The interplay between partisanship, forecasted COVID-19 deaths, and support for preventive policies

The study investigates whether partisan differences in support for COVID-19 preventive policies are mediated by partisan differences in beliefs about the pandemic’s severity, operationalized as forecasts of total COVID-19 deaths. The context is the global implementation of strong non-pharmaceutical interventions, which, despite effectiveness, entail significant social and economic costs. Prior U.S.-focused research shows partisan polarization in attitudes, behaviors, and beliefs about COVID-19. The authors articulate two competing hypotheses: (1) the mediation hypothesis—partisanship shapes beliefs about future deaths, which in turn drive policy preferences; (2) the independence hypothesis—partisanship independently correlates with both forecasts of deaths and policy support, with no direct relationship between forecasts and preferences. The purpose is to test these hypotheses across diverse political and cultural contexts through four experiments, thereby informing psychological theories of partisan behavior and implications for public health communication.

The paper situates its research within extensive work showing political leaders and elite discourse shape public opinion during crises and that the COVID-19 pandemic elicited partisan divides in the U.S. across preferences and behaviors. Studies cited indicate Republicans perceive lower risk and forecast fewer deaths than Democrats, and that partisan identity and elite cues can dominate over individual policy content (e.g., policy sponsorship and group influence effects). Theoretical accounts include rational-choice models where differing risk perceptions drive different policy preferences, and identity-based (tribal) models where social identity motives under collective threat shape policy attitudes independently of factual beliefs. The authors note the literature’s predominant U.S. focus and aim to test generalizability across Argentina, Uruguay, Brazil, and the U.S. by examining whether belief differences about death tolls mediate or are independent of policy support.

Design and approach: Four independent experiments (one primary and three quasi-replications) were conducted to test whether partisanship’s effect on support for COVID-19 policies is mediated by forecasts of COVID-19 deaths or operates independently. All studies asked participants to consider an exogenous anchor (low vs high) before forecasting national COVID-19 deaths (or cases in some U.S. conditions), then rate agreement with several COVID-19 policy interventions. Analyses were conducted separately by experiment to avoid cross-country comparisons due to design differences. Experiment 1 (Argentina): - Sample: N=640 university students from four universities (60.3% female; mean age 27.0±10.8 years; 41.3% with complete university education). Data collected in May 2020 (cumulative national deaths 382–589). - Partisanship measures: Support for the ruling party (Likert from strongly against to strongly support), reported 2019 vote, and a two-item affective polarization measure (emotions if one’s child supported ruling party vs opposition). - Anchoring and forecasting: Randomized to low anchor (600 deaths) or high anchor (60,000 deaths) when asking whether year-end deaths would be greater or lower than the anchor; then forecast total deaths by Dec 31, 2020 (actual 43,245). - Policy support: Nine statements rated 0–7 Likert; items included school reopening, bans on gatherings, exercise permissions, age-based movement restrictions, criminal records for violations, freedom to travel domestically without permission, tracking movements of COVID-19 positive individuals via cell data, fines for false app information, and mandatory geolocation sharing. Reverse-coded items 1, 3, and 6. Experiment 2 (Uruguay): - Sample: N=372, recruited via Offerwise panel (49.7% female; mean age 35.7±13.6 years; 52.9% ≥ complete middle education). Conducted June 2020 (cumulative deaths 25–27). - Anchors: Low 40 deaths; high 4,000 deaths (approx. +60% and +1600% relative increases from launch levels). - Policy items: Nine statements adapted to local context (e.g., universities reopening, bans on large gatherings, fines for distancing violations, age-based restrictions, criminal records for violations, domestic travel without permission, tracking via cell data, fines for false information, mandatory geolocation sharing). Reverse-coded items 1 and 6. Experiment 3 (Brazil): - Sample: N=353 via Offerwise (52.1% female; mean age 33.2±11.1 years; 65.4% ≥ complete middle education). Conducted June 2020 (cumulative deaths 51,271–58,314). Portuguese language. - Anchors: Low 80,000 deaths; high 8,000,000 deaths (approx. +60% and +1600% increases relative to launch levels). - Policy items: Nine statements analogous to other studies, including school reopening, bans on meetings/gatherings, age restrictions, criminal records for violations, reopening businesses without authorization, tracking via cell data, fines for false information, mandatory geolocation sharing. Reverse-coded items 1 and 6. Experiment 4 (United States): - Preregistration: https://aspredicted.org/wj8de.pdf - Sample: N=615 via Prolific, representative on age, gender, ethnicity (51.2% female; mean age 45.8±15.9; 58.1% ≥ complete college education). - Partisanship measure: Single 11-point scale from strong Democrat to strong Republican, plus 2020 vote intention. - Forecasting competition: Participants forecasted the number of new cases and deaths for the upcoming week (Jul 27–Aug 2, 2020), incentivized with a USD 2 bonus for the top 10% by least absolute error within condition. - Conditions (six between-subjects): • Deaths-anchored: Low anchor 40 deaths (then estimate deaths; then cases); High anchor 400,000 deaths; Control (no anchor on deaths; then cases). • Cases-anchored: Low anchor 8,000 cases (then estimate cases; then deaths); High anchor 8,000,000 cases; Control (no anchor on cases; then deaths). - Policy items (7): School reopening in 2020, bans on non-essential public events, federal tracking via app, mask wearing optional, age-based home restrictions (70+), federal permission for interstate travel, ban mass protests until vaccine. Reverse-coded items 1 and 4. Analytical strategy: - Hypothesis tests for anchoring effects on forecasts and on policy support (t-tests), and correlations between forecasts and policy support. - Power analysis via Monte Carlo simulations to detect small effects (d=0.2; r=0.15). - Equivalence testing (two one-sided tests, TOST) using ±0.5 Likert points as smallest effect of interest. - Bayes factor analyses (Matlab, klabhub/bayesFactor) to quantify evidence for no anchoring effect on policy support and for no association between forecasts and policy support; aggregate evidence by multiplying Bayes factors across experiments. - Structural equation modeling/path analysis with three simultaneous equations: (1) ruling party support as a function of age, gender, education; (2) forecasted deaths as a function of anchor (High vs Low) and ruling party support; (3) agreement with COVID-19 policies as a function of forecasted deaths and ruling party support. Models fit via maximum likelihood in Stata; compared full, mediation (β of partisanship→agreement = 0), and independence (β of forecasts→agreement = 0) models using BIC. For Experiment 4, only the death-anchor conditions (low, high) were included in SEM to harmonize with Experiments 1–3.

Across all four experiments, anchoring strongly shifted forecasts but did not change policy preferences, and forecasts were not associated with policy support; partisanship correlated with both forecasts and policy support. - Experiment 1 (Argentina, N=640): • Forecasts: Low vs high anchor (log10 deaths) 3.3±0.4 vs 4.1±0.5; Cohen’s d=1.55; t(638)=19.7; p=1×10⁻⁶⁷. Median forecasts differed five-fold (2000 vs 10,000). • Policy support: 3.54±1.50 vs 3.45±1.40; d=0.06; t(638)=0.87; p=0.39. • Forecasts–support correlation: r=-0.05; p=0.19. - Experiment 2 (Uruguay, N=372): • Forecasts: 1.7±0.3 vs 2.2±0.6 (log10); d=0.94; t(370)=9.1; p=8×10⁻¹⁸. • Policy support: 3.81±1.22 vs 3.60±1.23; d=0.17; t(370)=1.61; p=0.11. • Correlation: r=0.08; p=0.09. - Experiment 3 (Brazil, N=353): • Forecasts: 5.0±0.5 vs 5.3±1.0 (log10); d=0.43; t(351)=4.1; p=6×10⁻⁵. • Policy support: 4.80±1.13 vs 4.87±1.24; d=0.06; t(351)=0.54; p=0.59. • Correlation: r=0.10; p=0.06. - Experiment 4 (United States, N=615): • Deaths-anchored conditions: Forecasted deaths (log10) low 3.0±0.8; control 3.6±0.8; high 4.1±0.9; low vs high d=1.32; t(223)=9.8; p=2×10⁻¹⁹. Forecasted cases also shifted (low vs high d=0.70; p=4×10⁻⁷). Policy support unchanged (low 3.72±1.50; control 3.68±1.67; high 3.60±1.36; d=0.09; p=0.52). Correlation deaths–support r=0.03; p=0.57. • Cases-anchored conditions: Forecasted deaths and cases shifted (deaths low vs high d=0.64; p=2×10⁻⁵; cases low vs high d=1.03; p=3×10⁻¹¹). Policy support unchanged (d=0.21; p=0.13). Correlation cases–support r=-0.04; p=0.41. - Evidence for absence: • Power analyses indicated ≥75% power to detect small effects; probability of missing small true effects in all studies was <0.001% (anchoring on support) and <0.0001% (r=0.15 correlations). • TOST equivalence tests rejected meaningful differences in policy support across anchors in all experiments (e.g., Arg 95% CI [-0.13, 0.32], p=3×10⁻⁴). • Bayes factors favored the null (no anchoring effect on support): BF01=7.8, 2.5, 7.4, 5.7; aggregate BF01=835.6. For no forecasts–support association: BF01=13.2, 5.8, 3.8, 19.9; aggregate BF01=5556.2. - Structural equation modeling: • Partisanship predicted lower forecasted deaths (more pro-government → more optimistic) in all countries (β≈-0.06 to -0.13; all p≤0.006). • Partisanship predicted policy support with context-dependent sign: positive in Argentina and Uruguay; negative in Brazil and the U.S. (all p≤0.004). • Forecasted deaths did not predict policy support (all p≥0.06). • Independence model (partisanship independently influences forecasts and policy support) outperformed mediation and full models by BIC (ΔBIC vs mediation: 90, 5, 19, 10; vs full: 7, 2, 4, 3).

Findings directly address whether belief differences about the pandemic’s severity mediate partisan gaps in support for preventive policies. Despite large, experimentally induced shifts in death forecasts, policy preferences remained unchanged and forecasts were uncorrelated with policy support, refuting the mediation hypothesis. Structural models show that partisanship independently relates to both forecast optimism (pro-government partisans forecast fewer deaths) and policy preferences, with the direction of policy support aligning with each country’s political context (greater support for restrictions where governments endorsed strong measures; lower support where leaders minimized risks). These patterns align with identity-based or tribal accounts of partisan behavior, where group identity and elite cues dominate over individual factual beliefs in shaping policy attitudes. The cross-national replication across diverse samples, languages, and contexts suggests generalizability beyond the U.S. and that ideological left–right placement per se may be less central than alignment with the ruling party’s stance. The results imply that changing beliefs about case/death counts may not translate into support for interventions absent cross-partisan elite consensus, highlighting challenges for public health communication in polarized environments.

The study demonstrates that partisanship independently correlates with both forecasts of COVID-19 deaths and support for preventive policies, while forecasts themselves do not predict policy preferences. Across four experiments in Argentina, Uruguay, Brazil, and the United States, strong anchoring manipulations shifted forecasts but left policy support unchanged, with Bayesian and equivalence tests providing evidence for an absent forecasts–preferences link. These findings challenge rational mediation accounts and are more consistent with identity-based theories of partisan behavior. Implications include that communication strategies focusing solely on conveying pandemic severity are unlikely to increase support for interventions without cross-party elite alignment. Future research should examine additional drivers of policy support (e.g., media ecosystems, social networks, cultural factors, policy stringency, local context), explore causal pathways between partisanship and preferences, and test interventions that reduce identity-driven polarization to improve policy consensus.

- Cross-experiment comparability: Differences in sampling strategies, timing, and policy items across countries preclude direct cross-country comparisons of policy support levels. Instruments were adapted to local contexts, limiting comparability. - External validity and generalizability: Convenience samples in Argentina, Uruguay, and Brazil, and although the U.S. sample was representative on some demographics, it may differ on others (education, income, geography). Observed effect magnitudes may under- or overestimate population parameters. - Causality: While anchoring experimentally manipulated forecasts, associations between partisanship and both forecasts/policy support are correlational. Reverse causality (policy preferences influencing party support) cannot be ruled out. - Scope of variables and explained variance: The models focus on three variables and explain a modest proportion of variance (R²≈0.08–0.38). Other factors likely contribute (media consumption, social circles, cultural collectivism, policy stringency, neighboring policies), which were not directly modeled.