is standard for a test of this kind

# n=3 I believe is the lowest we should use.  Higher numbers seem to  
work better, but not good enough.
# I do not know why :(

# G1 is for the flap types
# G2 is for the vowel context.  (more later)

for(i in c(1:3))

# I run the same model 3 times for covariance testing - I may want to  
try different priors,
# but I barely understand how to build them and cannot even get  
convergance this way yet.

{	
    flap.glmm[[i]] <-MCMCglmm(firstflaptext ~ -1 + trait + context,
    random = ~us(trait):subject + us(context):subject,
    rcov = ~us(trait):units,
    prior = prior,
    family="categorical",
    burnin = 5000,
    nitt = 50000,
    data=flap,verbose=TRUE)
}

# firstflaptext ~ -1 + trait + context
# I would love to be able to use something like
# precedingContext * followingContext
# - but every time I tried I got a "bad prior" error and absolutely no  
hint as to how to fix it.
# So I just combined the preceding and following context into a single  
variable - this gave me
# VV (for "autumn"), VR (for "otter"), RV (for "Berta"), and RR (for  
"murder")
# This solution does not seem ideal, so any help would be appreciated

# random = ~us(trait):subject + us(context):subject
# This seemed standard based on on the HLP/Jaeger lab blog help

# nitt = minimum for raftery test below

flap.coda <- mcmc.list(
chain1=mcmc(flap.glmm[[1]]$Sol),
chain2=mcmc(flap.glmm[[2]]$Sol),
chain3=mcmc(flap.glmm[[3]]$Sol))

# I built a mcmc list based on the fixed effects
# Q: How do I do this for the random effects, or does it matter?

flap.gelman <- gelman.diag(flap.coda, transform = TRUE)
flap.raftery <- raftery.diag(flap.coda, q=0.025, r=0.005, s=0.95,  
converge.eps=0.001)

# Two covariance tests.
# Below are the results

 > flap.gelman

Potential scale reduction factors:

                      Point est. 97.5% quantile
traitfirstflaptext.H       1.05           1.13
traitfirstflaptext.L       1.03           1.09
traitfirstflaptext.U       1.03           1.07
contextRV                  1.27           1.79
contextVR                  2.41           4.45
contextVV                  2.29           4.14

Multivariate psrf

3.38

# Note, these numbers do not get better with higher burnin or nitt :(

 > flap.raftery

$chain1

Quantile (q) = 0.025
Accuracy (r) = +/- 0.005
Probability (s) = 0.95

                       Burn-in  Total Lower bound  Dependence
                       (M)      (N)   (Nmin)       factor (I)
  traitfirstflaptext.H 3        4212  3746         1.12
  traitfirstflaptext.L 3        4368  3746         1.17
  traitfirstflaptext.U 4        4615  3746         1.23
  contextRV            25       22940 3746         6.12
  contextVR            4        4968  3746         1.33
  contextVV            8        10200 3746         2.72


$chain2

Quantile (q) = 0.025
Accuracy (r) = +/- 0.005
Probability (s) = 0.95

                       Burn-in  Total Lower bound  Dependence
                       (M)      (N)   (Nmin)       factor (I)
  traitfirstflaptext.H 3        4449  3746         1.19
  traitfirstflaptext.L 2        3776  3746         1.01
  traitfirstflaptext.U 2        3987  3746         1.06
  contextRV            25       31880 3746         8.51
  contextVR            15       13680 3746         3.65
  contextVV            6        8844  3746         2.36


$chain3

Quantile (q) = 0.025
Accuracy (r) = +/- 0.005
Probability (s) = 0.95

                       Burn-in  Total Lower bound  Dependence
                       (M)      (N)   (Nmin)       factor (I)
  traitfirstflaptext.H 3        4061  3746         1.08
  traitfirstflaptext.L 3        4368  3746         1.17
  traitfirstflaptext.U 3        4289  3746         1.14
  contextRV            15       16749 3746         4.47
  contextVR            18       18096 3746         4.83
  contextVV            15       15918 3746         4.25


## In this run, the third chain suggested convergance.  I'm not sure  
it is real though :(
# Especially since it has never happened before, and higher burnin or  
nitt tend to not help.

//////

Any and all help is appreciated - and apologies for such a hideously  
long post!