[R-sig-ME] specifying crossed random effects for glmmPQL / lme

[Phillip Alday]

As an oblique not-quite-an answer ....

brms has an exponential Gaussian (exgaussian) option for the error
distribution / family, which as the documentation notes is "especially
suited to model reaction times". You have to go Bayesian, but you can
estimate crossed random effects and their associated correlations
without any special tricks.

Phillip

On 10/04/2017 09:55 PM, Van Rynald Liceralde wrote:
> Thanks for your response, Ben! The paper that argued the use of the
> identity link with Gamma for response time data is Lo & Andrews (2015) (doi:
> 10.3389/fpsyg.2015.01171 <https://dx.doi.org/10.3389%2Ffpsyg.2015.01171>).
> Would such a model still be computationally problematic if the observed
> values fall very much within the domain of the specified probability
> distribution (i.e., valid response times are always above 200 ms)?
> 
> Re: "allow for correlations of random effects to be estimated", I've been
> told that it's more tractable to estimate covariances between the random
> slopes and intercepts (as I want with my model) using PQL than
> Laplace/AGHQ. In fact, Lo & Andrews' demonstration using glmer explicitly
> specified the covariances between the slopes and the intercepts to be 0 due
> to the computational rigor of specifying a model with random
> intercept-slope covariances in glmer and due to theoretical reasons.
> 
> And thanks for pointing out the Pinheiro & Bates (2000) reference to
> specifying crossed effects!
> 
> Sincerely,
> Van Liceralde
>