Majority voted to skip this. Already pretty overwhelming by this point.

Might also reduce the number of alternative models in that exercise.

Will make the repository easier to follow without the presentations and make for a complete set of notes.

Maybe simulate some data? Might also make a good blog post.

Also demonstrate how this only works if the group-level intercepts are constrained as "random effects".

Was a point of confusion for many.

Something a bit like the variance structure exercises but ideally with simple real data sets. Simulated would also be okay if there is some believable context given like with the random effect grouping exercises.

Make sure to include exercises that deal with interpreting the coefficients.

Make sure to include some overdispersed examples.

Hi Sean,
Upon running your random-slopes and temporal-autocorrelation functions I discovered that the broom package throws an error which can be rectified by loading the library(broom.mixed) in its place.
Great course materials by the way, really helped with my analysis of temporally correlated fish movement dataset.
Cheers,
Ross

Consider stripping out the count data GLMM entirely to focus on just one.

Avoid paste() and custom functions as much as possible.

Demonstrate making usable GLMM coefficient plots with simple code.

• dplyr and ggplot2 probably first to go
• zero inflation?
• shorten count data GLMM?

In 02-ggplot.Rmd the link http://docs.ggplot2.org/ no longer works

Think about how to plot it.
Think about what could be modeled.
Gain familiarity with the data.

This way participants learn to work with a function that will be useful to them with their own data.

Maybe the sjPlot package as used elsewhere?

hey

i tried to follow the code in the zero-inflated markdown and got a tibble error. not sure why since you never call a function from tibble..

inverse_logit <- function(x) plogis(x)
set.seed(123)
d <- data.frame(f = factor(rep(letters[1:10], each = 10)), x = runif(100))
u <- rnorm(10, sd = 2)
d$eta <- with(d, u[f] + 1 + 4 * x)
pz <- 0.2 # probability of excess zeros
zi <- rbinom(100, size = 1, prob = pz)
d$y <- ifelse(zi, 0, rpois(100, lambda = exp(d$eta)))
d$zi <- zi

ggplot(d, aes(x, y, colour = f, shape = as.factor(zi))) + geom_point() +
    scale_shape_manual(values = c(20, 21))

all works

but

ggplot(d, aes(x, y, colour = f)) + geom_point() + facet_wrap(~zi)

fails, returning Error: 'as_tibble' is not an exported object from 'namespace:tibble'

Plotting residuals? Writing in the formulas? Exploring the data at the beginning?

Introduce it early on in the workshop. Why do we progress from a linear model to a GLM to a GLMM etc. GAMs too.

I'm wondering what the chopstick data really represent. The original reference defines food-pinching efficiency as:

Food-pinching efficiency. The subject would sit on an adjustable seat, and was required to pick up peanuts from a dish (150 mm diameter) in front of the subject (450 mm) to a cup (200 mm high and 70 mm diameter) under the mouth for 1 min. During pinching, the experimenter counted the numbers of peanuts in the cup. The reason for using
peanuts was that it was difficult to pick them up and hence was more representative as a measure of the effect of the length of the chopsticks on food pinching efficiency. Fig. 3 demonstrates the workplace layout and task of the food- pinching.

This would suggest count data to me?? Table 5 in the original paper gives mean values by chopstick length that are consistent with the data here, but the values given in this data set are continuous rather than count data. I don't have access to the paper on the Elsevier web site, so I can't see if there is supplemental material there (although I kind of doubt it in a paper from 1991?) The data sets themselves seem to be floating around in a variety of places:

remotes::install_github("jr-packages/jrModelling")
?jrModelling::chopsticks

credits https://bmdatablog.files.wordpress.com/2016/04/chopsticks.pdf , which refers to https://www.udacity.com/api/nodes/4576183932/supplemental_media/chopstick-effectivenesscsv/download; I think these are the same data referred to here: https://towardsdatascience.com/chopstick-length-analysis-2c4c7e9b6136

(The jrModelling data set only has 2 of the 4 chopstick lengths).

There are only 171 unique values out of 186 rows, so I guess it's possible these data are rescaled versions of integer counts?

Currently uses slopes only as an artifact of why I originally wrote this code. Intercepts would be simpler and more straightforward.

LMs, GLMs, LMMs

y = b0 + b1 * x
log(y) = b0 + b1 * x
log(y) = b0 + b1 * log(x)

Factor/binary predictors

Centered / scaled predictors.

Give output or coefficient values for hypothetical models and ask to answer specific questions about interpretation or quantifying values.

Things like:

• interpreting basic coefficients
• when to consider logging data
• basic residual diagnostics for linear models
• binary, factor, and continuous predictors
• an example with centering a predictor
• reference textbooks for further information