Evaluate an incentive offer from the people offered it

An incentive can appear effective because the people who claim it were already the most likely to participate. We would evaluate an offer from the moment it is assigned, keeping every eligible recipient in the comparison. That answers a practical question: what changes when we offer this program to this population?

The DXAP points page describes volume points and referral benefits. Those published mechanics give participation research a concrete setting. The experiment below proposes a separate feedback offer; its allocation and recognition terms would require explicit authorization before use.

Randomize the offer before anyone responds

Enroll 200 eligible owners who have consented to a four-week feedback study. Randomly assign 100 to a structured review invitation with a fixed recognition offer and 100 to the same invitation without that additional offer. Keep access, task instructions and support availability equal. Record assignment before opening the invitations and preserve declined or ignored invitations in the analysis.

The primary outcome is whether an owner supplies at least one distinct, reproducible instruction-carryover case within 28 days of assignment. A reviewer who cannot see assignment applies the same rubric to both groups. DXAP's instructions guide provides the behavioral boundary: confirmed instructions carry into future turns while configured policies retain their own role. A report needs enough authorized context to establish which instruction was active.

In an illustrative result, 24 of 100 offered owners and 15 of 100 comparison owners produce a qualifying case. The intention-to-treat difference is 24% minus 15%, or nine percentage points. Suppose only 40 offered owners claim the recognition. Dividing 24 by 40 gives 60%, but comparing that selected subset with all 100 comparison owners changes the question and loses the randomized comparison.

What the estimate identifies

Assignment estimates the effect of offering the entire package. It combines attention to the invitation, perceived recognition and willingness to submit. Isolating the effect of actually claiming would require additional assumptions, including whether the offer changes behavior through channels other than the claim. We would report the offer estimate first and show the claim rate as a mechanism measure.

Randomization needs a unit that contains likely spillovers. If owners share an invitation inside a close referral group, assign whole groups or measure exposure across arms. Cluster assignment reduces the number of independent units, so the uncertainty calculation must reflect groups rather than treating every owner as independent.

Publish confidence intervals, missing review outcomes and the number of distinct failure families beside the difference. Freeze the rubric and observation window before outcomes arrive. An offer that increases submissions while producing mostly duplicate cases may solve an attention problem without improving engineering coverage. That is a useful diagnosis: revise the review task, then test the revised offer with a fresh population. Predictive accuracy and trading economics remain separate outcomes requiring their own evaluation.