The language of crisis: spatiotemporal effects of COVID-19 pandemic dynamics on health crisis communications by political leaders

During times of crisis, communication by leaders is essential for mobilizing an effective and coordinated public response. The COVID-19 pandemic presented governors with one of the greatest public health challenges in modern times. Communicating specific public health guidelines and promoting compliance has been central, as individual actions directly impacted pandemic spread and outcomes. Specific word choice can affect compliance, and leaders’ wording is often designed to shape public opinion. Leadership communication varies with factors including political affiliation, demographics, linguistic style, and emotional state, and is also influenced by crisis dynamics and local population characteristics. Computational approaches enable systematic analysis of large text corpora to reveal patterns in semantics, grammar, and linguistic complexity; prior work has noted declines in speech complexity during crises. This study assembled over 1500 pandemic-related speeches from all 50 US governors during the initial months of COVID-19 to examine associations between case rates and speech properties across space and time, assessing whether pandemic dynamics affected topics and styles, and investigating relationships between pandemic intensity and linguistic markers of stress.

The paper situates its work within several strands of prior research: (1) Effects of communication and framing on compliance and behavior, including evidence that leaders’ downplaying of risks (e.g., during COVID-19 in Brazil) reduced social distancing in supportive areas. (2) Variation in leaders’ speech linked to political affiliation, demographics, individual linguistic style, and emotional state. (3) Corpus linguistics applied to study language evolution, cultural shifts, political trends, and historical epidemiology. (4) Analyses of political speech for semantics, grammar, and complexity, with findings that complexity often declines during crises. (5) Stress and language: public speaking as an acute stressor; prior work shows increased negation and simpler language under stress. The authors note gaps regarding spatiotemporal links between real-time crisis intensity (case rates) and leaders’ speech properties during a pandemic, which their study addresses.

Data collection: Pandemic-related speeches delivered by governors of all 50 US states were collected from February 27, 2020 through July 14, 2020. Sources included governors’ offices (websites or direct correspondence), Rev (commercial transcription), YouTube, and Facebook. Transcripts were curated to include only words spoken by governors. COVID-19 confirmed cases per 100,000 persons were obtained from the New York Times, aggregating state and local health authority data. Corpus: 1515 speeches containing 4,049,146 words. Median speech length 1984 words (range 100–12,557). Speeches spanned all 50 states, with availability varying by state and month. Preprocessing and inclusion: For each state and each speech, total counts for each word were tabulated. Included words were those spoken by at least 20 governors and constituting at least 0.02% of all words in at least one state. Words with multiple common meanings or not cleanly categorizable were excluded from semantic grouping on a case-by-case basis. Feature construction: Two independent raters grouped eligible words into semantic categories (details in Supplementary Table 1). Parts of speech were assigned using the NLTK POS tagger (Python 3.6.2). Linguistic complexity measures were calculated per speech: average word length (letters) and average syllable count. Normalization: For spatial analyses, word counts were normalized as the percentage of all words spoken by that governor across all speeches in that state. For temporal analyses, counts were normalized as the percentage of all words spoken by that governor across speeches in that state during a given week. Temporal analyses were performed only for states with at least 50,000 total words to ensure adequate sample size. Analytical approach: Associations between linguistic features (semantic categories, parts of speech, word length, syllable count) and COVID-19 case rates per 100,000 were assessed over space (across 50 states for the study period) and over time (weekly within each state). Spearman’s Rho correlations were computed. Confidence intervals for spatial analyses used the single Spearman’s Rho CI formula (with r and n). For temporal analyses (multiple Rho values across states), standard formulas for distributions were applied. Polynomial curves (numpy.polyfit) were used as visual aids in temporal plots.

- Corpus: 1515 speeches; 4,049,146 words; median 1984 words (range 100–12,557). - Semantic categories positively associated with case rates: - Hospital-related (HOP; examples: “ICU,” “ventilators”): Spatial Spearman’s Rho 0.56 [95% CI: 0.33, 0.75]; Temporal 0.15 [−0.01, 0.30]. - Negation (NEG; “can’t,” “no”): 0.28 [0.02, 0.52]; 0.07 [−0.01, 0.26]. - Strict instructions (ORD; “prohibited,” “compliance”): 0.31 [0.04, 0.54]; 0.13 [0.00, 0.26]. - Descriptive bad (BAD; “terrible,” “worst”): 0.30 [0.02, 0.53]; 0.046 [−0.07, 0.17]. - Religious (REL; “pray,” “God”): 0.19 [−0.10, 0.44]; 0.08 [−0.05, 0.21]. - Extreme descriptive (MST; “dramatically,” “extraordinarily”): 0.28 [0.01, 0.52]; 0.07 [−0.07, 0.22]. - Semantic categories negatively associated with case rates: - Job-related (JOB; “employment,” “workers”): −0.23 [−0.47, 0.06]; −0.08 [−0.20, 0.05]. - Travel-related (TRA; “tourism,” “hotels”): −0.32 [−0.55, 0.04]; −0.17 [−0.29, −0.05]. - Formal communication (COM; “announcement,” “declaration”): −0.30 [−0.54, −0.03]; (temporal association negative). - Help/assistance (HLP; “hospitality,” “assistance”): −0.34 [−0.57, −0.07]; −0.13 [−0.29, 0.02]. - Emergency (EMR; “crisis,” “disaster”): −0.21 [−0.46, 0.07]; −0.26 [−0.38, 0.14]. - Preventative measures (PRV; “sanitizer,” “quarantine”): −0.26 [−0.51, 0.02]; −0.24 [−0.36, −0.13]. - Parts of speech associated with case rates: - Positive: Past-tense verbs (VBD): 0.13 [0.00, 0.26]; 0.37 [0.10, 0.57]. Present-tense verbs (VBZ): 0.12 [0.01, 0.23]; 0.29 [0.02, 0.53]. Adverbs (RB): 0.32 [0.19, 0.45]; 0.21 [−0.07, 0.46]. - Negative: Possessive pronouns (PRP$): −0.23 [−0.36, −0.09]; −0.23 [−0.48, 0.05]. Plural nouns (NNS): −0.09 [−0.20, 0.03]; −0.23 [−0.48, 0.05]. Base-form verbs (VB): −0.07 [−0.22, 0.08]; −0.247 [−0.49, 0.03]. - Linguistic complexity: Average word length and average syllable count were negatively associated with COVID-19 case rates over both space and time, with strongest decreases in states experiencing the highest case rates. As cases surged, governors used shorter words with fewer syllables.

The study demonstrates strong spatial and temporal associations between COVID-19 case rates and governors’ speech properties. As cases rose, governors increasingly referenced hospital-related topics, used stricter instructions and negative/“bad” descriptors, employed more extreme descriptors, and increased negation—consistent with defensive framing, acknowledgment of uncertainty, and heightened urgency. Conversely, references to travel and jobs decreased, aligning with reduced travel and a shift in focus away from employment amid the acute public health crisis. Mentions of religion increased, potentially reflecting efforts to console and offer hope. Mentions of formal communication formats and helpful acts declined, consistent with emphasis on actionable guidance. Unexpectedly, words explicitly denoting “emergency” and preventative measures declined with higher case rates, potentially reflecting efforts to avoid panic and a relative shift in attention to other topics as crises intensified. Grammatical shifts from future-tense talk (base-form verbs) toward present and past tenses, and from nouns to verbs/adverbs, indicate a move from planning to action and reporting. Reduced use of possessive pronouns suggests a focus on impersonal, action-oriented messaging. The observed decrease in word length and syllable count aligns with known stress-related reductions in linguistic complexity during high-pressure public speaking, suggesting characteristic stress responses among governors at peak case rates. While alternative explanations (e.g., evolving political discourse) exist, the temporal co-movement of linguistic features and case rates across multiple states supports a dynamic linkage between crisis intensity and communication style.

By assembling and analyzing a large corpus of speeches from all 50 US governors during the initial months of COVID-19, the authors identify robust spatial and temporal relationships between pandemic intensity and leaders’ linguistic choices across semantics, grammar, and complexity. Several effects align with stress-related communication shifts (e.g., increased negation, simpler words), which may influence the effectiveness of public health messaging. The study provides an initial framework for understanding how crisis dynamics shape leaders’ language and underscores the need for further research across other leaders, government levels, crises, and later pandemic stages, including exploration of additional stress markers and application of advanced NLP methods to inform more effective crisis communication strategies.

- Coverage: Not all delivered speeches were available; the dataset includes over 1500 speeches but may omit some events. - Scope: Focus on US state governors and the initial months of the pandemic; findings may not generalize to other countries, government levels, or later pandemic stages. - Granularity: Analyses conducted at the single-word level; multi-word patterns and discourse structures were not explored and may yield additional insights. - Confounding: Spatial analyses may be affected by state-level attributes (e.g., population density, political affiliation). Temporal within-state analyses mitigate some confounders but cannot eliminate all. - Causality: The study examines associations; it cannot establish whether speech affected case counts or vice versa. Governors’ speech may also be shaped by political strategy, personal demeanor, or chosen information sources. - Data thresholds and categorization: Inclusion criteria (e.g., words used by ≥20 governors and ≥0.02% frequency) and semantic categorization decisions may influence results.