Short‑Term vs Long‑Term Effects

• Consumer: 6‑month retained buyers, subscription renewal, churn.

• Marketplace: 90‑day driver/merchant retention, supply stability, profitability.

• SaaS: expansion revenue, net dollar retention, seat growth.

“For a user to be a 6‑month retained buyer, what do they do in the first 1–4 weeks?”

• 6‑month retention → returning in week 2, 3, 4; number of sessions; number of successful “core actions”.

• LTV → early basket size, diversity of use cases (e.g., ordering multiple categories, using multiple surfaces).

• Marketplace health → early earnings stability for supply, cancellation experiences for demand.

Treatment → early behavioral change → structural change in relationship → long‑term outcome.

• “New onboarding checklist → more users complete 3 key setup steps → more users reach ‘aha’ moment → higher 3‑month retention.”

• “Better restaurant recommendations on home → more first orders from high‑quality merchants → better first‑week NPS → higher 90‑day order frequency.”

1. Proximal: They occur soon after treatment (days or weeks, not quarters).

2. Mechanistic: They are directly on the causal path you described.

3. Stable: Their relationship to the long‑term metric is reasonably stable across time and cohorts.

• Onboarding / activation
• Long‑term: 90‑day retained active users.
• Proxy candidates:
• “Completed 3 key actions within first 7 days.”
• “Returned in week 2 and week 3.”
• Why they work: The mechanism is “users who understand value and build habit early are much more likely to stay.”

• Subscriptions / paywalls
• Long‑term: 6‑month subscriber retention, ARPU.
• Proxy candidates:
• “Number of paywalled articles read in trial.”
• “Days with any consumption in first 14 days.”
• Why they work: These capture engaged subscribers vs. those who churn as soon as friction appears.

• Marketplaces
• Long‑term: sustainable supply, cross‑side retention.
• Proxy candidates:
• For drivers: “earned at least $X per online hour in first 2 weeks”, “fraction of sessions with unacceptable idle time.”
• For customers: “experienced at most Y bad ETAs or cancellations in first N orders.”
• Why they work: Income and experience early in the lifecycle shape whether participants stick around.

• Clicks on a shiny surface with no clear path to habit or revenue.

• Session length when the product is supposed to make users faster.

• Email opens when the real value is downstream purchases.

1. Retrospective cohort analysis
• Take past users, bin them by proxy value in early period (e.g., “0 core actions”, “1–2”, “3+”).
• Plot long‑term outcomes by bin (retention, LTV, churn).
• You want strong, monotonic differences that are robust to simple controls (cohort, region, device).

2. Simple causal checks
• Use observational tools (e.g., regression, matching) to see whether the proxy still predicts long‑term outcomes after controlling for obvious confounders like tenure, acquisition channel, geography.
• You are not trying to publish a paper, just to check the relationship is not entirely spurious.

3. “Proxy experiments”
• Sometimes you can run experiments whose primary goal is to move the proxy, then check long‑term follow‑up for units in those experiments.
• If interventions that reliably move the proxy also reliably move the long‑term metric in the expected direction (when measured months later), your confidence in that proxy increases.