derek-jones / eseur-code-data Goto Github PK

You cover LOESS, polynomials and breakpoint modelling nonlinear patterns, but oddly do not cover splines (breakpoint modelling is a subset of regression splines limited to linear, slope = 0 models) in linear models. These have many advantages. Check out the coverage in https://link.springer.com/book/10.1007/978-3-319-19425-7 (book is on libgen). This book is linked to the rms package, which implements a lot of stuff. See here for an example. https://rpubs.com/EmilOWK/rms_splines. You cover GAMs later on, but don't touch on how the mixing of splines and normal additive/linear terms in models can yield interpretable but powerful models.

For understanding predictions from models, try the ggeffects packages. Example here. https://rpubs.com/EmilOWK/ggeffects_examples, official examples https://strengejacke.github.io/ggeffects/

A more sophisticated approach looks at the distribution of predictions, and makes
an informed trade-off between true positive (in this context also known as recall and hit
rate), and accuracy (i.e., the false positive rate).

https://en.wikipedia.org/wiki/Sensitivity_and_specificity

I think you mean specificity, not accuracy. Accuracy is not the false positive rate (accuracy = (TP + TN) / TP + TN + FP + FN). The trade-off when selecting threshold is between TPR (sensitivity) and TNR (specificity).

In Figure 2.43 the caption says the fitted loess line is green, when in fact it is red.

Slightly hard to read, should probably make a big more whitespace upwards.

Seems like there is an error in the Y axis label here. It should be just "Effort" or maybe "Effort (central tendency)". I am guessing you wrote the label, and then later added the medians.

Page. 256

But this is false. The standard used normally is two-tailed test at .05, and this is sigma 1.96, not >2.

> pnorm(-1.96)
[1] 0.0249979
> pnorm(1.96, lower.tail = F)
[1] 0.0249979

These sum to ~5%. I also don't think higher impact journals generally have stricter standards. There is a review of studies of journal impact factor, and journal scientific rigor, finding generally no relationship. https://www.frontiersin.org/articles/10.3389/fnhum.2018.00037/full

I'd love to have a read of this on Kindle, PDF doesn't perform very well there. Might you consider a .mobi or .epub?

For some reason, there is a slight upwards spike in the pdf for the normal curve on the left tail here.

I like these data, but at first I was confused about them. I think they would be more clearly communicated with a survival curve type plot.