← Domains

Causal inference &
experimentation

Identifying a pattern in data is not the same as finding a lever you can pull. We find the lever.

A/B Testing Diff-in-Diff Synthetic Controls Power Analysis Observational Methods

Overview

A correlation in your dashboard is not a decision you can act on. Acting on the wrong one is expensive — in wasted spend, in misdirected strategy, in the credibility of your data function.

Correct causal analysis starts with a clearly stated question: what intervention, on what population, compared to what counterfactual. When randomisation is possible, we design experiments with proper power analysis up front, so a null result actually means something and interference or network effects between units don't quietly bias the estimate.

When randomisation isn't feasible — the intervention already happened, or can't ethically or practically be randomised — we reach for the right observational method: difference-in-differences, synthetic control, regression discontinuity, or instrumental variables, each with its own identification assumptions that we test rather than assume.

The output is always a defensible answer to "what happens if we pull this lever" — with the assumptions and their limits stated plainly, not buried in a footnote.

01

Causal question & identification strategy

Stating precisely what intervention, population, and counterfactual are in scope. Determining whether randomisation is feasible, and if not, which observational identification strategy fits the data-generating process.

02

Design & power analysis

For experiments: minimum detectable effect, sample size, and duration calculated up front so a null result is informative. Checking for interference, network effects, and contamination between treatment and control before launch.

03

Execution & monitoring

Randomisation integrity checks, sample-ratio-mismatch monitoring, and pre-registered analysis plans to avoid post-hoc metric shopping. For observational designs, construction of the comparison group and pre-trend validation.

04

Estimation & assumption testing

Estimating the treatment effect with the method matched to the design, and explicitly testing — not assuming — the identifying assumptions behind it: parallel trends, exclusion restrictions, or randomisation balance as applicable.

05

Interpretation & decision handoff

Translating the estimated effect, its uncertainty, and its limits into a decision recommendation stakeholders can act on — stated in terms of the business outcome, with the caveats that matter and none that don't.

Treatment vs. control, before and after

Intervention Treatment group Control group

Parallel pre-intervention trends, then divergence — the shape that makes an effect estimate defensible.

Need to know if something actually worked?

Tell us about the decision you're trying to make and we'll scope it together — no obligation.

Start a conversation →