# Amendment for a noncomparable positive event item

This amendment is registered after inspecting input quality and the first mechanical run, but before fitting the corrected four-category analysis. Original plan.md is unchanged. Initial registration: prereg:2ff1e719addc0adafdcaff907948938ab9c7cd9a95edf75645fdedacd90ba9a7, ledger239.

The five-category complete-case requirement produced only29 M2 training participants across30occasions, below the planned feasibility threshold. The later wave had528eligible people and789occasions. The first mechanical run did produce performance metrics before this sample-size problem was inspected: joint-vs-baseline MSE reduction was -23.3889% in M2 and -5.5613% in M3. Those preliminary metrics cannot answer the intended adequately trained comparison and are not the definitive analysis. Their existence and direction are disclosed so this amendment is not described as wholly outcome-blind.

The issue is independently visible in the supplied input metadata and frequencies: B2DF11, positive events involving a friend or relative, has1281yes,13631inapplicable(9),1264missing(8), and no recordedno(2). Other common positive-event items have normalyes/no distributions. We cannot justify treating all9 values asno. The complete-case rule consequently selects repeated yes responses and makes the five-category predictor unsuitable for the intended harmonization.

Replace positive-event count with the FOUR unambiguous common categories: social interaction, work/volunteering, home, and other. Use M2 DF8,9,10,12 and M3 DF8,9,10,13; omit network events from BOTH waves. Continue to require valid1/2 answers on every included category, and do not recode9 to zero. All other outcome rules, families/folds, models, penalty, bootstrap, forecast timing, tests, benchmarks, and sensitivity analyses remain unchanged. No model is selected using preliminary performance. Do not make confirmatory inference unless the M2 training source and evaluated cohort both meet the original50people/200occasions threshold.

The first implementation also halted on supplied mood-score discrepancies. Fully answered item sets match supplied averages. Some partial-answer averages differ, with negative-affect discrepancies consistent with treating code8 as numeric. The originally planned raw-item reconstruction remains unchanged; supplied composites are only a validation comparison. Do not interpret unverified composite causes as an empirical claim. Report that discrepancy as an implementation note.

The amended analysis is preregistered relative to its own corrected run, but the first outcome-bearing attempt was seen. Transparently label this study as amended after a data-quality diagnosis rather than presenting an unqualified untouched preregistration.