Philipp Bach

👋 Hello, causal inference people in Berlin. Nice to meet you 🎉

Thanks, Paul! Looking forward to CDSM

Thank you @janmarcus.de and thanks to everybody who joined the event yesterday. For me it was a great kickoff for all the upcoming projects and teaching activities at @freieuniversitaet.bsky.social !

You want to see @philippbach.bsky.social and @shushmargaryan.bsky.social in one session?

Come to the welcome event for our new colleague Philipp Bach at @fu-berlin-vwl.bsky.social on Thursday!

Economist of the Free University Berlin at the Annual Meeting of the Verein für Socialpolitik

Great to see such a strong presence of @fu-berlin-vwl.bsky.social at the @vfsecon.bsky.social's Annual Conference in Cologne – always an inspiring venue for research and exchange!

@danzernatalia.bsky.social @piotrlarysz.bsky.social @philippbach.bsky.social @phaan.bsky.social @simonvoss.bsky.social

Last day to register for our BENA Skills Camp in September in Berlin! #EconSky #EconConf #dataSkyence #CausalSky

Join us for a 2 days hands-on workshop on Causal Machine Learning taking place in September at @freieuniversitaet.bsky.social

#EconSky #Causality

We are looking for PhD students! More information at www.wiwiss.fu-berlin.de/fachbereich/...

🎉 #EconSky #EconConf

Thanks! I totally agree with @mcknaus.bsky.social. Also whenever I start some new Causal "ML" projects, the first benchmark is always OLS & logistic regression learners; it helps you to see the connection to standard approaches; not only for linear regression, but also for doubly robust etc.

Mehr denn je nach dieser Woche

innn.it/phoenix-muss...

Ich meine das mit dem Aufruf zum Wählen übrigens ernst. Laut Forsa könnten wir 28% (!) Nicht-Wähler_innen haben. Liebe Wissenschaftler_innen, liebe Wissenschaftsinstitutionen: Euch hören viele zu. Erinnert sie, wie wichtig Wählen ist. Für unsere Demokratie. & ermuntert sie, Botschaft weiterzutragen!

Flyer for 2025 Applied Causal Graphs Workshop in Berlin, to be held on March 4th, 9:00-17:30 at the Charité Virchow Klinikum. Accepting abstracts until Feb 7th, 2025. Additional information at applied-causal-graphs.de

We're looking to connect Berlin & Brandenburg researchers working with causal graphs from all disciplines!
➡️ "Direct" link: applied-causal-graphs.de ⬅️
⏱️ Abstracts due Feb 7th!
#CausalInference #DAGs #Berlin #CausalGraphs
⭐ Keynotes by @philippbach.bsky.social @pwgtennant.bsky.social & Simone Maxand

Two days left to submit your abstract to EuroCIM 2025!

If you want the chance to present your work at the European Causal Inference Meeting 2025 in Ghent, send in your abstract no later than Jan 15, 2025.

Submission form and more information here: eurocim.org/abstracts.html

Oh, here's the handle of Jan 😀: @janteichertkluge.bsky.social

The paper is joint work with (I guess almost all bsky-less)

Victor Chernozhukov
@svenklaassen.bsky.social
Martin Spindler
Jan Teichert-Kluge
Suhas Vijaykumar

Looking forward to your thoughts, comments and questions!

The causal part:

If you are a #causal #DAG enthusiast, you'll finde some causal diagrams and a discussion on causal aspects of demand analysis in the paper too 😀

#CausalSky #dataSkyence

The fun part (that's what you usually don't read in the papers):

Embedding text and image data makes demand analysis pretty accessible from an intuitive point of view. You can play around with the product embeddings, check for similarities and formulate/check hypotheses for various demand patterns

Our learnings:

We find that text and image data play an important role in predictive and causal demand analysis: Improved demand prediction and advanced heterogeneity analysis using product infos encoded in text and images, e.g., based on similarities and AI/data-driven product categorization.

Table 5 from the referenced paper showing the predictive performance of various models used for demand analysis. Deep learning based approaches that utilize both image & text data are found to substantially better predict the quantity and price signals than traditional (linear regression with tabular features only) and ML learners (boosted trees with tabular features)

A sorted-effects plot summarizing the heterogeneity in price elasticities as obtained from AI-based heterogeneity analysis (three different model specifications). More information, see Figure 7 in the linked paper.

Our approach:

1️⃣ Enhanced Predictions: AI-driven embeddings significantly improve the accuracy of sales rank and price predictions

2️⃣ Improved Causal Inference: By fine-tuning embeddings for causal tasks, we uncover strong heterogeneity in price elasticity linked to product-specific features

An AI-generated image showing a red and blue toy car with eyes. The figure has been obtained from summarizing a product category called "Iconic Movie-Inspired 1:55 Scale Diecast Cars Perfect for Storytelling and Roleplay". The categorization has been obtained in the referenced paper. More details in Table 4.

🆕 New year, new working paper: Adventures in Demand Analysis using AI 🆕

Our question: How can we advance demand analysis using recent tools from AI (Deep Learning, LLMs etc)?

Our idea: Use information from text & images in digital marketplaces like Amazon

Paper: arxiv.org/abs/2501.00382 #EconSky

This looks like a pretty useful paper and - probably more importantly - a pretty useful practical procedure to find our what happens when running Causal Machine Learning. Balancing checks etc are common in traditional approaches (like PSM), but are usually mor tricky to assess in ML-based estimation

I'm trying to compete with @stephenjwild.bsky.social's DAG People starter pack, because economists believe in competition Open to suggestions! go.bsky.app/Fa2XSDH

Join our team at the Econ Department of Uni Fribourg! We're hiring for a Post-doc position and a Ph.D. position in an
SNF-funded project on #NetworkScience & #Economics, led by Berno Büchel (visit berno.info). Duration: 40 months. #PostDoc #JobOpening #EconSky

Regarding the Riesz representers: we have the analytical RRs implemented for the sensitivity part, but the data driven RR are still to be added. That's a bit experimental and not 100% clear how to integrate them in the package

Haha thanks. We haven't planned to include it in DoubleML yet, but maybe that's a good idea. Yes we are working on the RR too, but that may still take some time. Thanks 😊

In case you like to learn more about the ideas behind all these new features, join our DoubleML trainings: trainings.doubleml.org

Next training starts in March: doubleml-training-mar-2024.eventbrite.de

#EconSky

3. Python API Updates:
- Added Utility Classes and Functions: docs.doubleml.org/stable/api/a...

2. Multiple new examples: docs.doubleml.org/stable/examp...
- First Stage and Causal Estimation Notebook
- Basic IV Notebooks for Python and R
- GATE and CATE Notebooks für PLR
- GATE Sensitvity Notebook (for IRM or weighted average treatment effects)

1. Updated Userguide:
- GATE and CATE for PLR: docs.doubleml.org/stable/guide...
- Weighted Average Treatment Effects: docs.doubleml.org/stable/guide...
- External predictions: docs.doubleml.org/stable/guide...
- Updated description of Sensitvity Analysis (IRM): docs.doubleml.org/stable/guide...