r-causal/causal_inference_r_workshop

Causal Inference in R Workshop

causal_inference_r_workshop's Issues

04 lecture: consider adding ggdag theme example in plots

Would help resolve potential confusion around what the numbers on the axes mean

typos in g-comp exercises

We create a dataset wait_times but we accidentally use seven_dwarfs instead for the subsequent sections

average over stratified mean differences in 03-causal-infernence-with...

Can probably be done with a second pivot and summarize. The benefit of this would be that we can show it's equal to the ordinary adjusted lm() estimate

Not all parks open at 8?

re: the continuous example. May need to switch to a later time.

Need to update the book example if so

Show example of SMD with dummy variables?

r-causal/tidysmd#7

Re-work continuous exposure section

• Use g-comp
• brief slides at end on using propensity scores
• for last part, include how to check, e.g. difference in bootstraps, bootstrap distribution, etc
• make current exercises for PS bonus?

need to get propensity's continuous functionality up and running to streamline the last bit

Organize slides folder

e.g. slides/raw, slides/html, slides/pdf

Weighted tables with gtsummary

When the add_difference() method for survey objects is ready

Re-order material

• Quartets after DAGs
• Selection bias to bonus
• Remove g-comp section (moving to cont.)

Closely related to the cont. section re-work which I will post a separate issue for

Improvements to DAG module

• Some simulated data to prove it
• use dagitty::adjustmentSets() directly. Stretch goal?

Update weights to use Li's tilting functions

http://www2.stat.duke.edu/~fl35/teaching/640/Chap3.4_observational_weighting.pdf

11 lecture: liver vs lung cancer

Seems appearance of "lung" cancer is a typo and should be "liver"

Add solutions folder

We'll have to be mindful of updating solutions any time we update the exercises

Workshop website

Down the line, we should set up a workshop website (something like this:https://rr-in-r-workshop.netlify.app/)

Suggestion: Interpretations of causal effects, diagnostics, etc

A student suggested we address interpretations a bit more. This might be better suited for the book, though, e.g. reporting causal analyses or something. @LucyMcGowan, what do you think?

Figure out of wack?

For some reason the DAG is rendering weird for me: https://causal-inference-r-workshop.netlify.app/12-whole_game-2.html#5

Does it look huge on your screen @malcolmbarrett?

Different estimates for boot strap of cont example?

Several people encountered this, and so did I!

rules for associational flow in DAG slides?

I've given the following 4 rules in the past which help identify when/why an association exists between variables. Not necessary to belabor it in the lectures, but could be a useful touchpoint to understand what dag_paths() is doing under the hood.

1. If no variables are conditioned upon, an association path is blocked when two arrowheads collide at some variable
2. A path that contains a noncollider that has been conditioned upon is blocked
3. A collider that has been conditioned upon does not block a path
4. Conditioning on a descendent of a collider does not block a path

Add table of estimands and when to use them

See r-causal/causal-inference-in-R#47

Switch whole game example?

There are pros and cons to revisiting the same question in different ways. We need to let this marinate

Change net data to be more complex

https://evalsp21.classes.andrewheiss.com/example/matching-ipw/

Mostly for the DAG part so students have to actually think it through

• update package data and attribution
• update slides and exercise file
• update solution

Switch to causaldata

Which supersedes cidata. Not urgent.

whole game: give further details on study question

Better enhance the question slide, in particular, specify the inclusion/exclusion criteria such as "must be current smoker" according to a target trial type question

Tweak tipr section

• Update to recent version of tipr
• Add adjust example
• Show with a simulated confounder?
• If simulating something like age, actually show how we would do the SMD for the argument

Group by and summarize exercises

maybe not add a separate set of exercises but have students just fill in group by or something

ggdag warning re: segment.alpha

This pops up in the whole game exercises, for example

> smk_wt_dag |>
Warning messages:
1: Unknown or uninitialised column: `segment.alpha`. 
2: Unknown or uninitialised column: `segment.alpha`. 
3: Unknown or uninitialised column: `segment.alpha`. 
4: Unknown or uninitialised column: `segment.alpha`. 
+   ggdag(text = FALSE, use_labels = "label")

I searched a bit through {ggdag} but wasn't able to find an obvious source. Not urgent since things work as-is, would just be good to fix to tidy and prevent questions.

Censoring with Disney data

@LucyMcGowan in the G-comp section we do a loss-to-follow-up model to account for censoring using the NHEFS data. Do we have anything we can use for that in the Disney data sets?

Switch to Quarto

• Write theme extension matching our slides
• Switch slides
• Tweak slides to make more consistent
• Switch exercises
• add hex stickers

Use smd package for SMDs and love plots

For basic calculation, it's not too rough to get SMDs with the smd pkg + dplyr

library(smd)
library(dplyr, warn.conflicts = FALSE)

mtcars %>% 
  mutate(wts = runif(n())) %>% 
  summarise(
    across(
      c(mpg, disp, hp), 
      ~smd(.x, am, wts)$estimate)
  )
#>    mpg      disp     hp
#> 1 -1.808428 2.198783 1.064067

^{Created on 2021-09-02 by the reprex package (v2.0.1)}

Additionally, gtsummary will have SMDs in a new update, potentially supporting weights. It also supports fully-weighted tables via survey and gtsummary::tbl_svysummary().

I think we should break this up into two steps 1) SMDs, weighted SMDS, love plots & friends with smd() 2) creating weighted tables with gtsummary + survey

05 lecture: use subscripts to denote time

slide 16:
e.g. $x$ is written twice, but it would be clearer to do something like $x_0$ and $x_1$

Add brief target trial section?

And make sure we use that approach throughout

Typo in love plot code

https://causal-inference-r-workshop.netlify.app/08-pscore-diagnostics.html#/smd-in-r-1

I think this should be color and group method not weights

add R code for PS models in 02-when-standard-methods...

Since the code which uses {propensity} is less verbose, it should probably now fit in the third pane in each of the example slides

Diagnostics for continuous and g comp

See r-causal/causal-inference-in-R#50