juliamixedmodels / embraceuncertainty Goto Github PK

Now that the release version of Julia is 1.10.0 I don't have a julia-1.9 IJulia kernel accessible through juliaup without adding it explicitly. @palday Would it be okay to bump the jupyter setting in the files in this repository to julia-1.10?

"These values are often considered as some sort of “estimates” of the random effects. It can be helpful to think of them this way but it can also be misleading. As we have stated, the random effects are not, strictly speaking, parameters — they are unobserved random variables. We don’t estimate the random effects in the same sense that we estimate parameters."

It could be worth pointing out that in Bayesian LMMs these actually have the status of parameters.

Are these bootstrapped samples of $\beta$ or $\hat \beta$? Also, the y-axis should be labeled "density", not pdf?

"To reiterate, this model can be reduced to a linear model because the random effects are inert, in the sense that they have a variance of zero." Is this sentence correct? The random effects don't necessarily have a variance of zero, the estimate of the variance from the data is zero.

Also, regarding the model being called degenerate, is it the model that is degenerate or the variance covariance matrix? In Pinheiro and Bates I think degenerate was only used in the context of the vcov matrix of random effects.

Finally, it may be worth mentioning that the 0 estimate of the random effect sd is just due to sparse data. If we had lots more data, the estimate would probably be non-zero. Simulation can demonstrate this point.

I probably missed it, but got stuck for a bit trying to figure out how to install EmbraceUncertainty when running the code here: https://juliamixedmodels.github.io/EmbraceUncertainty/intro.html#sec-dyestuff

It might be useful to reiterate here that you need something like the following code (even if it is mentioned elsewhere):

using Pkg; Pkg.add(url="https://github.com/JuliaMixedModels/EmbraceUncertainty")

@dmbates commented over in #18

I notice that figures are available in both .svg and .png formats. If would be good if one of these could be suppressed. Occasionally it is desirable to have .png, when e.g. the number of points to plot is very large, but generally .svg looks better.

The prefixes are described here.

But by default pandoc treats the top level as a section and not a chapter, as discussed in the caveats:

Because pandoc-crossref offloads all numbering to LaTeX if it can, chapters: true has no direct effect on LaTeX output. You have to specify Pandoc's --top-level-division=chapter option, which should hopefully configure LaTeX appropriately.

I propose that we have all the datasets that we use available through the EmbraceUncertainty.dataset generic and remove the data directory.

For the purposes of the book dataset(:foo) will mean EmbraceUncertainty.dataset(:foo).

Those datasets that are available through MixedModels.dataset will be grandfathered into EmbraceUncertainty.dataset by checking the names against the result of MixedModels.datasets().

Data tables from the English Lexicon Project are available in the arrow directory of https://github.com/dmbates/EnglishLexicon.jl

The datasets used in the longitudinal data chapter are available in the osf.io repository https://osf.io/sr46c/.

For the time being I think I will just customize the dataset retrieval for each name just so we can get it working. Later we can decide if there should be a unified storage and retrieval protocol.

• Wald tests on coefficients
• Likelihood ratio tests

everything in context of model structure

Currently we assign Grouping() contrasts to any grouping factors. In the past it was necessary to do this to avoid unnecessarily creating large, dense contrasts matrices when grouping factors had many levels.

After JuliaStats/MixedModels.jl@9ee4a0a we no longer need to so (thanks to @palday).

Shall we remove the explicit assignments of Grouping() contrasts? This will allow us to avoid creating a contrasts Dict until late in the second chapter when we apply a +/- 1 encoding to a binary experimental factor.

Also, the show_progress named argument to parametricbootstrap has been deprecated in favor of progress. It makes sense now to assign

const progress = false

and add a named argument progress to calls to fit and parametricbootstrap.

I will create a PR for these changes if there is no objection.