Causal inference · Experiments
Lift you can defend.
Run geo lift tests when you can't randomise users. Run A/B properly when you can. Send the result back into the next MMM run as a Bayesian prior — one click. Same tooling marketing-science teams build by hand, productised.
GeoLift / augmented SCM·PyMC Bayesian A/B·CUPED variance reduction·EconML HTE (advanced)
What's in the box
From design to readout, in the same workspace.
01
Geo experiment design
Pick treatment and control regions on an interactive UK / US map. Power analysis, MDE estimation, donor-pool diagnostics — before you spend a penny.
02
Augmented synthetic control
GeoLift under the hood: augmented SCM (Ben-Michael, Feller & Rothstein, 2021) with placebo diagnostics, pre-period fit reporting, and synthetic-DiD as an optional method.
03
A/B analysis
Frequentist baseline (statsmodels), Bayesian alternative (PyMC), CUPED variance reduction, multiple-comparison guardrails. Show calibration, not just p-values.
04
Heterogeneous effects
Causal Forests (Wager & Athey, 2018) and Double / Debiased ML (Chernozhukov et al., 2018) via EconML. Gated as advanced — easy to misinterpret, surfaced with caveats.
05
Send to MMM as prior
Click. The lift estimate attaches as a Bayesian prior on the relevant channel coefficient in the next MMM run. The MMM treats it as a calibration target with the credible interval.
06
Pre-registration
Lock the analysis plan before the test runs. Reduces the temptation to p-hack — and gives the readout credibility marketing-science teams trust.
Methodology
The canonical references.
Synthetic Control
Abadie, Diamond & Hainmueller — 2003 / 2010 / 2015
Foundational method. Understand the failure modes before you ship a result.
Augmented Synthetic Control
Ben-Michael, Feller & Rothstein — 2021
Robustness improvement on classical SCM. What GeoLift uses internally.
CausalImpact
Brodersen et al. (Google) — 2015
Bayesian structural time series. Available as a cross-check method.
CUPED variance reduction
Deng, Xu, Kohavi & Walker (Microsoft) — 2013
Massively reduces required sample sizes for unit-level A/B.
Causal Forests / GRF
Wager & Athey — 2018
Heterogeneous treatment effect estimation with CIs. Advanced mode only.
FAQ
Experiments, in detail.
Because most of your media spend is on broad-reach channels — TV, OOH, big-bid digital — where you can't randomise at the user level. Geo experiments use spatial variation as the randomisation. Augmented synthetic control gives you a credible counterfactual when classical pre/post DiD breaks down.
✦ Get started
Bring your spend, leave with a model you can defend.
Book a 30-minute working session with the Kleyr team. We'll model your last 12 weeks of marketing spend on the call — no slides, no abstract demo.