A causal inference newsletter


Encouragement Designs and Instrumental Variables for A/B Testing

Spotify researchers introduce IV for industry applications, when randomizing the treatment is not possible or desirable, for example because the feature is common knowledge. A trade-off that emerges in the industry is designing an encouragement that is sufficiently strong but does not violate the exclusion restriction.

Nixla Releases TimeGPT for Time Series Forecasting

TimeGPT isa generative model for time series trained over 100 billion rows of financial, weather, energy, and web data. It incorporates conformal prediction for uncertainty quantifications and should work with panel data (called multiple time series). Time series prediction is key in causal inference under unconfoundedness and for variance reduction.

Can I A/B Test That? Yes, with Few Exceptions

Ronny Kohavi shares practical advice on what to ask oneself when considering whether to test a feature. It also includes an insightful story from its time at Amazon, showing how testing constraints are usually not material harm but rather strong beliefs.

The Experimentation Platform Journey at Adevinta

Adevinta shares the story behind the evolution of their Experimentation Platform. At first it was just a Python package to unify AB testing in their Dutch market (Marktplaats), and then it evolved into a cross-marketplace AB testing platform, Houston.

Old Reads

The Balancing Act in Causal Inference

The propensity score is the unique balancing score, i.e. the unique function that balances any function of observables. However, in practice it needs to be estimated and inverted. The article compares inversity propensity weighting with different balancing approaches.

Using Balancing Weights to Target the Treatment Effect on the Treated when Overlap is Poor

The authors compare the performance of inverse propensity weighting (IPW) and balancing when overlap is limited. Across three sets of simulations they find that balancing produces unbiased results even with limited overlap, while IPW is biased.

Balancing Covariates via Propensity Score Weighting

IPW estimators can be very volatile because of propensity score inversion. The authors propose instead to weight by the probability of being assigned the opposite group. The estimator has the lowest asymptotic variance but the estimand, which they call “average treatment effect for the overlap population (ATO)”, is hard to interpret.

For more causal inference resources:

I hold a PhD in economics from the University of Zurich. Now I work at the intersection of economics, data science and statistics. I regularly write about causal inference on Medium.