Coverage, Not Volume: Measurement and Reporting Artifacts in Multi-Platform Social-Media Interrupted Time-Series Analyses of COVID-19 Infodemiology

Abstract Background Social-media infoveillance advertises “multi-platform” and “longitudinal” coverage, yet timestamp provenance, monthly coverage completeness, and multiple-comparison handling are rarely described—choices that decide whether a policy effect is real or artifactual. Methods We retrieved 58,750 verified Chinese COVID-19 posts across five platforms (Weibo, Douyin, Bilibili, Zhihu, Xiaohongshu) via retrospective archival queries (2019-07 to 2026-07). A naive pooled interrupted time-series (ITS) was fit first, then a disclosed post hoc eligibility rule (≥ 80% non-zero monthly coverage; ≥12 months each side of the Nov-2022 interruption). Qualifying platforms entered López-Bernal segmented regression with Newey–West HAC standard errors, Benjamini–Hochberg FDR (m = 2 primary; m = 8 sensitivity), and robustness checks (per-period composition, keyword-yield auditing on terms with ≥ 8 posts per period, leave-one-month-out, placebo breakpoints). Results Naive pooled ITS gave a spurious + 8,225.7 posts (P=.03), driven by incomparable pooling, a single Weibo scrape month (44,228 posts, 75.3%), and a composition flip at the interruption (pre 96.1% Bilibili+Zhihu; post 79.8% Weibo—Simpson’s paradox). After eligibility, only Zhihu qualified (Douyin’s breakpoint-density selection defeated identification); no Zhihu coefficient was policy-attributable (smallest q=.064, m = 2; .157, m = 8). The largest Zhihu level (+ 65.5) reflected period-biased vocabulary: 7 post-pandemic terms yielded 25 vs 1,089 posts versus 1.43× for 10 stable terms; it decomposes as + 65.54 = + 20.35 + 45.20, and the excluded term ranks 1/57 of candidate breaks yet is observationally equivalent to topic emergence. Douyin’s slope showed the same artifact (+ 4.78 = + 0.34 + 4.44; 92.9% from terms barely retrieved pre-breakpoint). Sentiment agreement was unusable (κ = 0.12). Conclusions Multi-platform ITS validity rests on temporal completeness, timestamp provenance, and stable retrieval yield—not corpus size. We propose six reporting standards and two prospective requirements (freeze and version-control the keyword list before crawling; store crawl_time and publish_time separately, ≥ 12 months comparable coverage per side). Without a pre-registered frozen vocabulary, ITS on platform search output cannot separate instrument change from topic emergence; single-platform restriction is insufficient.