Generosity during COVID-19: investigating socioeconomic shocks and game framing

The paper investigates altruism and the stability of social preferences during the COVID-19 pandemic, a major negative socioeconomic shock marked by lockdowns, economic contraction, and heightened mortality and media salience. Two research questions guide the study: (1) How does the pandemic’s progression relate to generosity toward different recipients (relatives, neighbours, strangers, and the state)? (2) Do framing effects, via the order of dictator games, influence generosity decisions? The context draws on historical and contemporary evidence that altruism can flourish during crises. The purpose is to measure dynamic generosity over the early weeks of the US pandemic and to test preference stability under randomized task ordering. This is important for understanding pro-social behaviour under stress and for informing policy and charitable strategies during crises.

The study builds on models of social preferences incorporating reciprocity, efficiency concerns, trust, and affective motives such as empathy and compassion, particularly salient in crises. It references warm-glow altruism and distributive fairness as drivers of baseline generosity, and critiques outcome-only models by emphasizing social distance, personal and social norms, and group identity as determinants of generosity. Prior work shows directed altruism toward known others is stronger than baseline altruism toward strangers and that framing can influence allocation decisions, though evidence on the sensitivity of dictator games to framing is mixed. Contextual factors (gender, demographics, earned vs. windfall endowments) and social framing can affect behaviour. The study positions its contribution within this literature by examining relationally differentiated dictator games during a real-world shock and explicitly testing framing through randomized order of games.

Design and data collection: An online experiment was conducted over eight weeks starting March 30, 2020, recruiting 1255 US-based participants via Amazon Mechanical Turk (approx. 156 per week). Recruitment focused on California (41.26%), New York (36.58%), and Washington (22.16%) to hold sociodemographic and policy environments broadly comparable while exploiting different pandemic intensities. Participants received $0.30. The dataset is a repeated cross-sections panel across weeks.

Tasks: Each participant completed four independent dictator games, presented in randomized order. Common prompt: “Imagine that today you have been given 1000 dollars. How much of this amount are you willing to give to …” Recipients: an anonymous person X, the current government (to support public services), a relative, and a neighbour. Responses were integers from 0 to 1000 for each game.

Survey measures: Post-game questionnaire collected socio-demographics (gender, age, marital status, education, state), employment status (current and one month prior), and attitudinal variables: concern about COVID-19, trust in government, self-reported anxiety/mood, perceived ability to manage daily tasks, and financial security.

External data: State-level COVID-19 indicators (total cases/deaths and daily percentage changes) and labour statistics (unemployment rate, unemployment insurance claims) were merged to respondents by state and timing of daily announcements.

Econometric framework: Guided by a utility specification where utility depends on own payoffs, known others’ payoffs (with altruism coefficient β for close contacts), and anonymous others’ payoffs (coefficient α), the analysis tests three hypotheses: H1: Pandemic shock increases donations over time; H2: Effects differ by recipient; H3: Game order (framing) affects donation amounts.

Primary model: Tobit regressions of amount donated (0–1000) for each recipient, motivated by mass at zero (corner solution). Covariate vectors: X1—objective shock measures (state daily/total cases/deaths, percentage changes; unemployment rate; benefits); X2—self-reported attitudes (concern for COVID-19, anxiety, trust in government), week fixed effects; X3—demographics (gender, age, marital status, education, state); X4—ordering variables (position 1–4 for the game of interest; indicator if the relatives game was played first). The partial effects are interpreted on E(y|y>0, X). Identification addresses potential endogeneity by controlling for concern about COVID-19, capturing emotional susceptibility correlated with generosity.

Robustness checks: (i) OLS on donation amounts; (ii) Logit on probability of donating (positive amount); (iii) Quantile regressions (Q0.50 and Q0.75) to account for skew and mass at zero; (iv) Tobit with week as a continuous variable. Attention checks were included to ensure data validity.

Descriptive patterns: Average donations by recipient show strong relational gradients: relatives ≈ $300; government ≈ $136; neighbours ≈ $120; anonymous ≈ $90. Donation distributions to non-relatives are skewed toward low amounts; approximately 50% donate zero to anonymous, ~40% donate zero to neighbours and government, and ~10% donate zero to relatives. Over eight weeks, mean donations increase for all recipients; anonymous and government donations roughly double, neighbours increase markedly, relatives increase more modestly. By state, New York respondents give the least across games; Washington shows higher generosity to neighbours and relatives; Californians give relatively more to anonymous and government.

Tobit regressions (main):

• Concerned by COVID-19: Positive and significant for relatives (increasing with stronger agreement) and generally positive for neighbours and government; not significant for anonymous. Strongly agree vs strongly disagree increases relatives by ≈ $358 and shows positive effects for neighbours and government (5–10% significance levels).
• Daily percentage change in deaths: Generally not significant across games; signs vary by recipient.
• Time dynamics (week dummies, baseline week 1): Significant increases over time for anonymous, neighbours, and government; peak effects at week 8: neighbours +$244 (1%), anonymous +$149 (1%), government +$200 (1%). Relatives show a positive but weaker and less consistently significant pattern; week 8 +$145 (5%). These results support H1.
• Framing/order effects (H3): For anonymous, playing second, third, or fourth significantly reduces donations relative to first (≈ −$60, −$44, −$40 respectively) and lowers probability of donating. If the relatives game is played first, donations to neighbours fall significantly (≈ −$42). Neighbour game played second/third increases neighbour donations and raises the likelihood of donating by ~9 percentage points vs first. Order effects are not significant for relatives and government amounts.

Robustness:

• OLS results mirror Tobit: week effects strong for neighbours and anonymous (and week 8 for government); concern significant for relatives.
• Quantile regressions: Week effects emerge from week 5 onward for neighbours (especially Q0.75). For relatives at Q0.75, percentage change in deaths raises high donations by ~3%.
• Logit on probability of donating: Probability increases over weeks, with week 8 the largest across all games; concern and percentage change in deaths significant for relatives only.
• Weeks as continuous: Positive and significant for government (5%), neighbours (1%), and anonymous (5%).

Other regressors: Married/domestic partnership associated with higher donations to anonymous, neighbours, and government. Females donate less to government and more to relatives than males. Higher state unemployment rates reduce donations to known recipients. Older respondents donate less than ages 18–24. Anxiety and trust in the current government positively correlate with amounts in all games except relatives. Washington residents donate more across all games; New York residents donate less than Californians.

Findings indicate that generosity increased as the pandemic progressed, despite heightened financial insecurity and unemployment, suggesting an increased perceived marginal benefit of donating during hardship. This pattern is strongest toward recipients with weaker social ties (anonymous and government), consistent with empathy elicited by media salience and the warm-glow of giving, as well as possible expectations of reciprocity from state support in government donations. The dynamic increases are not attributable to shifting sample composition, as demographics are controlled and relatively stable across weeks. Preference stability is challenged by framing effects: donations to anonymous recipients decline markedly unless the anonymous game is first, and playing the relatives game first appears to set a reference point that depresses subsequent donations to less familiar recipients, notably neighbours. This underscores instability in social preferences when social distance is high and highlights the role of task order in shaping giving decisions. H1 is supported (increasing donations over time), H2 is supported (magnitude differs by recipient, with largest time effects for neighbours/anonymous/government, weaker for relatives), and H3 is supported (order effects, particularly for anonymous and via relatives-first reference points). These results contribute to understanding pro-social behaviour under crisis and suggest practical implications: sequencing and framing of appeals can influence charitable outcomes, especially for donations to socially distant recipients.

The study shows that during the early COVID-19 lockdown, generosity increased across eight weeks, particularly toward anonymous individuals, neighbours, and the government, indicating higher perceived marginal benefits of giving in crisis. Generosity patterns differ by recipient, reflecting social distance and relational factors. Framing effects from task order significantly affect donations to anonymous recipients and reveal a reference point effect when the relatives game is played first, indicating instability in social preferences absent strong social bonds. Contributions include dynamic evidence on altruism under a real-world shock, recipient-specific heterogeneity, and identification of framing effects. Future research should validate external validity with real monetary stakes (revealed preferences), explore mechanisms of reciprocity (e.g., ultimatum or trust games for government-related giving), and test interventions leveraging order and framing to enhance donations to socially distant recipients.

Key limitations include the use of hypothetical allocations (stated preferences) with a relatively high endowment ($1000), which may induce lower giving via endowment or stake-size effects; potential biases from online MTurk sampling and demographic over-representation (younger, highly educated), although prior research supports MTurk validity and attention checks were used; internet users may be more self-regarding; unobserved deep relational dynamics (e.g., direct communication with relatives) are not captured and may drive relatives’ donations; observational timing and state-level data may not fully capture individual exposure to shocks. External validity would be improved with real-payoff experiments and additional behavioural tasks to disentangle mechanisms.