In LM_intro_course.html:

Slide 46, we fit a model with planet + sex but don't yet show the coefficients. Do we want to show them?

GLM interval slide 60: do not mention bats. Checks similar sentences elsewhere!

Results for dfbetas and dffits are somehow a little off for extreme outliers showing that I must not correct with the right SE... To figure it out.
A good example is to focus on the following for testing:
{r} fit_davis <- lm(weight ~ height, data = Davis) influence.measures(fit_davis)$infmat[12, ]

Check that and see if it is worth mentioning.

• pkg: https://cran.r-project.org/web/packages/gvlma/index.html
• paper: https://www.tandfonline.com/doi/pdf/10.1198/016214505000000637?needAccess=true

As now, which() breaks the flow...

for now they are part of the GLM exercises...

This package seems to have a good method for testing random effects, to explore.

Not sure that PI for binomial and binary make any sense. Check.
Check that it does work for Poisson!

Zimai had a good example.
spaMM does not handle the beta family for the fixed effect, since it implies the estimation of an extra parameter (so not a GLM technically).
glmmTMB does that.

Fix title in slide 57 in LM_test_intervals_course.Rmd

In Introduction_course.Rmd:
We used negative indexing in our plotting examples (slide 34) but didn't introduce in the Introduction section.

I should show how to compute marginal predictions and explain the issue (see course for stats club).

After discussing the interactions in abstract, we could also add a plot to show how such an interaction would look.

To make it clearer for the students to understand what is there "just" to show them how things work from what they need to do in practice, I could perhaps split slides like that and use some graphical way to make it clear (different background? Different color for the titles?).

In his presentation (https://github.com/eco4cast/Statistical-Methods-Seminar-Series/tree/main/bolker_mixedmodels), I spotted several ideas that I could use, especially for checking model assumptions and plotings.
He also uses gamm4 to somehow fit spatially autocorrelated random effects.

In LM_fitting_course.Rmd

Slides on QR decomposition (pg 35-36) are assigned 🎓 but QR decomposition is surely the nerdiest part of the whole set of slides!! 🤓

Would it detect difference in variance in the absence of difference in means?
In other words, does it only test mean-variance relationship...?

I don't think that what I called marginal predictions are actually marginal...

There is a method I know nothing about that can be used for goodness of fit in GLM(M)s.
I should give it a try: https://www.jstatsoft.org/article/view/v081i10

Since I use these functions often, it would be good to introduce them.
Same for sapply(), replicate() and perhaps other functions that I keep using.

In Introduction_course.Rmd:

Comments for matrix multiplication on pg 28 are wrong. Should be:

## = [1 * -1 + 3 * 1][2 * -1 + 4 * 1]

simulate does not work correctly for gaussian() in link != "identity"

In GLM_intro_course.Rmd, slide 28. Use control = list(trace = TRUE) to show the issue occurs when it finds a very large deviance.

It would be better to phrase all slides testing whether tests are correct under the null hypothesis as tests of false positive or something like this rather than as robustness. Still looking for the right terminology...

The session takes 2 hours (with no practice) and covers a lot of ground but it is difficult to understand design matrices from the get go. Perhaps I should try to first present predictions informally (as done by hand element per element) and then introduce the design matrices. That would imply restructuring the material...

Alia's remark: some removed files are still there in old commits...
That's right, I need to rewrite the history.

Check and show in spaMM: summary(..., list(p_value = "Wald"))
but that is a chi2 based test, not a F or t-test using df...

• Explain ==, != (also & and |?)
• Explain how logical responses can coerce to numeric (maybe show what happens if you use sum() on logical data)

In GLM_intro_course slide 41:

Have a single legend rather than a new legend on each plot.

Notation not homogeneous...

It would be interesting to show different plotting tools in details such as coplot(), interaction.plot(), scatterplot() and scatter3d().

Instead of returning the full data frame, only return those with major influence

get_ranPars(fit , which="lambda") allows for extracting lambda.

This seems useful:

fit_davis <- lm(weight ~ height, data = Davis)
foo <- influence.measures(fit_davis)
foo$infmat[apply(foo$is.inf, 1, sum) > 0, ]

LM_intro_course.Rmd:

Around slide 26

In LM_intro_course.html:

Extends #33. Include once we introduce residuals (around slide 52-53)

Predictions with fit_UK1 -> that this refers to the model in the course is not clear.

In LM_intro_course.html:

Around slides 39-40, add a slide that shows how coefficients would change if we use a different reference level.

Slides about interpreting treatment contrasts.

In GLM_intro_course.Rmd slide 7.

the R code is in the folder sources, but it needs to be turned into a vignette

There are a huge number of dependencies, and that could be greatly reduced. Perhaps I should absorbe all the datasets used in the package inside it (citing their origin).

Why do we say it's 'better' when humans are removed.

Perhaps, all the fitting procedure (LM, GLM, LMM...) should all be moved into an appendix.

I should externalize everything that uses spaMM in a clean document for testing.

I need to export the dataset on the nodes:
anova(fit_spaMM2, fit_spaMM2_H0, boot.repl = 10, fit_env=list(dat.bcg=dat.bcg) )