On this page
- How it works
- Statistical model
- Counterfactual and lift
- Incremental ROAS
- Why it was developed
- Experiment types
- Meridian GeoX (2026 implementation)
- Design-aware geo selection
- Design-aware placebo inference
- Key terms
- Design factors and power
- Assumptions and failure modes
- Benchmarks (as-of 2026-09-01)
- Meridian GeoX vs. unnamed alternative (vendor-run)
- Competitor landscape
- Haus (synthetic control)
- Measured (lasso selection)
- GeoLift (Meta/Facebook)
- What practitioners report
- Software implementations
- Privacy and cookieless positioning
- Industry warnings
- Related concepts
Time-Based Regression
Time-Based Regression
Time-Based Regression (TBR) is a geo-experiment analysis method that uses variation across time (pre-period vs. test period) as the counterfactual, rather than variation across geographic units. It was designed for settings where few geographic units are available — such as smaller countries, subregions, or single-market matched-market tests — where geo-based regression (GBR) is inapplicable. As of 2026 it is the default analysis engine in Meridian GeoX, Google's open-source geo-experiment framework.
How it works
Statistical model
TBR fits a linear regression on the pre-experiment period relating the treatment group's time series to the control group's time series:
Y_treat,t = α + β · Y_control,t + ε_t
where:
Y_treat,t= treatment group average KPI at time tY_control,t= control group average KPI at time tα= baseline volume difference between groupsβ ≥ 0= structural correlation between groups (non-negativity constraint for physical plausibility)ε_t= residuals, assumed independent
Parameters are estimated with OLS on pre-experiment data. The 2017 paper uses non-informative Bayesian priors (uniform on α, β, log σ), yielding a posterior for the cumulative effect that is a shifted, scaled t-distribution with n−2 degrees of freedom — no simulation required. The 2026 Meridian GeoX framework replaces classical inference with Design-Aware Placebo Inference (non-parametric) after finding t-tests produced 10× the expected false positive rate in a 192-advertiser A/A study.
Counterfactual and lift
During the test period, the counterfactual (what the treatment group would have produced without the intervention) is:
Ŷ_treat,t = α̂ + β̂ · Y_control,t
The daily average lift is δ̄_t = Y_treat,t − Ŷ_treat,t. The cumulative incremental conversions through day k are:
Δ_k = |G_T| × Σ(t=1 to k) δ̄_t
where |G_T| is the number of treatment geos. This scales group-average effects back to totals.
Incremental ROAS
Incremental ROAS (iROAS) at time t is the ratio of cumulative response effect to cumulative incremental cost. When pretest cost is a known constant, the iROAS posterior is also a scaled t-distribution. When cost is variable, the ratio of t-posteriors is nonstandard and must be simulated. Cumulative iROAS can be unstable early in the test and should be monitored for stabilization before choosing cooldown length.
Why it was developed
GBR (Vaver and Koehler, 2011) regresses across geos, requiring many geographic units for statistical power. TBR was developed specifically because GBR is "less applicable, or not applicable at all" in settings such as smaller countries, subregions, or single-market matched-market tests comparing just one test region to one control region.
Experiment types
TBR supports three experiment types (terminology from GeoX, 2026):
| Type | Treatment geos | Control geos | Use case |
|---|---|---|---|
| Holdback | Campaign withheld | Campaign runs | Measure impact of not running |
| Go-dark | Spend cut to zero | Normal spend | Measure effect of pausing (sign is negated) |
| Heavy-up | Spend increased | Normal spend | Measure iROAS of additional investment |
Meridian GeoX (2026 implementation)
Meridian GeoX is Google's open-source Python framework for geo experiments, with beta disclosure May 5, 2026 (Gaurav Bhaya blog post) and general availability September 9, 2026. TBR is the default analysis method (as-of 2026-09-09). Two alternatives are available in the same framework:
- Synthetic Control (SC): finds non-negative control-geo weights summing to 1; predictions stay within the convex hull of control geos; recommended for sparse or highly volatile data.
- Synthetic Difference-in-Differences (SDID): adds time weights and a DID correction; references Arkhangelsky et al. (2021) for asymptotic consistency.
Google's internal simulations (38,400 simulated experiments across 192 advertisers) found SC or SDID can outperform TBR when geos are bespoke and outsized, or when time series are highly noisy; TBR's efficiency declines as noise increases (as-of 2026-09-01). Both comparisons were made by an interested party.
Design-aware geo selection
GeoX defaults to stratified sampling rather than traditional matched markets. Geos are featurized on mean, coefficient of variation, linear slope, lag-1 autocorrelation, and DTW-based similarity; then clustered via K-means into strata; treatment and control are randomized within strata. Google's internal simulations found stratified sampling + TBR achieved the best overall MDE gain across scenarios (as-of 2026-09-01).
The old matched-markets approach used in-sample fit, which can underestimate live-test noise; GeoX switched to out-of-sample validation across historical windows for more buffer against volatility.
Design-aware placebo inference
Classical t-tests applied to geo experiments showed more than 10× the expected false positive rate (5% nominal level) in a large-scale A/A study. Design-Aware Placebo Inference builds a null distribution by repeatedly drawing placebo treatment groups from control geos, using the same stratified assignment procedure as the original design. Effects are studentized by dividing by out-of-sample RMSE.
Relative to random placebo selection, design-aware placebo inference reduced mean MDE by 26.8% and median MDE by 31.3% (as-of 2026-09-01; Google internal study, not independently replicated).
Key terms
| Term | Meaning |
|---|---|
| TBR | Time-Based Regression; uses time (pre vs. test period) as the counterfactual dimension |
| GBR | Geo-Based Regression; uses cross-geo variation as the counterfactual |
| iROAS | Incremental return on ad spend; ratio of cumulative response effect to cumulative incremental cost |
| MDE | Minimum Detectable Effect; the smallest lift the design has sufficient power to detect |
| Holdback | Experiment where the control geos do not receive the campaign |
| β | Structural correlation coefficient between treatment and control time series; constrained ≥ 0 |
| Stratified sampling | Geo assignment: featurize → cluster → randomize within strata |
| Design-aware placebo | Non-parametric inference using the same assignment mechanism as the original design |
Design factors and power
The 2017 TBR paper documents the following design sensitivities:
- Spend intensity is the most influential factor. Scaling intensity by factor f scales the credible-interval half-width by 1/f.
- Longer pretest narrows uncertainty at roughly 1/√n but cannot shrink below a floor set by σ₀/(c̄√T), where σ₀ is baseline noise, c̄ is mean incremental cost per period, and T is test length.
- Longer cooldown with fixed budget widens the CI (grows roughly with √T) because spend is spread more thinly.
- Number of geos has no direct effect on precision, but very few geos worsen fit and raise residual variance.
- Worked example: 210 randomly assigned US DMAs, 8-week pretest, 4 intervention weeks, 1-week cooldown.
Assumptions and failure modes
TBR assumes the linear relationship Y_treat = α + β·Y_control holds throughout both the pre-period and the test period. Known failure modes:
- Low treatment-control correlation (<0.5) combined with exponential baseline growth: coverage was systematically low in simulations.
- Strong seasonality or trends in the relationship between treatment and control: violates the stability assumption.
- Measurement error in the control series: degrades TBR performance; an errors-in-variables (orthogonal regression) extension (TBR-OR) is discussed in the 2017 paper but is "less stable when treatment-control correlation is near zero."
- Structural-break tests (Zeileis et al., 2002) can flag potential violations before the experiment runs.
TBR estimates an intention-to-treat effect, equated to the ATT at the geo level. Users within treated geos may not all see the ads, so individual exposure cannot be fully controlled. Matched (non-randomized) geo designs may also generalize less well than randomized ones.
Benchmarks (as-of 2026-09-01)
Meridian GeoX vs. unnamed alternative (vendor-run)
A benchmark across 69 studies (40 enterprise, 29 SMB), comparing GeoX against an unnamed open-source augmented synthetic control library:
| Metric | All studies | Enterprise |
|---|---|---|
| Better out-of-sample RMSE | 73% of studies | 80% |
| Median RMSE improvement | 71% (IQR 25–87%) | — |
| Lower weekly budget | 64% of studies | 83% |
| Median budget savings | $23k vs $28k (~17%) | ~31% (p=0.01) |
Competitor landscape
Haus (synthetic control)
Haus uses Synthetic Control Method|synthetic control rather than TBR as its primary geo holdout method and claims synthetic control produces results "4× more precise than matched market tests" — with no primary methodology or sample provided for this figure. Haus also offers "Time Tests" for measuring incremental impact when geo splits are not feasible; the methodology is not publicly disclosed and may or may not use TBR.
Measured (lasso selection)
Measured uses three prediction models for geo testing — linear regression, lasso, and flat average — selecting whichever achieves best A/A validation, with lasso most commonly selected. Results are based on actual sales from the brand's transaction source (not platform-reported conversions), making attribution window settings irrelevant.
GeoLift (Meta/Facebook)
GeoLift uses Synthetic Control Methods to estimate incremental lift at the geo level and includes power calculators and market-selection tools; it does not use TBR.
What practitioners report
The following findings come from Haus CSO Olivia Kory via podcast appearances (not TBR-specific but relevant to practitioners choosing between methods):
- Minimum spend thresholds: Incrementality testing becomes operationally useful at approximately the low-to-mid eight-figure annual paid marketing stage (~$5M+).
- Test duration: YouTube tests require 4–6 weeks minimum; view-based channels (YouTube, CTV) need 6–8 weeks because much of the lift appears late or in the post-treatment window.
- YouTube incrementality factor: From 190 geo holdout tests across 74 brands, YouTube showed a 3.42× incrementality factor versus platform-reported conversions (as-of publication date — unknown).
- Measurement scope: Including Amazon and retail sales in geo holdout measurement roughly doubles the measured lift from YouTube vs. D2C-only measurement.
- Creative as lever: Creative is the biggest lever for improving incrementality in paid channels; account structure and traffic composition are secondary.
- Privacy positioning: Geo holdout tests avoid user-level tracking entirely, making the method privacy-resilient and applicable across all channels after iOS 14.5 signal loss.
(eCommerce Evolution Podcast ep. 339; Marketing Intelligence Show, Supermetrics — exact dates unknown)
Software implementations
| Tool | Language | Method | Status |
|---|---|---|---|
| GeoexperimentsResearch | R | TBR + GBR | Archived 2022-12-29 |
| Meridian GeoX | Python | TBR, SC, SDID | Active (GA 2026-09-09) |
| GeoLift (Meta) | R | Synthetic Control | Active |
Privacy and cookieless positioning
TBR and GeoX are described as cookieless and publisher-agnostic: they measure aggregate regional outcomes and require no user-level identifiers, making them viable in a post-cookie measurement environment.
Industry warnings
CIMM (Coalition for Innovative Media Measurement) warned on 2026-07-31 that default assumptions in open-source geo-experiment frameworks (including TBR defaults) can have material budget consequences for advertisers who adopt them without adjustment (as-of 2026-07-31; primary CIMM document not retrieved).
Related concepts
- Incrementality Testing — parent methodology
- Geo-Based Regression — predecessor method (many geos, cross-sectional)
- Marketing Mix Modelling — GeoX v2.0.0 (September 2026) enables TBR-derived causal estimates to feed into MMM as priors
- Synthetic Control Method — alternative analysis in GeoX; Haus's default
- Design-Aware Placebo Inference — inference method replacing t-tests in GeoX
- Meridian GeoX — 2026 open-source Python implementation
- Haus — measurement vendor using synthetic control instead of TBR
- Minimum Detectable Effect — key design output
- Incremental ROAS — primary output metric
- Holdback Test — most common TBR experiment type
- Difference-in-Differences — related econometric method
- Conversion Lift Studies — alternative (platform-native) incrementality method
References
- Kerman, Wang, Vaver, Google Research, 2017 — GeoX documentation, Google, 2026 — https://developers.google.com/meridian/geox/time-based-regression — storage.googleapis.com/pub-tools-public-publication-data/pdf/45950.pdf
- Google Meridian GeoX Developer Docs, 2026 — — developers.google.com/meridian/geox/time-based-regression
- Kerman et al., 2017 — — research.google/pubs/pub45950
- PPC Land / Luís Rijo, 2026 — — ppc.land/meridian-geox
- Meridian GeoX framework paper, Google Research, 2026-09-01 — PPC Land / Luís Rijo, 2026 — https://ppc.land/meridian-geox/ — storage.googleapis.com/gweb-research2023-media/pubtools/1043733.pdf
- Measured / Trevor Testwuide, 2025-03-05 — — www.measured.com/blog/what-do-you-know-about-geo
- Meta GeoLift GitHub — — github.com/facebookincubator/GeoLift