Notes

[mvanrongen]

• ## Context section at the very start for each chapter: 1 or 2 sentences to put into context what we're going to deal with.

• ## Section setup instead of Libraries & functions. See example Ch7. These should get evaluated, which will make rendering more robust than when relying entirely on the setup scripts.

• Figure caption. Currently done up to & including Ch8

• Ch2: update the setup instructions > link to software_installation repo

• Ch3: need to add information on how data is generated

• Ch5: we need to mention the phrase 'beta coefficients'. It's not mentioned & then comes up first in Ch9, followed by Ch11

• Ch8.5: Warning in eval(family$initialize): non-integer #successes in a binomial glm! >> needs addressing in text. Chapter also needs a finishing sentence before moving on to exercises

• Ch9 the table in 9.2 is super helpful, nice job. The whole explanation on significance testing flows a lot better than it did.

• Ch9.1 move the text into ## Context

• Ch 9.5 "Helpfully, this likelihood ratio approximately follows a chi-square distribution, which we can capitalise on to calculate a p-value." << this is quite a crucial point & needs a bit more explaining: When you have a test statistic and you know the distribution of that test statistic, then you can use this to calculate a p-value. We're using a different type of chi-square test. Not the traditional 2x2 contingency table variant.

• Ch10.4: "In most hypothesis tests, we want to reject the null hypothesis, but in this case, we’d like it to be true." << explain why

• Ch10.5 Very nice explanation! I changed the orange text to purple - it was hard to see on a white background.

• Ch11 Really good addition to the materials.

• Ch11.4 Checking assumptions: in 11.4 the Python part has no diagnostic plots. Might not be possible, but it's not mentioned & the section now abruptly finishes. The R narrative also finishes quite abrupt with 'Much better!'. It needs a concluding sentence, such as "Much better! This shows that the model is much improved by dropping the interaction term. MvR: add a comment that in Python we can't do the diagnostic plots, but that we assess all the assumptions individually with code.

• Ch11.4.2 if you want to make specific comments about the predictors, then collinearity is an issue because dropping one causes swings in beta estimates. If all you care about is the overall accuracy of your model (ML approach = maximum prediction) collinearity becomes irrelevant. The model + residuals don't change.

• Ch11.4.2 in R: the warning Cannot simulate residuals for models of class glm ... not explained/addressed

• Ch13 I like how you defined the coefficients in a separate chunk, then used them in the formula. Much cleaner than doing it inline - I'll be using that method from hereon :)

Slides:

• Maybe have a slide in the count data part, where we discuss the different ways that count data can be tricky (including zero-inflation, even if it's not in the practical course materials) Two types of zeros (number of publications, non-academics vs academic who haven't yet published) >> Vicki-H to do.