Work

Estimand-dependent associations between fragmented short-video exposure and subsequent behavioral viewing persistence in a longitudinal computational audit

Year: 2026
Type: article
Abstract: Background Fragmented short-video exposure is relevant to competing accounts of processing demand, motivational resource allocation, and reward-based persistence. Platform logs cannot identify those states or mechanisms. We therefore conducted an observational methodological audit of whether the within-person association between fragmented short-video exposure and subsequent behavioral viewing persistence was stable across consequential alternative features, outcomes, time lags, and regression estimators. Methods We analyzed 11,713,045 KuaiRand-1K interaction records. The primary analysis used 25,560 complete longitudinal user-day observations from 1,000 users. A post hoc audit used a non-identical recovered processed table containing 26,499 complete-predictor rows; it is reported as a specification audit and is not pooled with the primary sample. Longitudinal regressions used user-clustered standard errors. The audit crossed alternative FSVE, switching, and outcome constructions; interaction adjustment; lag definitions; and hybrid versus user fixed-effects estimators. Selection scenarios examined single-exposure rows, end-of-window days, and selection on observed FSVE. FCE-Net was evaluated with user-separated partitions, conventional machine-learning comparators, and a BVPI-only ablation. Results In the archived interaction-rate-adjusted model, within-person FSVE was negatively associated with subsequent BVPI ( b = −0.0582, 95% CI −0.0771 to −0.0394, p = 1.40 × 10^-9). In the separate recovered-table audit, the hybrid model produced b = 0.0311 (95% CI 0.0129 to 0.0493) and the user fixed-effects model produced b = −0.0178 (95% CI −0.0336 to −0.0019). At Table S21’s displayed precision, the 96 coefficients comprised 31 positive values, 64 negative values, and one value displayed as −0.0000; 16 positive and 50 negative confidence intervals excluded zero. The direction therefore depended on estimator, feature construction, and lag. Across all assumption-based selection scenarios, hybrid recovered-table coefficients ranged from 0.0118 to 0.0367. The maximum-assignment confidence interval included zero, and none of these scenarios addressed the user fixed-effects result. On the user-separated test set, FCE-Net achieved R 2 = 0.6466, random forest achieved 0.6418, and BVPI-only FCE-Net achieved 0.6437; the full-versus-BVPI-only interval included zero. Conclusion The available evidence did not establish a stable directional association between fragmented short-video exposure and subsequent behavioral viewing persistence. The archived primary estimate was negative, but a non-identical recovered-table audit produced opposite signs for the hybrid and user fixed-effects estimators. Because sample provenance, estimator choice, and outcome-dependent selection cannot be jointly resolved with the available records, the study establishes estimand dependence rather than a population-level direction. The results concern observable platform behavior and do not identify cognitive load, attentional decline, reward mechanisms, causal effects, or enduring psychological consequences.
Language: English
Subfield: Communication
Open Access status: gold
Estimated APC paid: $3,801
OpenAlex ID: W7235440704
FWCI:
Cites: 35
Cited by:
Related to: 10
Parts: 2