I Designed a Delete-and-Repost Survey to Survive Regression to the Mean

The pair is eligible only when the comparison is fair enough to read

An eligible pair contains one public original and one later upload of the same underlying concept on the same account and platform. The creator must be able to document the original even if it was deleted or made private. Both uploads need the same observation window, such as the first 24 or 168 hours, and the window must begin at each upload's publication time.

Paid promotion, known upload corruption, missing audio, accidental private visibility, cross-platform reposts, and mismatched windows belong in separate analyses or the exclusion log. A technical correction within minutes of publication does not answer the same question as a strategic repost after a creator judged performance.

Cap submissions per creator and model repeated pairs as clustered observations. One agency running the same workflow fifty times should not count as fifty independent creator decisions.

What the survey records for each pair

Changes are bundles until the data proves otherwise

The creator can select every changed element, but the analysis must resist turning a checkbox into an effect. A repost with a new first frame, shorter runtime, different audio, new caption style, and evening publication tests the bundle. It does not show that the first frame caused the outcome.

Create mutually intelligible groups before looking at results: exact-file retry; metadata-only change; opening-only structural change; broader recut; new recording of the same concept; and mixed change. Preserve the detailed checkboxes for exploratory analysis, but correct for multiple comparisons and label them exploratory.

The exact-file group is not a perfect no-treatment control. Reposting is itself a new upload at a new time, and the group is still selected after an underperforming first result. It can help describe retry variance and rebound, but it cannot isolate pure algorithmic randomness.

Regression to the mean is a design problem, not a footnote

If creators enter the study because the original fell far below baseline, the sample is selected on an extreme value. A second observation will often be less extreme even without an effective intervention. Comparing only original and repost exaggerates the temptation to attribute that rebound to the change.

The study should retain several pre-event comparable uploads, contemporaneous videos that were not reposted, and repeated exact-file or minimal-change cases where available. A hierarchical model can estimate creator-level baselines and platform/date variation while keeping uncertainty wide enough to reflect the data. Sensitivity analysis should test different baseline definitions and exclude pairs whose 'flop' label was not extreme relative to the registered history.

No statistical adjustment repairs every unobserved difference. The honest output is an association: repost bundles of a declared type were followed by a stated distribution of changes in this eligible sample.

Outcome reporting without arbitrary winners

Continuous measures remain primary. Creator-friendly labels are secondary and must be sensitivity-tested.

Recruitment and verification

Recruit the most recent eligible pair, not the creator's best repost story. The form should ask for failures and neutral outcomes as explicitly as wins. Publish the number screened, excluded, withdrawn, and accepted at each evidence tier.

Self-reported pairs can inform motivations and beliefs, but primary outcome models should use records where matching fields can be verified. Ask creators to redact audience identities and unnecessary account information before upload. Strip file metadata on ingestion and set a short, declared deletion schedule for screenshots and exports.

Use phased targets. First run a feasibility pilot to discover missing fields and platform mismatches. Then simulate the precision of the planned comparisons using pilot variance. A round recruitment goal such as 250 or 600 pairs is an operational scenario, not a power guarantee until the assumptions are published.

Platform policy belongs in the risk box, not the causal model

No first-party source reviewed for this protocol says that one thoughtful repost automatically triggers a universal penalty. The risk depends on the content, pattern, account, monetization program, and current rules. The page therefore does not repeat the original memo's invented strike counts or guaranteed consequences.

YouTube's current spam policy prohibits manipulative, high-volume, repetitive, scraped, and detection-evasion behavior, while its channel monetization policies separately address inauthentic and reused content. TikTok's Creator Rewards requirements emphasize original, high-quality material for eligible videos. Instagram's recommendation guidance lists largely repurposed content without material value among low-quality publishing signals.

Those documents are not evidence that deleting and reposting improves or harms the second upload. They explain why a research article must not encourage spam, content theft, minor changes designed to evade detection, or repeated attempts to manipulate distribution.

Publication rules

• Report the eligibility flow, evidence tiers, missingness, baseline rule, matched windows, and every outcome definition.
• Publish exact-file, metadata-only, opening-change, recut, rerecord, and mixed bundles separately where sample size permits.
• Show negative, neutral, and inconclusive pairs alongside wins.
• State that a repost is a different upload shown under different conditions and the study is not a randomized experiment.
• Do not name creators or quote authored comments without explicit publication permission.
• Link current first-party platform policies and date the verification instead of preserving stale enforcement claims.

Sources and statistical foundations

The regression-to-the-mean guardrail is informed by the methodological review Assessing regression to the mean effects in health care initiatives. The domain differs, but the selection problem is the same: units enter after an extreme baseline value and later improvement can be misattributed to an intervention.

YouTube's current recommendation guidance says an individual video's underperformance does not automatically penalize the whole channel and directs creators to their own audience and retention data: Good to know about recommendations.

For Shorts-specific fields, YouTube's Content tab analytics guide defines shown in feed and how many chose to view. Those platform fields should be stored under their own definitions rather than merged into generic engagement.

Instagram states that recommendation eligibility does not guarantee recommendation in its official eligibility guidance. That distinction is central to interpreting every pair.