Time-based regression is a method for analysing geo experiments that regresses the treated markets' aggregate metric on the control markets' over the pre-period, then projects that relationship through the test window. The gap between actual and projected values, accumulated over the window, is the estimated lift.
Time-based regression, usually shortened to TBR, was set out by Google researchers in 2017 as a way to analyse geo experiments. Rather than modelling each market separately, it aggregates the treatment markets into one series and the control markets into another. Over the pre-period it fits a single linear regression of the treated series on the control series. Through the test period it applies that fitted relationship to what the control markets actually did, which gives the counterfactual forecast — the path the treated markets would have taken had nothing changed — and the accumulated difference between actual and forecast is the lift.
The appeal is that the controls carry everything the treated markets share with them. A national promotion, a holiday, a heat wave or a competitor's launch moves both series, so it moves the forecast along with the actual and drops out of the difference. What remains is what happened only in the treated markets during the window, which in a well-run test is the treatment. Compared with difference-in-differences, which assumes the gap between the groups stays constant, TBR lets the treated series scale with the controls through the slope. Compared with synthetic control, it fits one coefficient on an aggregate rather than a weight per donor market, which makes it hard to overfit and easy to inspect.
Uncertainty is the part implementations differ on. The original paper used a Bayesian regression and reported a predictive interval on the cumulative effect, which widens through the test window because each day's forecast error accumulates. Meridian GeoX keeps the regression, constrains the slope to be non-negative, makes no distributional assumption about the errors, and takes its intervals from placebo tests on periods where nothing happened. Either way the interval should grow with the length of the test; one that stays flat is treating the forecast as if it were known.
It breaks in three recognisable ways. Contaminated controls: if control markets receive spillover from the treatment — overlapping broadcast coverage, or national campaigns that reach some regions harder than others — the counterfactual absorbs part of the effect and the lift is understated. A relationship that drifts: the method assumes the pre-period slope holds through the test, so a short pre-period, or one that misses a seasonal peak the test falls in, can produce a forecast that is wrong for reasons unrelated to the treatment. And aggregation hides imbalance: one large market in the treated group can dominate the aggregate, so a local disruption there reads as a campaign effect.
None of this is exotic, and that is the method's main strength. The model is a slope and an intercept, its assumptions fit on one line, and its pre-period fit can be plotted and argued over by people who are not statisticians. It is the analysis behind Meridian GeoX, and in a randomised design — where assignment rather than the model is what makes the groups comparable — that simplicity costs very little.
The pre-period fit, the counterfactual it produces, and the two quantities read off the test window.
Ȳ_treat,t = α + β · Ȳ_control,t + εₜOrdinary least squares on the aggregated daily or weekly series. Meridian GeoX constrains β ≥ 0.
Ŷ_treat,t = α̂ + β̂ · Ȳ_control,t through the test windowDriven by what the controls actually did, so shocks common to both groups cancel.
Δ = Σ ( Ȳ_treat,t − Ŷ_treat,t ) over the test windowAccumulated over the window and scaled to the treated markets' total. A go-dark test reverses the sign.
widens with the length of the test windowEach day's forecast error accumulates. A placebo run on a quiet window is the cheapest check that the interval is honest.
A subscription brand runs a five-week heavy-up of online video in 30 randomly assigned treatment DMAs, with 30 control DMAs, and reads the result with time-based regression on daily sign-ups.
The heavy-up added about 3,200 sign-ups, a 6.5% lift, at an incremental cost of $75 each, and the placebo window showed nothing.
The slope of 0.97 says the treated markets moved almost one for one with the controls before the test, which is what a randomised split should produce and means the counterfactual is close to a straight copy of the control trend. The placebo row is what makes the headline believable: on five weeks where nothing changed, the same machinery found nothing, so it is not manufacturing effects out of weekly noise. The number to act on is the interval rather than the point estimate. Dividing the $240,000 by the two ends of it puts the incremental CPA anywhere between about $45 and $218 — a range wide enough to justify scaling the channel if the target is $250, and nowhere near wide enough to call it if the target is $80. A longer window or more markets would narrow it; reading the $75 as if it were exact would not.
When your advertising efforts are up and sales are increasing, that’s great news, but it doesn't necessarily mean the ads are the reason. Here are three popular tools to measure the lift of a campaign.
Figuring out an ad's real effect is tricky. Clicks don't tell the whole story and attribution models fall short. The Solution: Geo Testing.
Applying it to a live measurement problem is the part that goes wrong. If you are designing an experiment, reading a result you do not trust, or trying to work out what your marketing actually caused, that is the work we do.