Pandemic in the digital age: analyzing WhatsApp communication behavior before, during, and after the COVID-19 lockdown

The paper examines how the COVID-19 pandemic, and especially the March–April 2020 lockdowns, affected communication behavior on the leading mobile instant messaging (MIM) app WhatsApp. Context: Massive global restrictions and physical distancing shifted education, work, entertainment, and social interaction online, with documented increases in Internet and MIM usage. Research questions: RQ1 asks whether WhatsApp communication behavior changed during the March/April 2020 lockdown compared to the same months in 2019; RQ2 asks whether there are lasting effects on WhatsApp communication beyond the lockdown. Purpose and importance: Using large-scale anonymized chat histories, the study aims to quantify short- and long-term behavioral changes, offering insights for behavioral research and for network operators to anticipate and manage traffic during crises.

The related work surveys (1) usage behavior in MIM applications and WhatsApp specifically, showing MIM’s prominence relative to SMS and its roles in social and work coordination; (2) impacts of COVID-19 on Internet traffic and online behavior, documenting rapid traffic increases, shifts to weekend-like patterns, and changing use of messaging, live streaming, and online learning. Prior COVID-19 studies on MIM reported increased messaging volumes early in the pandemic but lacked detailed, longitudinal analyses of communication behavior within MIM apps. The gap identified is an in-depth, pre/during/post comparison of WhatsApp communication patterns at scale.

Data collection: Private WhatsApp chat histories were voluntarily submitted via the WhatsAnalyzer tool (Schwind & Seufert, 2018), leveraging WhatsApp’s export-by-email feature. Anonymization removed names, phone numbers, and message content, retaining metadata such as timestamps, message type (text, image, audio, video, location, contact, document, gif), number of characters, and emoji counts. A normalized representation was stored due to format differences by OS, app version, and system language. System messages were reliably detected for English, Spanish, German, and Italian; users could switch phone language to conform. Users were informed that anonymized data would be used for research. Dataset: The overall database contained 7312 entries; after filtering duplicates and incomplete chats, 5891 chats remained (116,441 users; 75,910,808 messages). For this study’s time-focused analyses, a trimmed subset included 2577 private chats with 25,378,093 messages from 51,973 users. Study design: Two analyses address the research questions. - RQ1 (lockdown effects): Paired comparison of chats active across both periods—Mar/Apr 2019 (pre-pandemic) vs Mar/Apr 2020 (lockdown). Inclusion: chats started before March 2019 with messages in both periods; N=345 chats (250 group chats). Active users: 8150 (2019 period) vs 4378 (2020 period; −46% unique senders due to inactivity/leaving). Metrics: total and average messages per day, per-chat average daily messages, message-type shares, and hourly distribution of activity (in local time). Statistical tests: Kolmogorov–Smirnov (KS) tests for distribution changes; Chi-square tests for message-type distributions; paired t-tests per media type with Bonferroni correction. - RQ2 (lasting effects): Full-year comparison of March 2019–February 2020 (pre-pandemic; 2039 chats; ~15M messages; ~38k users) vs March 2020–February 2021 (pandemic; ~9.7M messages; ~27.2k users; 29–39% lower data density). Analyses use relative metrics due to differing chat counts. Metrics: monthly mean of average daily messages per chat with 95% CIs and linear trend from pre-pandemic year; annual media shares with paired tests (paired by days) using Chi-square and paired t-tests with Bonferroni correction; hourly distribution differences; regularity via share of days per chat where daily messages exceed mean + 2 standard deviations (outlier frequency), with KS test across periods. No user-level demographics or country-of-origin data were available; analyses assume stable geographic composition over time.

Lockdown period (Mar/Apr 2020 vs Mar/Apr 2019, paired chats): - Message frequency: Daily total messages spiked to 21,785 (2.58× the 2019 period’s average). Average daily total rose from 7,087 (2019) to 13,787 (2020). Total messages nearly doubled from 432,348 to 827,241 despite a 46% drop in different senders. KS tests reject equal distributions (p<0.001). - Per-chat changes: 29% of chats sent fewer messages (some as low as 5.35% of prior), but median increase was strong—50% of chats exceeded 173% of prior activity; >45% at least doubled (>200%). Maximum observed increase: 454-fold (single data point). - Media composition: Overall media share rose from 15.35% to 18.14% (+18% relative). Notable changes: video 2.27%→4.21%; images and audio slightly increased; locations increased nearly 4×; documents 0.23%→0.39%; contacts decreased. Chi-square and paired t-tests with Bonferroni show significant differences (p<0.001). - Time-of-day shifts: Significant redistribution (Chi-square p<0.001). Night-time (12–3 am) increased by +0.27% to +0.77%; mid-day (8 am–3 pm) increased with a peak +1.45% at 11 am; evening (4–10 pm) decreased by −0.70% to −1.46%. Lasting effects (full-year pandemic vs pre-pandemic): - Message frequency: Monthly means show a large lockdown peak, followed by a drop below the pre-pandemic trendline but remaining above same-month 2019 values; toward period end, values approach the prior trend. Excluding lockdown months, the mean average daily messages per chat increased from 41.02 to 50.65. - Media usage over the year: Overall similar levels but with composition shifts: images −0.34%; audio ~constant; video +0.14%; locations 1.69%→0.75%; contacts 0.06%→0.01%; documents up 36% to 0.50%; more users never sent media. Chi-square p<0.001; paired t-tests (Bonferroni) p<0.05. - Persistent time-of-day shift: Night-time +0.1–0.2%; mid-day peak +0.89% at 11 am; decreased evening messaging from 4 pm to midnight, indicating durable behavioral change. - Regularity of messaging: KS p<0.001 for distributions of outlier-day shares. Chats with 0% outlier days more than doubled (about 15%; elsewhere reported 16.5%), indicating more regular, stable messaging patterns during the pandemic year. Increases in higher outlier shares (≥7% of days) are largely attributable to the initial lockdown spike; excluding Mar/Apr removes that trend.

Findings directly answer RQ1 and RQ2. RQ1: WhatsApp communication changed markedly during the March–April 2020 lockdown—message volumes nearly doubled with a sharp daily spike, media shares (especially video, locations, documents) increased, and activity shifted from evening toward mid-day and slightly into late night. RQ2: Some effects persisted throughout the first pandemic year. While overall media usage normalized, the daily activity pattern remained shifted toward mid-day and away from evenings, and messaging became more regular with fewer outlier days, suggesting integration of WhatsApp into routine professional/educational and day-time social communication. These results underscore the role of MIM during crises and offer actionable insights for network operators to anticipate diurnal demand shifts and potential spikes, and for social/behavioral scientists studying how physical distancing reshapes digital social interaction.

The study contributes a large-scale analysis of private WhatsApp communication across pre-pandemic, lockdown, and pandemic-year periods. It identifies short-term lockdown effects (surges in message volume, increased media—especially video and locations, and a mid-day activity shift) and lasting changes (persistent mid-day focus, increased late-night messaging, more regular communication patterns). Post-lockdown, message volumes stabilize near pre-pandemic trends while diurnal usage shifts persist. Future research should assess which effects endure long-term beyond the first pandemic year, include cross-country comparisons once metadata are available, and extend analyses to other MIM platforms and facets of online/offline social behavior to understand broader societal adaptations.

- Volunteer, convenience sample via a public tool; no targeted recruitment or representativeness guarantees. - No user demographics or country-of-origin data; cannot analyze national differences or demographics; assumes stable geographic composition over time. - Anonymization removes content, limiting semantic/contextual interpretation; only metadata-based behavior is analyzed. - System message detection supported only English, Spanish, German, and Italian; other languages required users to switch system language. - For RQ1, only chats active in both periods were included (N=345), potentially biasing toward persistent groups; for RQ2, differing chat counts between years required reliance on relative metrics; data density decreased in the pandemic year. - Observational design; cannot ascribe causality beyond associations with pandemic/lockdown timing. - Data set not publicly available until project completion; replication depends on access to similar data.