Next steps to promote the pacakge

[TianyiPeng]

For people who are interested in improving this package (e.g, @fedemolina, @arushi-jain-27), I am wondering what your suggestions are for making this package more impactful and useful.

For now, I think the first priority is to have a readable documentation, like the one used by causalML (https://causalml.readthedocs.io/en/latest/about.html), linearModels (https://bashtage.github.io/linearmodels/iv/index.html), and even scikit (https://scikit-learn.org/stable/).

But I am also looking for more suggestions, e.g., what methods do you think we should include, should we have better and more organized tutorials, etc.

Look forward!

I just realized the download of this package has exceeded 13K! Congrats and thank you for everyone who ever contributed to this project!

[fedemolina]

Hi @TianyiPeng,

Some thoughts on my experience using the library, and I’ll likely have more to share.

I completely agree that improving the documentation, API reference, examples, and tutorials should be the first priority. Additionally, providing (more) mathematical intuition and explanations of the results would be valuable, especially for new users who are not familiar with causal inference. Although to me the api provides a clearly comparison showing the objective functions of every method the econml provides a deeper explanation of the problem setup.

For reference, both the gsynth and fect packages offer extensive documentation. Expanding examples and tutorials with datasets from different domains—such as retail (as mentioned in the paper), healthcare, and more—could make the package more accessible and widely applicable.

Some Feature Suggestions. Both gsynth and fect include plot methods and test methods, which inspire a couple of suggestions:

• Dataset Handling Class: A class to manage wide-format datasets ($N \times T$), with built-in imputation methods if necessary.
• Visualization/Analysis Tools: A class to generate plots and analyses that help investigate estimation results and model performance. Alternatively, it could be structured as an 'analyzer' class that takes the results from each method as input and performs various tasks (e.g., plotting, statistical analysis, etc.). Since the result object already contains key information such as $\tau$ (as a scalar or vector), baseline, standard deviation, and other relevant metrics, integration should be straightforward

Related to Counterfactual Matrix & Model Comparisons. When using the package, I estimated $\tau$ using different methods, but I was also interested in how the counterfactual matrix compares against the observed matrix.

I understand that the "counterfactual" matrix depends on the method used (e.g., Sections F and N of the Debiased Convex paper). However, in Matrix Completion methods, the baseline itself serves as the counterfactual. I found that the Matrix Completion baseline did not perform as well as Debiased Convex when visually inspecting individual time series.

It would be helpful to have built-in plot methods or recommendations on how to inspect results effectively.

With respect to Additional Methods to Consider. Other methods that could be interesting to include:

• Interactive Fixed Effects (IFE, as proposed by Xu) → Allows estimation of individual treatment effects.
• Causal Forest → Could be useful for heterogeneous treatment effect estimation.
• Causal Impact Analysis → Given matrices Y and Z, this could be applied row-by-row, using other rows as controls.
• Advancements in Difference-in-Differences (DiD) → Integrating recent developments in DiD methods.

Benchmarking & Expanded Focus on Treatment Effects. Including benchmarks with common datasets (such as the tobacco dataset) would help to compare the implemented methods and provide a useful reference for users.

Additionally, I believe the package should emphasize not just ATT, but also:

• Individual Treatment Effects (ITE) → Once the baseline counterfactual matrix is estimated, it becomes possible to analyze treatment effects at a more granular level.
• In Debiased Convex methods, when using τ as a vector and Z as a list, we obtain row-wise debiased effects, but this is not clearly highlighted in the current documentation.

And lastly, a method for determining how to define the spectrum cut would be really useful. This could be automated based on a metric or include recommendations from the literature. And as I mentioned in #19 maybe a fit method to Matrix Completion so that all methods are use in the same way? Also the non negative restriction could be useful also for datasets with small values where negative values are not allow (maybe even a warning could be interesting).

Maybe I’m a bit biased since I work with retail data, but I truly believe CausalTensor has a lot of potential across various domains!

[arushi-jain-27]

I can take up the documentation work to make the package more structured. Regarding enhancing the package, Federico has some great ideas especially related to additional methods that we can incorporate—we can prioritize and implement them based on what would be most impactful. I don’t have many suggestions, but I’m happy to take up whatever is needed. Contributing to this will help me learn more along the way!

[TianyiPeng]

@fedemolina Apologies for the delayed reply! I'm really impressed by your enthusiasm and thoughtful suggestions—your comments make a lot of sense and could truly help take the package to the next level.

@arushi-jain-27 Thank you as well for your support! I hope everything is going great on your end.

I’d love to see if you both might be interested in catching up online to explore ways you could contribute to (or even lead) the continued development of the package. If you're open to it, feel free to drop me a note at tianyipeng95@gmail.com.

By the way, I just noticed we’ve hit 19K downloads—super encouraging!

[fedemolina]

Hi @TianyiPeng , no worries at all, and thank you for your kind words!

I’m really excited to hear that the package has reached 19K downloads (20 now!)—congratulations!

I’d definitely be interested in catching up and exploring how I could contribute further and learn more about the topic. I’d love to exchange ideas and see how I can best support the continued development.

I’ll follow up via email!