NickCH-K
commented
2 years ago Hi! Thanks for inviting me to this. I hope commenting here is the right way to interact.
Many of the bullet points refer to areas I don't know much about, but some are for ones I do! Some notes:
- I disagree with this in the opener: "There do not exist basic R functions to do causal analyses straight away" - not true! You can (and people do) do good causal analysis with
lm(), or heck evenmean()oraggregate()or a million other functions. If you mean that there aren't basic R functions that directly implement standard CI designs with all the bells and whistles, that's more accurate. - The package fixest should probably be included under "Regression models where the causal estimand is a regression parameter". This package is in most ways an improvement over estimatr for that task and also for instrumental variables. fixest also includes a number of tools for working with difference-in-difference designs, including the Sun & Abraham estimator, and should be included in the DID section. Keeping track, fixest should probably take estimatr's place in "Regression models" and "instrumental variables", and should be added to DID.
- How does estimatr do IPW designs? Is it just that it accepts weights? Most regression models do, including
lm(), so I'm not sure this is a useful entry there. Unless I'm missing a specific application they have. - There should probably be an entry for fixed effects designs. fixest pops up there again, for both linear and GLM fixed effects models, and fixed effects combined with instrumental variables, for which plm and lfe are also options. There is also alpaca for GLM fixed effects.
- In the RDD section I'd probably mention rdrobust first, and then afterwards go into the rest of the package family, since
rdrobustis the main one you'd use. Also suggest the addition of RDHonest for inference under poorly-behaving running variables - Minor typo: extra space in the DAGitty link.
- Additional matching packages: cem for coarsened exact matching, and ebal for entropy balancing
- Also in Matching: the Matching package, already listed, also does k-nearest-neighbor matching.
- Recommend an addition of tidysynth to synthetic control for its ease of syntax.
- Might be worth the addition of estudy2 for event study estimation, although perhaps this is too niche
zeileis
tuxette
imkemayer
julierennes
davidjohannesmeyer
eddelbuettel
Scope
This CRAN Task View proposal contains a list of packages useful for causality related analyses, including causal discovery and mainly causal inference. On top of the list of reviewed packages, we plan to include a list of selected books and articles that make use of some of these packages and that are helpful to get a better understanding of the topic of causality on concrete examples implemented in R code. Causal analyses can be performed in many different ways, depending on the context, the data and the assumptions made by the analyst. With this task view we aim at assisting the user in the choice of adequate implemented methods for his/her problem and data to analyze.
Since some of this proposed CTV maintainers are also maintainers of the already existing MissingData CTV, we followed a similar approach to select packages into the task view. To be considered for inclusion, the package must be useful for running causal analyses, be it causal discovery, causal inference or causal prediction, either by providing analysis tools or relevant datasets. Furthermore the package should satisfy the following technical criteria: the date of first release of the package should be more than one year old and the date of last update or the frequency of updates should reflect a regular maintenance of the package.
Packages
Please see https://www.imkemayer.com/files/causalinference for a comprehensive list and preview of the Task View. It currently consists of 150 packages (including 9 core packages).
Overlap
There might be some overlap between packages in the ClinicalTrials and ExperimentalDesign CTVs, but to minimize this overlap, our CTV proposal focuses on issues that arise mainly in observational (or quasi-experimental) data analysis and less on experimental design. Since causal inference is an important topic in econometrics, there is also some overlap with the Econometrics CTV but we aim at targeting an audience from a wider range of domains, for example with problems from social and health care applications. Missing data handling and causal inference share some similar methods and certain packages can be used in the two contexts but with different goals. Therefore overlap between this present proposal and the MissingData CTV is possible but does not imply redundancy.
Maintainers
The primary maintainer is Imke Mayer (@imkemayer). Current co-maintainers are Pan Zhao (@panzhaooo), Nick Huntington-Klein (@NickCH-K) and Julie Josse (@julierennes).
Collaborators of these maintainers have already agreed to assist in the maintenance of the task view to ensure constant updates and corrections where necessary.