[R-sig-ME] Resume terminated lmer fit if verbose=TRUE?

[Mike Lawrence]

The data come from a dozen or so published experiments using the
Attention Network Test (ANT), which is a video-game like test of human
attention. The data contain about 1000 participants, each observed in
a series of 288 "trials" that last only a couple seconds and terminate
when the participant responds to a target. We record response time and
accuracy of these responses and analyze these variables as a function
of the characteristics of the trial. Trials are characterized by the
kind of cue that preceded the target (4 kinds), the location of the
target (2 locations), the identity of the target (2 identities), and
the kind of distractor items presented concurrently with the target (3
kinds). My colleagues and I are furthermore interested in how the
features of trial N affect performance on trial N+1. For the model of
accuracy, we therefore have:

acc ~ (1|participant) +
cue*previous_cue*distractor*previous_distractor*target_location_repeat*target_identity_repeat

In theory, I could drop the "cue" and "previous_cue" variables because
my colleagues haven't generated predictions for the influence of those
variables. (There's also a 7th variable that distinguishes a slight
methodological difference between some experiments, hence my mention
of 7 fixed effects variables before)

When the model fitting completes, I aim to generate a posteriori
samples from the model (as implemented in the dev version of
ezPredict: https://github.com/mike-lawrence/ez/blob/master/R/ezPredict.R)
with which my colleagues will investigate a series of comparisons they
have highlighted that distinguish competing theories that aim to
account for performance in this task. (To facilitate such comparisons,
the output of ezPredict is formatted to match the input format
required by ezBootPlot:
https://github.com/mike-lawrence/ez/blob/master/R/ezBootPlot.R).

Indeed, this has already been achieved for the response time data, and
using the a posteriori data one of my colleagues has rather elegantly
proposed a handful of "rules" that appear to account for a large
number of phenomena in the data. After checking that the accuracy data
don't diverge from this account, my next step is to compute likelihood
ratios for each rule (as well as their combination) to quantify the
degree to which they account for the data.

Mike


On Wed, Jul 27, 2011 at 9:05 PM, Dennis Murphy <djmuser at gmail.com> wrote:
> Just out of curiosity, why would you need a seven factor interaction
> (assuming V1-V7 are all factors)? For that matter, why would you need
> more than 2fi's? Is there some scientific reason why that might make
> sense?
>
> Dennis
>
> On Wed, Jul 27, 2011 at 11:39 AM, Mike Lawrence <Mike.Lawrence at dal.ca> wrote:
>> Thanks to Harold for pointing out the start argument and indicating
>> that the verbose output can be supplied as its value. Thanks also to
>> both Harold and Ben for warning on over-parameterization, though it
>> may not fully apply here as my model is of the form "y ~ (1|random) +
>> V1*V2*V3*V4*V5*V6*V7". (well, I might have a priori reason to
>> eliminate one or two fixed effects...)
>>
>> Mike
>>
>> On Wed, Jul 27, 2011 at 10:47 AM, Ben Bolker <bbolker at gmail.com> wrote:
>>> -----BEGIN PGP SIGNED MESSAGE-----
>>> Hash: SHA1
>>>
>>> On 07/27/2011 09:29 AM, Doran, Harold wrote:
>>>> Yes. See ?lmer and the start argument. You can provide the function
>>>> with starting values, which can come as the last iteration of the
>>>> output from verbose. 100 hours is a ridiculous amount of computing
>>>> time.
>>>
>>>  Oops. I was wrong (thanks).  I was thinking of lme.
>>>
>>>  A common theme on this list seems to be that people set up models of
>>> the form
>>>
>>> response ~ (a lot of fixed effects) + (a lot of fixed effects|grouping)
>>>
>>>  The problem here is that n fixed effects  interacting with the
>>> grouping variable means n*(n+1) variance-covariance parameters, which is
>>> often slow to fit ... see e.g.
>>> http://article.gmane.org/gmane.comp.lang.r.lme4.devel/6308 ...
>>>
>>>
>>>>
>>>> You might want to check your model and make sure it isn't
>>>> overparameterized.
>>>>
>>>>> -----Original Message----- From:
>>>>> r-sig-mixed-models-bounces at r-project.org
>>>>> [mailto:r-sig-mixed-models- bounces at r-project.org] On Behalf Of
>>>>> Mike Lawrence Sent: Wednesday, July 27, 2011 9:16 AM To:
>>>>> r-sig-mixed-models at r-project.org Subject: [R-sig-ME] Resume
>>>>> terminated lmer fit if verbose=TRUE?
>>>>>
>>>>> Hi folks,
>>>>>
>>>>> I ran a binomial lmer on a large data set with lots of levels of a
>>>>> random effect (1000+) and a large fixed effects structure (7
>>>>> variables all interacting). My local machine wasn't up to the task
>>>>> (it needs about 16GB of memory so far as I can tell), so I put it
>>>>> on a serial node on a supercomputer to which I have access. I asked
>>>>> for 100 hours of compute time, but it seems that the model went
>>>>> over time and the process was terminated by the system's queue.
>>>>> However, I ran it in verbose mode and have all the output, so I'm
>>>>> wondering if there's any way to use this information to resubmit
>>>>> the job and have it resume where it left off. Any ideas?
>>>>>
>>>>> Cheers,
>>>>>
>>>>> Mike
>>>>>
>>>>> -- Mike Lawrence Graduate Student Department of Psychology
>>>>> Dalhousie University
>>>>>
>>>>> Looking to arrange a meeting? Check my public calendar:
>>>>> http://tr.im/mikes_public_calendar
>>>>>
>>>>> ~ Certainty is folly... I think. ~
>>>>>
>>>>> _______________________________________________
>>>>> R-sig-mixed-models at r-project.org mailing list
>>>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>>
>>>> _______________________________________________
>>>> R-sig-mixed-models at r-project.org mailing list
>>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>
>>> -----BEGIN PGP SIGNATURE-----
>>> Version: GnuPG v1.4.10 (GNU/Linux)
>>> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>>>
>>> iEYEARECAAYFAk4wFuMACgkQc5UpGjwzenMU8gCfewTblJqXdJKHfPwfbxF72PfK
>>> o+AAn0ASTeqSPx/N2wr5i/+CvSVGyDYw
>>> =3r0G
>>> -----END PGP SIGNATURE-----
>>>
>>> _______________________________________________
>>> R-sig-mixed-models at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>