This page was translated automatically from Turkish. Read the original

This page is an AI-assisted Turkish summary of the study presented at the 26th World Congress of Psychiatry. The study belongs to the presenting authors; please refer to the original source when citing it.

SO466

A confidence-tiered framework for reconstructing imperfect multiwave adolescent mental health cohorts

A CONFIDENCE-STRATIFIED FRAMEWORK FOR RECONSTRUCTING IMPERFECT MULTI-WAVE ADOLESCENT MENTAL HEALTH COHORTS

In multiwave data collected with imperfect identifiers, record linkage decisions determine who enters the cohort more than how shared records are partitioned into individuals, and should therefore be treated as an auditable component of study design.

The aim is to develop a confidence-tiered reconstruction framework that transforms multiwave cohort data compromised by nonresponse, duplicate submissions, and identifier errors into analyzable samples while preserving linkage uncertainty. The framework was applied to a four-wave online survey of seventh-grade students in Ninglang County, Yunnan Province, China; the workflow included within-wave deduplication, cross-wave candidate linkage, person-level consolidation, confidence-tier assignment, and auditing. Following the removal of 203 duplicates and the handling of 184 identifier conflicts, the 3,442 raw records were reduced to 3,239 person-wave records and 2,023 reconstructed individuals; the high-evidence cohort included 754 individuals with 1,876 observations, and the high-plus-moderate-evidence cohort included 844 individuals with 2,060 observations. Cohort inclusion was shaped most strongly by identifier data quality: the absolute standardized mean difference was 0.558 for a complete identifier with a valid checksum, 0.490 for a within-wave review flag, and 0.326 for survey completion, compared with only 0.089 for baseline PHQ-2. Primary PHQ-2 estimates remained very similar across different constrained linkage strategies (high evidence, -0.048 SD; high plus moderate evidence, -0.050 SD; strict deterministic, -0.051 SD; globally constrained forest, -0.047 SD); when the set of analyzed records was held fixed, the current framework and the Fellegi–Sunter approach yielded nearly identical results on the same 152 person-wave records: -0.312 SD and -0.307 SD, respectively. Validation and robustness tests on the FEBRL4 benchmark yielded a pairwise F1 score of 0.996, a B-cubed F1 score of 0.998, and 0 false merges; perturbation tests involving the deletion of 10% of accepted links and the addition of 5% plausible competing links produced a median absolute change of 0.0105 SD, with no prespecified scenario exceeding 0.05 SD or reversing the direction of the effect. In conclusion, confidence-tiered reconstruction enables longitudinal analysis under real-world field constraints by organizing linkage uncertainty in a transparent and auditable manner; joint assessment of identity, cohort selection, person partitioning, and outcome estimates is recommended.

All WCP 2026 summaries