Data & decisions / 05
causal inference
Causal impact lab
Two programs that were never randomized. The naive read gets the sign wrong; three causal methods recover the true effect.
The decisionDid a program work when nobody ran an experiment, and which estimate should the business act on?
01 / Overview
The problem, and the approach.
AThe challenge
Most programs are never randomized: managers choose where to launch, and customers choose whether to join. A dashboard comparison then mixes the program’s effect with who received it, and can point in the wrong direction.
BThe approach
A loyalty price lock launched in eight manager-chosen markets and an opt-in autopay discount, with a true effect planted in the synthetic data. Difference-in-differences with an event study, a demeaned synthetic control checked with placebo runs, and propensity matching and weighting, each scored against the truth, with E-values for what unmeasured confounding could still do.
02 / The decision journey
5 decisions, one story.
Each stage asks one business question and answers it with a model you can open and change. Any card opens that stage of the live demo.
-
01
Frame
What does the dashboard say?
Both naive comparisons say the program raised churn. It cut it.
Open this stage
-
02
Diff-in-diff
Did the treated markets change more?
Within 0.03 of the truth, with the event study that earns the trust.
Open this stage
-
03
Synthetic control
What would have happened without it?
A weighted twin of untreated markets, checked with placebo runs.
Open this stage
-
04
Matching
Did the discount work, or did loyal customers join?
Matching removes measured differences, and a toggle shows what it cannot.
Open this stage
-
05
Decide
Which number goes to the business?
Every estimate against the truth, and a rule for choosing among them.
Open this stage
03 / Under the hood
How it’s built.
Algorithms
- Two-way fixed effects
- Event study
- Synthetic control
- Propensity scores (logistic)
- Nearest-neighbor matching
- Inverse probability weighting
Stack
- Python
- NumPy · pandas
- Fixed-effects regression
- scikit-learn (propensity)
- Plotly
- Streamlit
Methods, in order
- Difference-in-differences
- Event study
- Synthetic control
- Placebo tests
- Propensity matching
- E-values
Data
Synthetic weekly churn for treated and untreated markets, with a known effect of −0.40.
04 / Outcomes
What it changed.
- 01
The naive comparison said +0.16: the program looked harmful.
- 02
Difference-in-differences recovered −0.43 against a true −0.40.
- 03
A rule for which estimate goes to the business, and what each method cannot fix.
Results come from synthetic or public data with known answers. They show the method, not a client result.
05 / Try it
The live demo.
The real app, running on synthetic data. Change the inputs and watch the decision move.
Interactive, synthetic data. If nobody has used it for a while it takes about 30 seconds to wake.
Open in a new tab ↗06 / Related
Keep exploring.
Data & decisions / 06
Marketing mix optimizer
A mix model with carryover and saturation, calibrated to a geo test, moves the same budget to where it earns the most.
View project
Interactive lab / 02
Subscriber value lab
A fictional internet and mobile provider with 60,000 subscribers: who will leave, who an offer actually saves, and how far prices can rise.
View project
Applied AI / 03
Analytics copilot
Ask the warehouse a question in plain English and get governed SQL, a chart and a one-line answer, or an honest decline.
View project
Open conversations