# Worked Example: Collaboration Feature and Retention

## Decision Context

Decide whether to invest in surfacing collaboration prompts during the first week of a new workspace.

## Estimand

- Treatment: using the collaboration feature during week 1
- Control: not using the collaboration feature during week 1
- Population: new eligible workspaces
- Outcome: retained at 30 days
- Time horizon: signup through day 30
- Target estimand: ATT for workspaces that adopted collaboration in week 1

## Causal Model

DAG in words:

```text
Acquisition channel -> Collaboration use
Acquisition channel -> Retention
Plan -> Collaboration use
Plan -> Retention
Team size -> Collaboration use
Team size -> Retention
Baseline sessions -> Collaboration use
Baseline sessions -> Retention
Invites sent -> Collaboration use
Invites sent -> Retention
Collaboration use -> Retention
Support contact <- Collaboration use
Support contact <- User confusion -> Retention
```

Confounders:

- acquisition channel
- plan
- team size
- baseline sessions
- invites sent before collaboration use

Potential colliders or post-treatment variables:

- support contact after feature use
- project activity after feature use
- paid conversion before day 30 if affected by collaboration

## Identification Strategy

Primary strategy: inverse-propensity weighting for the ATT, conditioning only on pre-treatment covariates.

Backup strategy: matching with caliper on propensity score and exact matching on plan where feasible.

Preferred future strategy: randomized collaboration prompt exposure, using prompt assignment as an instrument for feature adoption if exclusion is plausible.

## Assumptions

Exchangeability:

After conditioning on channel, plan, team size, baseline sessions, and early invites, potential 30-day retention is independent of first-week collaboration use.

Positivity:

Comparable workspaces must exist with and without collaboration use across the covariate distribution.

SUTVA/interference:

Workspace-level treatment is acceptable because collaboration occurs inside the workspace. User-level analysis would violate interference assumptions.

Measurement:

Feature-use events and retention events are logged consistently across channels and plans.

## Diagnostics

- Balance before and after weighting
- Propensity score overlap
- Extreme weights
- Retention placebo outcome measured before collaboration exposure, if available
- Negative-control outcome unrelated to collaboration
- Leave-one-covariate-family-out sensitivity

## Interpretation

If the weighted ATT is positive, the claim should be:

> Among new eligible workspaces similar to observed adopters, first-week collaboration use is associated with higher 30-day retention under a no-unmeasured-confounding assumption.

This should not be reported as proof that collaboration causes retention for all workspaces. Large enterprise teams and workspaces with weak overlap should be excluded from the headline estimate or analyzed separately.

## Strongest Critique

Motivation and team intent are only partially measured. Workspaces that choose collaboration may have higher latent intent to retain even after observed adjustment. The result is useful for prioritizing a randomized prompt experiment, but not sufficient by itself for a confident causal claim.