Syntes: Li & Yan (2024) — paper vs digital reading, meta-analysis 2000–2022

Källa: Li, Y., & Yan, L. (2024). Which reading comprehension is better? A meta-analysis of the effect of paper versus digital reading in recent 20 years. Telematics and Informatics Reports, 14, 100142. https://doi.org/10.1016/j.teler.2024.100142

Metadata

• ID: MEDIA-012
• Tier: meta
• N / design: 46 effects from 46 controlled experiments in 37 papers (Chinese- and English-language databases, published 2000–2022, searched 9 Jan 2023); each group read on one medium only; same text content in both groups, longer than one sentence; students "with basic reading abilities", no special groups (p. 2). Random-effects model, Hedges' g, CMA 3.0 (p. 5). Several digital conditions in one experiment (computer, tablet, phone) were merged into one effect (p. 3).
• Population: kindergarten to university. Stage-coded effects: kindergarten 8, primary 6, middle school 8, university and above 22 (Table 3, p. 7). 12 effects from China, 34 from other countries.
• Sign convention: positive g = digital better, negative g = paper better (pp. 5–6).
• Full text: read in full 2026-09-28 (PDF supplied by Jonatan; publisher page was Cloudflare-blocked on 2026-09-27 and 2026-09-28 04:05).

Finding

1. Overall: no significant difference. Random effects g = 0.046, 95 % CI −0.192 to 0.283, p = 0.706 (fixed effect g = 0.004). Heterogeneity very high: Q = 548.119, df = 45, I² = 91.790 % (Table 2, p. 6; text p. 5). Egger's test p = 0.5476, funnel plot symmetric (p. 5).
2. Learning stage (Q = 16.147, p = 0.001): university and above g = −0.432 (CI −0.654 to −0.21, p < 0.001, k = 22) — paper significantly better. Kindergarten g = 0.866 (p = 0.067, k = 8), primary g = 0.688 (p = 0.081, k = 6), middle school g = 0.204 (p = 0.487, k = 8) — digital-leaning, none significant (Table 3, p. 7; text p. 6).
3. Country (Q = 12.768, p < 0.001): China g = −0.66 (CI −1.108 to −0.212, k = 12), paper better; other countries g = 0.279 (CI 0.025 to 0.534, k = 34), digital better (Table 3, p. 7).
4. Text type (Q = 11.570, p = 0.001): literary g = 0.968 (CI 0.321 to 1.616, p = 0.003, k = 12), digital better; informational g = −0.218 (CI −0.439 to 0.002, p = 0.053, k = 26), paper-leaning, marginal (Table 3, p. 7).
5. Text length (between-group Q = 3.189, p = 0.203, not significant): ≤1000 words g = 0.125 (ns, k = 8); 1000–2000 words g = −0.391 (CI −0.633 to −0.150, p = 0.001, k = 6); >2000 words g = −0.524 (CI −1.114 to 0.065, p = 0.081, k = 9) (Table 3, p. 7; text p. 6).
6. Reading time (Q = 9.871, p = 0.002): time-limited g = −0.468 (CI −0.75 to −0.186, p = 0.001, k = 10), paper better; free reading g = 0.192 (CI −0.108 to 0.493, p = 0.209, k = 26), ns (Table 3, p. 7).
7. Interactivity of the digital text (Q = 9.96, p = 0.002): interactive (AR, multimedia, online annotation/feedback) g = 0.959 (CI 0.297 to 1.622, p = 0.005, k = 11), digital better; non-interactive g = −0.172 (CI −0.404 to 0.061, p = 0.149, k = 33), paper-leaning, ns (Table 3, p. 7).
8. Reading strategies (pen marking, reading guidance, word definitions; Q = 7.291, p = 0.007): with strategies g = 0.781 (p = 0.008, k = 12); without g = −0.172 (ns, k = 8) (Table 3, p. 7).

Not analysed (checked against the full text)

• Calibration / overconfidence: no calibration or confidence outcome was pooled. Overestimation on screen appears only in the Discussion as a narrative account of the shallowing hypothesis (p. 7), not as data — although several included studies measured calibration (Ackerman & Goldsmith 2011; Singer & Alexander 2017; Singer Trakhman et al. 2019; Norman & Furnes 2016; Chen & Catrambone 2015).
• Comprehension depth: the outcome pools any of three NEPS levels (locate/recall, integrate/interpret, reflect/evaluate) into one "reading comprehension effect" (p. 3); no depth moderator.
• Device: tablets, desktops, phones and e-readers were not distinguished (limitation, p. 8).
• Publication year: no year moderator or meta-regression.
• Scrolling vs pagination: not coded; scrolling is mentioned only as a possible explanation (p. 6).

Mechanism

Authors lean on the shallowing hypothesis (less effort, more mind-wandering on screen) for the paper advantage under time limits, with informational text and longer texts, and on interactivity and reading supports for the digital advantage (pp. 6–7).

Limits

• Very high heterogeneity (I² ≈ 92 %); moderators are one-at-a-time subgroup contrasts, no meta-regression, so they are confounded with each other: the literary, kindergarten, interactive and "other countries" subgroups largely contain the same children's e-book / AR storybook studies (Table 1, pp. 4–5).
• Many digital conditions add features (AR, multimedia, annotation, word definitions) to the same text, so the overall null is not a clean "same text on screen vs paper" comparison.
• Small subgroups (k = 6–12 for most school-age and text-length cells); authors warn that the uneven distribution (mostly university students) calls for caution (p. 8).
• The subgroup "Total" rows in Table 3 (e.g. learning stage total g = −0.229, k = 44) do not reproduce the overall random-effects estimate (g = 0.046, k = 46), and the paper does not explain the difference.
• Journal is a communication/informatics outlet; data "available on request" (p. 8).

This claim dies if … (for C3)

Meta/RCT in the relevant age band where screen ≥ paper for deep comprehension under equal time and the calibration gap disappears. Li & Yan does not meet this: its in-scope subgroups favour paper (university, time-limited, 1000–2000 words; informational marginal), and it pooled no calibration data. Kill-condition test on C3 survived 2026-09-28.

Swedish implication

Do not read "no overall difference" as "medium does not matter": in this meta the paper advantage is clear for university readers, under time limits and for 1000–2000-word texts, while the digital advantage comes from interactive or supported e-books, mostly for young children. For plain continuous informational text, the direction still favours paper; for school-age readers this meta cannot separate medium from added digital features.

Links

• DOI: https://doi.org/10.1016/j.teler.2024.100142
• Related: delgado-etal-2018-paper-vs-screen, clinton-2019-paper-vs-screens, salmeron-etal-2024-handheld-meta, furenes-kucirkova-bus-2021-children-paper-screen, florit-etal-2025-first-grade-paper-screen, singer-alexander-2017-reading-across-mediums, mangen-etal-2013-paper-vs-screen