Attributing A Subscription Launch Lift Without A Clean Control Cohort
Subscribers self-select for loyalty, so the post-launch cohort curve always looks better. Here's how to back out the true incremental lift using propensity matching and synthetic controls.
Quick answer
If you launched subscription without a randomized holdout, the raw subscriber-vs-everyone-else lift is almost entirely self-selection. Build a propensity-matched non-subscriber cohort from pre-launch behavior (RFM, AOV, category mix) and compare its forward 90-day curve to the subscriber cohort's. If matching isn't feasible at the customer level, use a synthetic control on pre-launch weekly revenue. The gap between the two curves is your real lift.
Subscription launch attribution without a control cohort
Estimating the incremental lift from a subscription launch when no randomized holdout exists, using propensity matching or synthetic controls to approximate counterfactual behavior.
When a subscription program launches without a pre-designed holdout, you cannot directly observe what subscribers would have done without the program. The cohort that opts in is structurally different from everyone else — higher purchase frequency, higher AOV, deeper category affinity — so any raw comparison overstates impact. The job is to rebuild a counterfactual: a group of non-subscribers who looked identical to subscribers immediately before launch, then measure how their forward behavior diverges. The two workhorse methods are propensity-matched cohorts (customer-level) and synthetic control (time-series-level). Both estimate the same thing — incremental lift net of self-selection — under different assumptions.
The trap is the chart your CEO is already looking at: subscriber 90-day repeat rate of 62% versus 24% for the rest of the base. It feels like a 38-point lift. It is not. Most of that gap existed before anyone subscribed to anything.
Sign-up correlates with intent. Customers who choose a subscription have already decided they want the product on a schedule. The same customers, without the program, would still have repeated at 50%+. The honest question is what fraction of the gap subscription itself created.
Why the raw post-launch cohort curve lies
Subscription programs are a self-selection filter. The opt-in step screens for customers who already buy more often, spend more per order, and skew toward replenishable categories like skincare, supplements, pet food, or coffee. Their forward curve was always going to be steeper.
Quantifying that filter is its own exercise — see how much of subscriber repeat rate is mechanical self-selection rather than program impact. On most DTC brands we look at, 60-85% of the raw lift evaporates once you condition on pre-launch behavior. The remaining 15-40% is the real number worth defending to a CFO.
The CEO chart problem
If the only artifact circulating is 'subscribers repeat at 62%, non-subscribers at 24%', you have a comms problem before you have a measurement problem. Replace it with the matched-cohort chart before someone uses the inflated number to justify a budget commitment.
The propensity-matched non-subscriber playbook
Propensity matching builds a synthetic twin for every subscriber from the non-subscriber pool. The twin is a customer who, on the day before subscription launched, looked statistically indistinguishable on the features that predict opt-in: recency, frequency, monetary value, category mix, tenure, and channel of first purchase.
Fit a logistic regression or gradient-boosted model predicting subscription opt-in from pre-launch features. Score every customer. For each subscriber, find the nearest non-subscriber on the propensity score (1:1 or 1:3 matching, caliper of 0.05 SD). The matched non-subscribers are your control cohort.
Then run the same 90-day forward analysis on both cohorts: repeat rate, revenue per customer, orders per customer. The difference is your propensity-matched lift. This is the spine of any defensible incrementality memo, and the method most CFOs will sign off on when a clean holdout wasn't designed up front.
What the numbers actually look like
Raw vs propensity-matched 90-day lift across DTC verticals (illustrative)
| Vertical | Raw lift (subs vs all non-subs) | Matched lift (subs vs twins) | Self-selection share |
|---|---|---|---|
| Coffee & beverages | +34 pts repeat rate | +8 pts repeat rate | 76% |
| Skincare & beauty | +41 pts repeat rate | +11 pts repeat rate | 73% |
| Supplements & wellness | +38 pts repeat rate | +14 pts repeat rate | 63% |
| Pet food & treats | +45 pts repeat rate | +18 pts repeat rate | 60% |
| Apparel basics | +22 pts repeat rate | +3 pts repeat rate | 86% |
Apparel is the cautionary tale. The raw subscriber curve looks compelling, but matched lift collapses to near zero — subscription mostly attracts customers who were going to repeat anyway. Replenishable categories like pet food retain a real durable lift even after matching, which is what you'd expect from a category where the program solves a genuine convenience problem.
When matching isn't enough: synthetic control
Propensity matching fails when your non-subscriber pool is thin, when opt-in features aren't observable, or when subscription was rolled out to a whole market segment at once. In those cases, switch to a synthetic control built on pre-launch weekly revenue (or orders) of the affected cohort versus a weighted combination of unaffected geos or customer segments.
The choice between methods isn't ideological. Customer-level matching is sharper when you have rich pre-launch behavioral data; synthetic control is more honest when the unit of treatment is a market or a segment. Many teams run both and report the range — and if a geo holdout was the only viable design, that's a legitimate fallback rather than a compromise.
Novelty vs durable lift, and the zero-lift case
Even a clean matched lift in the first 30 days can be novelty — subscribers buying extra in the excitement of joining, then reverting. Split the 90-day window into 0-30, 31-60, and 61-90 buckets and check whether the gap holds. Durable lift looks flat or growing in the back half; novelty lift decays toward zero.
And occasionally the matched lift is zero or negative. That's not a measurement failure — it's the answer. If the program isn't driving incremental behavior, the right move is to model what killing the subscription program looks like, including the discount margin you'd recover. That conversation is easier when the matched-cohort number is already on the table.
Designing for the next launch
If you're reading this before the launch happens, do the obvious thing and design a randomized subscription holdout up front. Hold out 10-15% of eligible customers from the program for 90 days. The arithmetic becomes trivial, the memo writes itself, and you avoid every method on this page.
If launch is already live, set up the matched-cohort pipeline this week — propensity scores get harder to estimate cleanly the further you drift from launch day, because subscribers start influencing non-subscriber behavior through referrals and word-of-mouth. Lock the pre-launch feature snapshot now.
Frequently asked questions
Because the two groups were already different on the day before subscription launched. Subscribers had higher pre-launch frequency, AOV, and category engagement. A direct comparison measures who opted in, not what the program did. The honest measurement requires matching subscribers to non-subscribers who looked identical before anyone subscribed.
Pre-launch RFM (recency, frequency, monetary), AOV, tenure, number of distinct categories purchased, share of replenishable SKUs, channel of first purchase, email engagement, and discount usage. Five to ten features is typical. Avoid post-launch variables — those leak the treatment into the score.
1:1 matching maximizes balance but discards data. 1:3 matching with a tight caliper (0.05 SD on the propensity score) is the usual compromise — better statistical power, still credible balance. Always check covariate balance after matching; if standardized mean differences exceed 0.1 on any feature, retighten the caliper.
Synthetic control wins when the treatment unit is a market, region, or segment rather than an individual customer, or when pre-launch behavioral features for non-subscribers are sparse. It's also more robust when subscription was rolled out simultaneously to everyone eligible. The two methods answer the same question with different assumptions.
Snapshot pre-launch features the day before launch and you can match indefinitely. If you didn't snapshot, you can reconstruct features from order history up to about 6 months post-launch before referral spillover and changed buying patterns contaminate the non-subscriber pool. After that, synthetic control on aggregate revenue is more defensible.
That's normal — typically 15-40% of the raw lift survives matching. If it survives, you have a real program. If matched lift is near zero across a 90-day window, the program is mostly a discount being given to customers who would have repeated anyway. That's a margin question, not a measurement question.
Split the post-launch window into 0-30, 31-60, and 61-90 day buckets and compute matched lift in each. Durable programs show flat or growing lift in the back half; novelty-only programs decay toward zero by day 60. Report all three buckets in the memo so finance sees the shape, not just the headline.
Yes, if subscription is launching market-by-market. Hold one comparable market back by 60-90 days and compare cohort curves between treated and untreated geos. It's lower-precision than customer-level matching but doesn't require pre-launch behavioral features and is easier to explain to non-technical stakeholders.
Lead with three numbers: raw lift, matched lift, and the share that's self-selection. Show the 0-30, 31-60, 61-90 split. Add one paragraph on the matching method and one on assumptions. Keep the propensity-score diagnostics in an appendix. The memo should fit on two pages.
Propensity matching builds individual twins for each subscriber from the non-subscriber pool, using pre-launch behavioral features. Synthetic control builds a weighted average of untreated time series (geos, segments) that tracks the pre-launch trajectory of the treated cohort. Matching is sharper when you have rich customer data; synthetic control is more honest when treatment is at the market level.
Test ideas before you ship them
Run unlimited A/B tests, attach hypotheses to outcomes, and build a searchable archive of what works — and what doesn't.