1. Data Ingestion
After the treatment window closes (including any planned lag), we append new data to the original panel, allowing the synthetic-control model to analyze the complete pre- and post-period as a unified dataset.
2. Synthetic Baseline Generation
We process the fresh time series through the synthetic-control model to create a counterfactual using the same selection criteria in GeoMatch Design.
3. Effect Calculation
Day-level lift is calculated as
We aggregate these values over the treatment window to determine total incremental revenue/conversions and percentage lift.
4. Statistical Inference via Permutation Testing
Residuals form a single vector spanning both pre0- and post-periods.
Our engine automatically detects autocorrelation using Durbin-Watson and Ljung-Box tests.
If no autocorrelation is found, residuals are shuffled independently; otherwise, they're permuted in blocks (weekly by default).
For each of 10,000 permutations, we sum the "fake" treatment window residuals to build a null distribution.
The two-sided p-value equals the proportion of permuted sums whose absolute value equals or exceeds the observed sum.
Results are considered significant when p < 0.05 (this threshold is configurable).
5. Confidence Interval Estimation
A 10,000-draw bootstrap on day-level lifts generates the 5% and 95% quantiles, establishing our 90% confidence interval.
6. Business Metric Derivation
Incremental ROAS = Incremental Revenue / Incremental Spend (or inverse for spend-cut tests).
We compare iROAS against the target that informed the original MDE.
Both absolute and percentage changes are presented per-day and cumulatively.
7. False-Positive Protection
We run the same permutation engine on a placebo window immediately preceding the actual treatment period; an unusually low p-value here signals potential bias not captured during design.
Important Outputs:
Significance: A significant result is one where the observed effect is unlikely to have occurred by random chance (typically p < 0.05), indicating the treatment had a real impact.
Absolute Lift: The increase or decrease in our output in raw value since the start of our experiment.
Percentage Lift: The increase or decrease in our output as a percentage since the start of our experiment.
Calculated iROAS/CPIC/CPA: The computed metric we calculated after the experiment.
Confidence interval: A range of values that has a specified probability (typically 90%) of containing the true effect. It quantifies the uncertainty in our estimates due to sampling variability.