Reading Shanghai's lockdown through four newspapers

The most striking number in my undergraduate thesis is 91%. In the hand-coded local-media sample from May 2022, the peak month of Shanghai's COVID lockdown, that was the share of articles using a resilience frame.

This is an informal companion to my Signature Work at Duke Kunshan, Framing the Crisis (2024), under Prof. Andrew Macdonald. The thesis itself is technical: 1,161 articles, LDA topic modeling, hand-coded content analysis. The point I want to make here is small enough to make without any of that.

What I actually read

Across the April-to-June 2022 lockdown period, Shanghai was locked down. Twenty-six million people, food shortages, a collapsing testing apparatus, the whole thing. I scraped four newspapers across that window. The full corpus covered April 1 through June 1; the hand-coded time sample also included June 2 so it could compare the beginning, midpoint, and end:

• People's Daily (人民日报) and Global Times (环球时报), the central organs.
• Jiefang Daily (解放日报) and Xinmin Evening Post (新民晚报), Shanghai's two big local papers.

Same event, same language, different audiences. 1,161 articles in total, 386 central and 775 local. The local corpus was bigger because local news has to operate at local resolution: which neighborhood is sealed, which testing site is overwhelmed, which residential committee is coordinating grocery delivery. The central organs don't cover those things. The local papers have no choice.

What the papers were doing differently

I ran two passes. First, an LDA topic model, letting the machine find what each corpus was actually about before I labeled the topics. Second, a hand-coded content analysis on a stratified sample of seventy-six articles, scored along four dimensions: who they cite, what kind of article it is, what the tone is, and what the frame is.

Reading the numbers off the bar charts:

• Sources. Central media cited government officials in 67% of pieces; local media in 25%. Local media cited ordinary people in 51% of pieces; central media in 11%.
• Nature. 78% of central articles were information-sharing or government announcements, and only 11% were narrative. Local ran 34% information-sharing and 55% narrative.
• Tone. Central media leaned authoritative (56%); local media leaned warm and sympathetic (61% warm, 37% sympathetic).
• Framing. Both used resilience heavily, but local papers also concentrated on local community response (46% against central's 11%), while central concentrated on the national economy (67% against local's 22%).

What that adds up to is two halves of a single apparatus doing complementary work. Hard propaganda upstream, with an authoritative voice, official sources, and national-cohesion themes. Soft propaganda downstream, with a warm voice, ordinary-people sources, and community-resilience themes. They weren't contradicting each other. They were covering different shifts.

The one number I keep coming back to

The temporal evolution of the local papers is what stayed with me. In April, when the lockdown began, local coverage was relatively balanced: roughly even amounts of struggle (45%), resilience (45%), health (50%), and local community response (55%). A city telling itself what was happening to it.

By May, the worst month, when public morale was visibly cracking, that balance collapsed. In the hand-coded local-media sample, 91% of articles used a resilience frame and almost everything else got crowded out. Coverage became one long narrative about endurance, survival, holding on.

You can read that as cynicism, the apparatus pivoting into morale management when reality stopped cooperating. You can read it as obligation, the local press doing what it always does in a crisis its readers are physically inside. Probably it was both. The 91% figure comes from the hand-coded sample; its direction echoes the community-resilience themes found in the topic model. By June, as the city began to unlock, the coded sample's resilience share fell back to 43% and local response, meaning community efforts, neighborhood organizing, and recovery, rose to 48%. The story shifted again.

What I take from it

Two things. First, that "Chinese state media" treated as a single object is the wrong unit of analysis. The system is stratified: central does signaling and information control, local does empathy and morale management, and the coordination between them is more interesting than either piece alone.

Second, that what propaganda omits and re-weights over time is often more revealing than what it asserts. Nobody in May 2022 was lying about resilience. In the coded sample, they were talking about almost nothing else. When ninety-one percent of a sample uses one frame, the frame has stopped merely describing the situation and started doing a job.

The thesis page explains the methodology and findings. The underlying newspaper corpus is not redistributed here because the source articles remain copyrighted.