Trust, social protection, and compliance: Moral hazard in Latin America during the COVID-19 pandemic

Abstract

Political trust is an important predictor of compliance with government policies, especially in the face of natural disasters or public health emergencies. During the COVID-19 pandemic, for example, multiple studies related political trust to increased compliance with mobility restrictions. Yet these findings come mostly from high-income countries where political trust and wealth correlate positively. In Latin America, both variables correlate negatively, allowing for better testing of competing explanations. Using a difference-in-differences design, we find that in Latin America wealth and, counterintuitively, low political trust predict increased compliance. To understand mechanisms, we decompose political trust and wealth into underlying predictors (social protection, corruption, and education) and reinsert them into the model. While education, as a wealth proxy, predicts decreased mobility across all periods, social protection, which was the strongest predictor of political trust, relates significantly to increased mobility, but only at the beginning of the lockdown prior to distribution of emergency support. This suggests the existence of a public health moral hazard early in the pandemic, whereby citizens who benefited previously from government benefits may have been more risk tolerant in the face of the COVID-19 threat. We interpret these findings within the context of the region's recent “inclusionary turn.” Future studies should explore the distinct relationships between political trust, risk perception, and compliance, especially in low- and middle-income countries, and their implications for policy responses to national emergencies.

1. Introduction

Understanding citizen responses to government policies is critical for crafting strategies to enhance collective welfare, especially in the face of natural disasters or public health emergencies. At the start of the COVID-19 pandemic, for example, mobility restrictions were intended as temporary measures to mitigate virus spread and give health systems time to prepare for the expected rise in cases. Yet compliance required citizens to trade off economic activity for physical health. Although emergency social protection policies were created to offset economic harm and decrease the cost of staying at home (Gentilini et al., 2020), we still observed within and between country variation in compliance with mobility restrictions.

One explanation offered for compliance in high-income countries is differing levels of pro-sociality, be it civic capital (Barrios et al., 2020; Durante et al., 2021) or trust in politicians (Bargain and Aminjonov, 2020a), science (Eichengreen et al., 2021; Bicchieri et al., 2020; Plohl and Musil, 2021; Brzezinski et al., 2020), or the press (Brodeur et al., 2020).1 In Singapore, however, the relationship between government trust and compliance may have also created a moral hazard whereby citizens with higher levels of political trust were more risk tolerant and less careful about social distancing because of faith in the government's ability to manage the crisis (Wong and Jensen, 2020; Wachinger et al., 2012). Furthermore, we lack evidence to establish the external validity of these findings outside of high-income settings.

An alternative explanation of compliance, applied more readily to low- and middle-income countries (LMICs), acknowledges that economic constraints hinder the ability to adhere to stay-at-home orders because of the lack of resources to smooth consumption (Ravallion, 2020; Robalino, 2020) or less opportunity to perform remunerative work at home (Dingel and Neiman, 2020; Garrote et al., 2020). While there is a positive relationship between wealth and compliance in the United States (Wright et al., 2020) and globally (Bargain and Aminjonov, 2020b; Aminjonov et al., 2021), the ex ante positive correlation between education (and other socioeconomic proxies) and political trust in high-income countries may hinder causal comparison of these two competing explanations. In LMICs, such as those in Latin America, a negative correlation exists between political trust and socioeconomic proxies (Hakhverdian and Mayne, 2012), thus enabling better testing of the two major explanations of compliance.

This study examines the role of political trust and wealth in citizen compliance with government mobility restrictions in Latin America during the COVID-19 pandemic. We employ a difference-in-differences (DD) design, like that used by Bargain and Aminjonov (2020a), henceforth BA, in their analysis of political trust and compliance across 233 regions (19 countries) in Europe between March and April 2020. BA found that ex ante political trust predicted variation in region-level compliance as measured by Google mobility data. Similarly, we use data from the annual Latin America Public Opinion Project survey (LAPOP, 2019), Google mobility (Google LLC, 2021), and the Oxford COVID-19 Government Response Tracker (OxCGRT) (Hale et al., 2022) to compare trust, wealth, and other underlying explanations of compliance with mobility restrictions across 282 regions (15 countries) in Latin America.

We find that the relationship between political trust and compliance during COVID-19 is different than in high-income countries because of the distinct ex ante relationships with government services across socioeconomic gradients. In Latin America, political trust predicted less compliance (more mobility), unlike in high-income countries. This relationship holds when controlling for epidemiological factors, mask mandates, and socio-demographic variables, including age and labor formality. When adding wealth to the model, it predicted increased compliance (less mobility), with the negative effect of political trust outweighing the positive effect of wealth during the period after lockdown until early April. Yet by June 2020, the magnitudes of the effects of political trust had waned, while the impact of wealth, especially for workplace mobility, remained consistent and significant throughout, highlighting two distinct mechanisms affecting compliance. These results contradict findings from high-income countries and may appear counterintuitive, suggesting alternative channels through which trust affects compliance in LMICs.

To explore these channels, we began with a stepwise procedure to identify underlying predictors of political trust and wealth. In line with the literature, we found links to education, social protection, and corruption, measured as experience with police bribes. We then replaced political trust and wealth with these variables and implemented a new DD analysis. For the first two weeks after lockdown, having benefited from social protection before the pandemic predicted increased mobility. Since emergency transfers and other protection schemes took longer to implement, we rule out the direct effect of government transfers on mobility immediately after lockdown (Gentilini et al., 2020), though we cannot discard the anticipatory effect of the government's announcement of support on compliance. Regardless, the effects of social protection began to fade after the first two weeks. Having benefited from government services may thus help explain the initial and large negative effects of political trust in the original model. Meanwhile, results for education replicated those of wealth, predicting increased compliance throughout. As for corruption, there was some evidence that having experienced bribes was related to increased workplace mobility two weeks after the lockdown, though the effect reversed the longer mobility restrictions were in place and became marginally significant.

One interpretation of these results aligns with findings from studies of natural disasters which predict that high political trust can decrease risk perception, which in turn reduces individual protective actions (Wachinger et al., 2012). In Latin America, political trust driven by increased social protection may have created a public health moral hazard whereby beneficiary citizens were more risk tolerant, especially during the first two weeks of the lockdown period when (i) governments announced but had not yet disbursed social support and (ii) information uncertainty about the virus was greatest. As the lockdown continued, socioeconomic variables (e.g., wealth or education) became more predictive because economic constraints hindered sustained compliance, despite emergency support from governments. Finally, previous experience with bribes may have initially contributed to decreased compliance because of a lack of respect for government authority and later increased compliance because of the inconsistent enforcement of mobility restrictions, which could have made citizens more susceptible to bribe requests when leaving the home.

We argue that our results are grounded in the Latin America's “inclusionary turn” (Kapiszewski et al., 2021), in which the region became a global pioneer in cash transfer programs both before and during the pandemic, although it remains an open question whether the generous nature of Latin America's social protection programs may have unintentionally spurred increased mobility or contributed to it, given that the region registered the globe's highest per capita case and mortality rates (Ritchie and Ortiz-Ospina, 2021). Despite some cross-country evidence of the effects of income supports on decreased mobility (Aminjonov et al., 2021), more focused randomized evaluations of the impact of emergency cash transfers in Peru (Bird et al., 2023) and Kenya (Brooks et al., 2022) in the first months of the pandemic suggest that support may have also stimulated small business activity and thus could have made recipient households more mobile and vulnerable to contagion and death.

This study makes contributions on two levels. First, it adds to our evolving understanding of citizen responses to the unprecedented government actions taken during the COVID-19 pandemic. While political trust and civic norms predicted compliance to COVID-19 shelter-in-place and social distancing policies across and within high-income countries (Bargain and Aminjonov, 2020a; Allcott et al., 2020; Barrios et al., 2020; Bazzi et al., 2021), the result does not hold among middle-income countries in Latin America, even when controlling for wealth and other covariates.2 Rather, these findings lend more robust support to the paradox of trust hypothesis (Wachinger et al., 2012) initially observed in Singapore (Wong and Jensen, 2020). While political trust increases compliance in some contexts, it may create a moral hazard in others, a possibility acknowledged for natural disasters but not fully recognized in the public health compliance literature (Van de Weerd et al., 2011; Blair et al., 2017; Vinck et al., 2019; Christensen et al., 2020).

More broadly, these results contribute to our general understanding of trust and compliance (Letki, 2006; Marien and Werner, 2018) by demonstrating the importance of taking into account welfare state configurations. For example, previous work indicates that the relationship between political trust and welfare spending among European countries resembles that of “twin peaks,” whereby two distinct mechanisms drive welfare spending in high-trust versus low-trust countries in Europe (Algan et al., 2016). In low-trust European countries, welfare spending is large because “uncivic” people expect to free ride or benefit without assuming their share of the cost. In high-trust European countries, “civic” people may support high taxes and benefits expansion when they are surrounded by other high-trust citizens. Likewise, our findings suggest that political trust and compliance among high-income versus low- and middle-income countries may operate differently based on distinct ex ante relationships with government services across socioeconomic gradients. More studies could explore this hypothesis.

Our findings are robust to different estimation models, including controls and measures of mobility, trust, and wealth (see Supplemental Material). Regardless, the region-level mobility data prevent us from testing the individual-level relationships between mobility, political trust, wealth, social protection, and corruption experiences. Though suggestive evidence is offered, we were also not able to test more directly whether political trust driven by social protection experiences fostered a public health moral hazard after the initial lockdown. Data permitting, future studies could examine these individual-level relationships.

The rest of this paper is organized as follows. Section 2 contextualizes trust dynamics in Latin America compared to high-income countries. Section 3 introduces the data and Section 4 describes the empirical approach. Section 5 shares the main results, while Section 6 explores mechanisms. Section 7 concludes.

2. Political trust in Latin America

Social and political trust are related but distinct constructs. Social trust runs along a particularized (i.e., people you know and people like you) to generalized (i.e., people you come across) continuum (Delhey et al., 2011; Newton et al., 2018). As for political trust, it includes two components. The first is institutional trust, which refers to confidence one has in supposedly impartial government entities like the police or courts. The second is trust in political actors, such as the president/prime minister, cabinet, or congress/parliament (Zmerli and Newton, 2017). Global cross-country evidence indicates that particularized social trust forms the foundation for generalized trust, and the latter is a basis for the formation of political trust (Newton and Zmerli., 2011).

The relationship between particularized and generalized trust is visible when mapping results from 77 countries using the World Values Survey (see Fig. 1 ). Notably, most Latin American countries sit in the bottom left quadrant (low particularized and low generalized trust) while European countries are located in the top right quadrant (high particularized and high generalized trust). A pattern is also seen for Eastern and Southern European countries located lower on the regression line, while Northern European countries sit at the top.3 Relatedly, Latin America has consistently registered among the lowest levels of political trust in the world (Catterberg and Moreno, 2006; Bargsted et al., 2017; Mattes and Moreno, 2018; Letki, 2018).

Fig. 1.

Generalized and particularized Trust in Europe and Latin America.

Source: World Value Survey 2017–2020.

Differences in social and political trust have been connected to per capita income and long-term economic growth (Knack and Keefer, 1997; Algan and Cahuc, 2013), short-term recessions (Searing, 2013), inequality (Algan and Cahuc, 2013), and ethnic fractionalization (Alesina et al., 1999; Alesina and La Ferrara, 2002; Miguel and Gugerty, 2005). Trust and corruption perceptions also correlate across countries (La Porta et al., 1997), although the relationship is complex and has been posited as both cause and effect (Morris and Klesner, 2010; You, 2018). Furthermore, simultaneous equation models indicate a feedback mechanism between social trust, corruption, and income inequality (Uslaner, 2018).

While these factors help explain low levels of social and political trust in Latin America, other particularities come into play and help contextualize how trust may influence compliance with COVID-19 shelter-in-place orders and related mobility restrictions. Latin America is a region of mostly middle-income countries seeking to consolidate its democratic institutions amid high degrees of historic ethnic fractionalization, while registering among the highest rates of inequality and perceived corruption in the world. Political trust in the region therefore may not operate in the same way as in high-income countries with more consolidated democracies (Bargsted et al., 2017). Globally, high-income countries, where average education levels are higher, have more absolute levels of political trust. However, within-country differences between political trust and wealth, especially in Latin America, do not follow similar patterns between countries. In fact, Latin America registers more within-country than between-country differences in political trust (Bargsted et al., 2017). Fig. 2 depicts the relationship between region-level political trust and two socioeconomic proxies in Latin America grouped by the country's human development index (HDI) category (very high, high, and medium). Although overall trust levels do not differ between the three HDI categories in our sample, regions within “very high” HDI countries exhibit a positive correlation between political trust and education or wealth, similar to Europe, while there is negative correlation within “high” and “medium” HDI countries.4

Fig. 2.

Socioeconomic Proxies and Political Trust in Latin America by Country HDI.

Note: Very High HDI: Argentina, Chile, and Uruguay. High HDI: Bolivia, Brazil, Colombia, Dominican Republic, Ecuador, Jamaica, Mexico, Paraguay, Peru. Medium HDI: El Salvador, Guatemala, and Honduras. Years of education are the average number of years of schooling for respondents in each region within the respective countries. Wealth index is the percentage of the people in each country region above the national median wealth index, based on LAPOP.

Several factors may explain the negative within-country correlation among LMICs in Latin America. Education can enhance political sophistication or the ability to discern and judge differences in corruption between institutions (Weitz-Shapiro and Winters, 2017). In this sense, Hakhverdian and Mayne (2012) find that education has both a conditional and conditioning effect on political trust (i.e., institutional trust and trust in political actors) and the relationship depends on the level of corruption in a society. In “corrupt societies” (e.g., Latin America), education is negatively related to political trust, while in “clean societies” (e.g., Europe) it is positively related. Furthermore, the negative effect of corruption on trust increases with more education. Thus, while some studies have found that country-level perception of corruption in Latin America does not decrease country-level rates of trust (Catterberg and Moreno, 2006; Bargsted et al., 2017), its effects are better seen when examining within-country differences.

The negative correlation between education (and by proxy wealth) and trust in Latin America may also relate to the region's “inclusionary turn.” Historically, Latin America has been characterized by extreme inequality and social exclusion. Social policies exhibited large gaps in coverage, failing to give rural, informal, and unemployed workers access to social protection. However, recent decades under democratic rule have fostered an unprecedented expansion of inclusionary reforms and policies in Latin America, engendering an “inclusionary turn” (Kapiszewski et al., 2021). In the early 2000s, social programs were extended in several policy areas, including health care, pensions, and income support (Huber and Stephens, 2012; Pribble, 2013; Garay, 2016), with Latin American governments becoming global pioneers in the design and implementation of conditional cash transfer (CCTs) programs (Sugiyama, 2011; De la O, 2015). These targeted programs not only helped to increase household consumption and reduce poverty in the region (Fiszbein and Schady, 2009; Stampini and Tornarolli, 2012), but also linked poor and excluded citizens with government institutions in a sustained way, allowing for the potential establishment of trust relations.

Evidence of the moderating influence of country type has been found in Europe, where the relationship between political trust and welfare spending resembles that of “twin peaks,” highlighting two mechanisms driving welfare spending in high- versus low-trust countries (Algan et al., 2016). In low-trust countries, welfare spending is large because “uncivic” people expect to free ride or benefit without assuming their share of the cost. In high-trust countries, “civic” people support high taxes and benefit expansion when they are surrounded by other high-trust citizens. Similar mechanisms may be at work in Latin America - a region with high levels of inequality, different reach of social protection programs, and weaker state capacity, all of which encourage middle- and high-income households to opt out of government services.

Fig. 3 highlights the relationship between government social expenditure and political trust by education terciles in Latin America from 2004 to 2018, grouped by country HDI category. Until 2015, a positive correlation existed between increased government social spending and political trust. Yet the country HDI groups diverge thereafter, with medium HDI countries experiencing reduced spending and trust, high HDI countries maintaining spending and trust levels, and very high HDI countries exhibiting an inverse relationship. Notably, very high HDI countries overall have higher rates of government social spending, largely because of more robust social protection systems that incorporate, akin to those in Europe, larger proportions of the population than in medium and high HDI countries.

Fig. 3.

Social Expenditure and Political Trust by Education Levels.

Note: Own elaboration based on LAPOP survey. Country social expenditure data come from United Nations Economic Commission for Latin America and the Caribbean (ECLAC) and the United Nations Development Programme's Human Development Index.

The different configuration of trust in Latin America, how it operates, and for whom, raises the empirical question of how political trust may affect compliance differently in Latin America compared to Europe during the COVID-19 pandemic.

3. Data

This study employs a difference-in-differences (DD) design, similar to that used by Bargain and Aminjonov (2020a) in the European COVID-19 context. Four data sources were used. The annual Latin America Public Opinion Project (LAPOP) survey included trust, demographic, and socioe-economic variables for 23,299 individuals in 282 regions across 15 countries in Latin America and Caribbean (LAPOP, 2019). Daily COVID-19 mortality figures were sourced from Our World in Data (Ritchie and Ortiz-Ospina, 2021). Lockdown stringency data came from the Oxford Covid-19 Government Response Tracker (OxCGRT) (Hale et al., 2022). Google mobility data provided region-level mobility (Google LLC, 2021).

3.1. Dependent variable: mobility

As in previous studies (Bargain and Aminjonov, 2020a; Barrios et al., 2020), we use Google COVID-19 mobility data which provide indicators for countries and regions worldwide. Mobility fields include (i) residence, (ii) workplaces, (iii) grocery and pharmacy, (iv) retail and recreation, (v) parks (public gardens, dog parks, beaches, etc.), and (vi) transit stations (public transport hubs such as subway, bus, train stations, etc.). Residence captures the percentage of time spent in the home (lower values equal more mobility), while the latter five categories measure mobility to designated places outside the home (higher values equal more mobility). The mobility types outside the home could be further categorized as labor (workplace), primary consumption (grocery and pharmacy), secondary consumption (retail and recreation, parks), and transportation (transit stations). The smallest geographic unit available is regions within the country.

Google calculated mobility using pre-COVID-19 levels in January and February 2020. In some cases, mobility data were not available for all regions or certain days. In these situations, we only included regions with 15 percent or less missing values or one day per week on average. Table 8.2.1 in the Suplementary Materials details the number of regions per mobility type and missing value threshold. The last column calculates the percentage of days without missing values for regions meeting the minimum 15 percent missing value threshold. Total missing values by days range between 1.2 percent (workplaces) to 4.2 percent (transit stations).5 We report results using residence and workplace mobility because they represent the upper and lower bound of the total number of regions and are the two mobility types with the least number of missing values. Furthermore, given realities in Latin America, parks and transit stations were not considered as suitable mobility measures, while retail and recreation confounded distinct activities. Meanwhile, grocery and pharmacies (primary consumption) mirror the results for residence (home) and workplaces (labor).6

Fig. 4 depicts the level of mobility patterns for residence and workplaces beginning on February 15. In mid-March, time in residence increases considerably while workplace mobility plummets.

Fig. 4.

Mobility trends in Latin America.

Note: Own elaboration based on Google Mobility Data.

3.2. Explanatory variables: trust, wealth, and stringency

Political Trust. Following established measures of political trust (Zmerli and Newton, 2017), we used three questions from the LAPOP survey to measure the construct. The survey asked to what extent (1–7, “not at all” to “a lot”) the respondent trusted political parties, the president, and parliament. Responses were added and standardized. Following BA, we aggregated the variable at the regional level by calculating the share of respondents in the region above or below the mean.7 We also conducted a series of robustness checks testing different constructions of the trust variable.8

Wealth Index. A principal component analysis (PCA) was used to construct a wealth or poverty index variable (McKenzie, 2005; Vyass and Kumaranayake, 2006), a method commonly used for the LAPOP survey. The following household items reported in the LAPOP survey were included: television, refrigerator, traditional telephone, cellphone, vehicle, washing machine, microwave oven, indoor plumbing, indoor bathroom, and computer (Cordova, 2009). The first component generates a variable that gives more weight to assets that vary across households, so that the most common assets have zero weight (McKenzie, 2005). Since wealth has relative features and distribution is not uniform throughout Latin America, we followed the same method used for aggregating the trust variable at the regional level and calculated relative wealth for each region as the percentage of people above or below the median wealth for each country. The median was used because of the wealth measure's non-normal distribution (see Supplemental Materials for details).

Policy Stringency. Policy stringency refers to the mobility restriction level mandated by national governments in response to the COVID-19 pandemic, as measured by the Oxford COVID-19 Government Response Tracker (OxCGRT) (Hale et al., 2022). A series of indicators were measured and rescaled to generate a score between 0 and 100, with 100 representing the highest degree of stringency. The composite index used is the daily average value. In Latin America, stringency jumped in mid-March after the World Health Organization (WHO) declared the pandemic, effectively initiating lockdown. Fig. 5 reports country-specific stringency trends across time for countries included from the LAPOP survey.9

Fig. 5.

Stringency Trends in Latin America.

Source: Own elaboration based on Stringency data from OxCGRT.

In Latin America, the lockdowns occurred even more abruptly than in Europe and were among the strictest and most protracted in the world early in the pandemic. All countries in the sample implemented severe restrictions around mid-March, irrespective of their levels of trust, wealth, and the number of COVID cases. Once these restrictions were implemented, the countries did not relax them until July, except for Uruguay. This finding suggests that pre-existing country wealth, trust, and COVID prevalence did not affect the probability of implementing mobility restrictions and that the mobility restrictions may be considered as exogenous.

We examined effects until the lockdowns were eased and mobility returned in the region, selecting August 2020 as the cutoff.

3.3. Covariates

COVID-19 Deaths. Daily reported deaths may serve as a public signal about the risk of non-compliance with shelter-in-place or social distancing orders. We obtained daily deaths from Our World in Data (Ritchie and Ortiz-Ospina, 2021). Following BA, we include this covariate in our model, using the country-level measure. We did not expect particularized trust to be predictive of mobility, which the results confirmed. Finally, we replicated the model with a political trust variable generated from the Latinobarometer survey, which included 14 countries for LAPOP, except Jamaica. Since the latter survey did not enable the creation of a robust wealth index, the estimations were made both with and without years of education as a wealth proxy. Regardless, we found consistent results and replicated the signs and significance of results for political trust.

Population Density. Population density is known to affect virus spread because of the likelihood of less physical distancing, with urban environments at higher risk of contagion. Like BA, we control for this factor by calculating regional population density. Since population figures are relatively stable, we constructed the variable based on the most recent estimates from each country's statistical department. To ease interpretation, we converted values to population density per 10,000 inhabitants.

Age. Age could affect mobility in at least two ways. First, the older the person is, the less economically active they may be. Second, the older the population, the more susceptible they are to complications and mortality from the virus, leading them to take more precautions. To control for these possible effects, we constructed a region age variable by following the same procedure as for trust and wealth and calculated the proportion of people in each region above the country median age, given the variable's non-normal distribution.

Mask Mandates. Although early in the pandemic there was some uncertainty regarding the effectiveness of masks, many countries implemented mask mandates. As the weeks and months passed, however, mask use became more common. Regardless, the existence of mandates may affect the citizen's risk aversion towards the virus and mobility. We therefore included a dummy control variable for mask mandates, capturing whether mask use was obligatory either in all public spaces or simply to leave the home.

Formality. The unemployment rate in high-income countries may reflect the health of the economy and other labor-related conditions and was thus used in the BA model. However, in LMICs where welfare schemes are historically weaker, the unemployment rate may not proxy economic health given the role of the informal sector, especially in Latin America, in absorbing the un- and under-employed. Using unemployment rates in Latin America is also problematic because of different methods of calculation in each country and region. Regardless, for comparison with the BA model we sought to include a labor formality measure. The LAPOP survey asks if the respondent or respondent's employer makes pension contributions for retirement. This question has been demonstrated as a strong proxy for formality (LAPOP, 2019). The aggregate region-level variable represents the percentage of people in each region who work in the formal sector.

4. Graphical evidence and empirical approach

We initially examine the graphical evidence to establish the existence of parallel trends prior to the lockdowns and then specify the DD models used to confirm the observed relationship.

Fig. 6a and b depict the parallel trends in mobility from one month before country lockdowns (February 15, 2020) to five months after (August 5, 2020). For ease of interpretation, high and low political trust and wealth were calculated using the variable mean as a cutoff. Across the four major mobility types, we see parallel trends until the first half of March, when trajectories diverge. On March 11, the World Health Organization's made an official declaration of the global pandemic. Immediately after this announcement, governments in Latin America swiftly imposed quarantines, as seen in Fig. 5. In our sample, the median OxCGRT stringency index was 30 on March 13, 50 on March 16, and 70 on March 19. All countries in our sample surpassed 50 between March 3 and March 19. We use March 16 as the regional lockdown rate because it was the day the country-level median in our sample reached 50. Regardless, results are robust to different lockdown dates between March 13 and March 19 (see Supplemental Material).

Fig. 6.

Note: Local polynomial fit of the daily variation across regions of Latin America and the Caribbean and with 95 percent confidence intervals (CI).

Following the BA model, our DD estimation exploits the panel nature of the data and tests for the impact of political trust on mobility over time using the following specification:

$$\text{Mobilit}{\mathrm{y}}_{\text{it}}={\alpha }^{\mathrm{T}}+{\beta }^T\text{Trus}{\mathrm{t}}_i+{\gamma }^T\text{PostTrus}{\mathrm{t}}_i+{\delta }^T\text{Deat}{\mathrm{h}}_{\mathit{it}-1}+{\eta }^T{X}_i+{\lambda }^T\text{Post}{\mathrm{X}}_i+{\zeta }_t^T+{\pi }_C^T+{\epsilon }_{\mathit{it}}$$ (1)

The treatment variable is the level of political trust for region i, constructed as previously detailed. Post is the treatment period. For direct comparison with BA, our initial estimation is for March 1 to April 5. To examine longer-term effects, we extended the treatment period to June 5 and August 5, respectively. γ^T denotes the double-difference estimator of interest, while α^T is the constant selection bias between regions. We control for day dummies ζ^T and country dummies π^T . To make our model comparable to BA, we include additional region-level controls represented by vector X_i, including population density, the number of deaths reported the previous day T, and a region-level formality indicator, as discussed in the previous section. However, to further strengthen the model we include two additional covariates not used by BA. In the months after the onset of the pandemic, it became apparent that mortality risk was greater the older the population and that masks could mitigate virus spread.10 The model was run with and without weighting for region and standard errors were cluster-bootstrapped at the region level since the panel data of daily mobility includes multiple region-level observations.

A further specification included region effects in the model, creating the following equation:

$$\text{Mobilit}{\mathrm{y}}_{\mathit{it}}={\alpha }^T+{\gamma }^T\text{PostTrus}{\mathrm{t}}_i+{\delta }^T\text{Deat}{\mathrm{h}}_{\mathit{it}-1}+{\lambda }^T\text{Post}{\mathrm{X}}_i+{\zeta }_t^T+{\pi }_i^T+{\epsilon }_{\mathit{it}}$$ (2)

The region effect is captured in dummies π^T, which absorb the pre-lockdown time invariant characteristics of population density, COVID-19 deaths the previous day, formality, age, and mask. Once again versions of the model were run with and without region weighting.

Subsequently, we rerun both estimations replacing formality with wealth:

$$\begin{array}{ccc}\hfill \text{Mobilit}{\mathrm{y}}_{\mathit{it}}& =& {\alpha }^T+{\beta }^T\text{Trus}{\mathrm{t}}_i+{\gamma }^T\text{PostTrus}{\mathrm{t}}_i+{\sigma }^T\text{Wealt}{\mathrm{h}}_i+{\theta }^T\text{PostWealt}{\mathrm{h}}_i\hfill \\ & & +{\delta }^T\text{Deat}{\mathrm{h}}_{\mathit{it}-1}+{\eta }^T{\mathrm{X}}_i+{\lambda }^T\text{Post}{\mathrm{X}}_i+{\zeta }_t^T+{\pi }_C^T+{\epsilon }_{\mathit{it}}\hfill \end{array}$$ (3)

$$\begin{array}{ccc}\hfill \text{Mobilit}{\mathrm{y}}_{\mathit{it}}& =& {\alpha }^T+{\gamma }^T\text{PostTrus}{\mathrm{t}}_i+{\theta }^T\text{PostWealt}{\mathrm{h}}_i\hfill \\ & & +{\delta }^T\mathit{Deat}{h}_{\mathit{it}-1}+{\lambda }^T\text{Post}{\mathrm{X}}_i+{\zeta }_t^T+{\pi }_C^T+{\epsilon }_{\mathit{it}}\hfill \end{array}$$ (4)

As with the original model, country and region effect models were examined for each period, with and without region reweighting and cluster-bootstrapped standard errors at the regional level. Given consistency of the results, we report the region fixed effects model without region reweighting.

5. Results

5.1. Political trust and mobility

Table 1 shares results for the effects of political trust on residential and workplace mobility over time in 2020 - March to April, March to June, and March to August. Results for columns A, D, and G offer a direct comparison with the BA results for Europe, while columns B, E, and H include age, and columns C, F, and I incorporate both age and mask mandates.

Table 1.

Residence & workplaces - political trust.

	Continuous trust panel DD (using daily regional mobility)
	01 march - 05 april			01 march - 05 june			01 march - 05 august
	(A)	(B)	(C)	(D)	(E)	(F)	(G)	(H)	(I)
Residence
Post x Political Trust	−14.171***	−14.531***	−13.982***	−10.330***	−10.512***	−10.301**	−9.182***	−9.397**	−8.759**
	(4.685)	(4.626)	(4.718)	(3.949)	(3.925)	(3.991)	(3.606)	(3.599)	(3.634)
Controls X:
# daily deaths (t-1)	−0.111***	−0.110***	−0.117***	0.004***	0.004***	0.004***	0.008***	0.008***	0.007***
	(0.018)	(0.018)	(0.017)	(0.001)	(0.001)	(0.001)	(0.001)	(0.001)	(0.002)
Population density	6.676***	6.473***	6.415***	7.361***	7.262***	7.341***	6.353***	6.235***	6.506***
	(1.992)	(2.031)	(1.994)	(1.790)	(1.802)	(1.824)	(1.811)	(1.835)	(1.911)
Age		14.271	12.702		7.001	6.757		8.394	7.569
		(11.362)	(11.261)		(8.594)	(8.652)		(7.215)	(7.366)
Mask mandate			1.926*			−0.477			−1.632
			(0.984)			(0.967)			(1.002)
Formality	−5.811**	−5.953**	−3.397	−10.331***	−10.403***	−10.634***	−8.714***	−8.796***	−9.616***
	(2.662)	(2.668)	(2.938)	(2.067)	(2.063)	(2.191)	(1.860)	(1.861)	(1.951)
Observations	6336	6336	6336	17,072	17,072	17,072	27,808	27,808	27,808
R-squared	0.885	0.886	0.889	0.878	0.878	0.878	0.836	0.837	0.837
Workplaces
Post x Political Trust	15.961**	18.364**	11.822**	15.732**	13.291**	14.494**	12.371**	13.431**	13.941**
	(7.420)	(7.557)	(7.185)	(5.939)	(6.088)	(6.013)	(5.646)	(5.777)	(5.712)
Controls X:
# daily deaths (t-1)	0.003	0.002	0.035	−0.008***	−0.008***	−0.009***	−0.014***	−0.014***	−0.014***
	(0.033)	(0.032)	(0.036)	(0.003)	(0.003)	(0.003)	(0.003)	(0.003)	(0.003)
Population density	−14.982***	−14.232***	−13.693***	−17.001***	−16.554***	−16.082***	−16.001***	−15.671***	−15.502***
	(4.281)	(4.456)	(4.270)	(3.302)	(3.336)	(3.199)	(3.596)	(3.640)	(3.560)
Age		−35.971**	−34.842**		−22.020*	−22.491**		−15.862	−15.912
		(15.251)	(14.433)		(11.251)	(11.091)		(9.909)	(9.846)
Mask mandate			−7.006***			−1.208			−0.372
			(1.599)			(1.529)			(1.645)
Formality	7.907**	8.511**	−0.639	11.431***	11.812***	12.001***	12.092***	12.364***	12.533***
	(3.508)	(3.535)	(3.849)	(2.697)	(2.710)	(2.705)	(2.581)	(2.597)	(2.596)
Observations	10,145	10,145	10,145	27,346	27,346	27,346	44,544	44,544	44,544
R-squared	0.884	0.885	0.891	0.853	0.853	0.854	0.806	0.806	0.806

Note: Difference-in-differences (DD) estimation of Google mobility index (for different types of activity as indicated) or index of time spent in private residence on trust data (LAPOP) using daily regional variation for the period from March 1 to April 5, June 5, and August 5, respectively, with continuous trust (regional trust measure, calculated as the proportion of people with trust scores above national average). We report the coefficient on Post x Trust, with Post a dummy indicating the average lockdown date (March 16, 2020). Estimations include the lagged daily number of COVID-19 fatalities, day dummies, region fixed effects and Post interacted with regional control variables. Robust standard errors in parentheses, cluster-bootstrapped at region level (1000 replications). Significance level: *** p < 0.01, ** p < 0.05, *p < 0.1.

Across all model versions and time periods, as the proportion of people in the region with political trust above the national average moves from 0 to 1, mobility increases on the 100-point mobility scale between 8.8 and 14.5 for residence (negative coefficients mean less time spent in the residence) and 11.8 and 18.4 for workplace mobility. While the magnitudes and significance were consistent for workplaces across the three periods, the magnitudes for residential mobility faded, though significance remained. These signs and magnitudes are the reverse of that found by BA in Europe. While more political trust decreases mobility in Europe, it increases mobility in Latin America.

During Europe's initial lockdown period, an increase in the deaths reported the day before predicted less mobility. In Latin America, the relationship was more complex. In the first two weeks after lockdowns, daily deaths predicted more time spent outside that home, but by the second and third periods daily deaths became predictive of less mobility with increasing magnitudes. As for workplace mobility, in the first period deaths were not predictive but became so with increasing magnitudes in the second and third periods. These results suggest that information on deaths may have initially been a noisy signal which did not take effect until reporting systems improved and the population learned the full risks of the virus.

As in Europe, population density consistently relates to less mobility across all time periods in Latin America. These effects are roughly twice as large for workplace mobility compared to residence. Although the effect of age on residential mobility was not significant at any point, it had a significant negative effect on workplace mobility in the first period, after which magnitudes waned, with significance disappearing by the third period. While we expected mask mandates to relate to increased mobility, they initially predicted less mobility for both residential and workplace mobility after the first two weeks before fading thereafter. Two interpretations are possible. First, the existence of mandates may have sent a signal of the seriousness of the virus in the first two weeks after the lockdown. Second, increased country adoption of mask mandates after the initial period may have muted detection of effects in this analysis. Finally, although the BA model did not reveal clear patterns for the impact of unemployment in Europe, regions in Latin America with higher formality rates were related to increased mobility.

5.2. Wealth, political trust, and mobility

Table 2 shares results for the effects of political trust and wealth on residential and workplace mobility from March to April, March to June, and March to August 2020.

Table 2.

Residence & workplaces - political trust and wealth.

	Continuous trust panel DD (using daily regional mobility)
	01 march - 05 april			01 march - 05 june			01 march - 05 august
	(A)	(B)	(C)	(D)	(E)	(F)	(G)	(H)	(I)
Residence
Post x Political Trust	−11.961**	−12.241**	−11.882**	−8.595**	−8.719**	−8.948**	−7.853**	−8.005**	−7.879**
	(4.992)	(4.974)	(4.919)	(4.277)	(4.284)	(4.242)	(3.863)	(3.888)	(3.893)
Post x Wealth index	7.271**	7.574**	7.400**	6.284**	6.417**	6.606**	4.907*	5.069*	5.011*
	(3.613)	(3.513)	(3.489)	(2.950)	(3.000)	(2.957)	(2.648)	(2.590)	(2.606)
Controls X:
# daily deaths (t-1)	−0.126***	−0.125***	−0.129***	0.000	0.001	0.002	0.005***	0.005***	0.005***
	(0.018)	(0.018)	(0.017)	(0.002)	(0.001)	(0.002)	(0.001)	(0.001)	(0.002)
Population density	4.726**	4.441**	4.527**	5.299***	5.174***	4.945***	4.703***	4.552***	4.566***
	(1.928)	(1.941)	(1.917)	(1.510)	(1.508)	(1.491)	(1.520)	(1.536)	(1.553)
Age		14.821	13.942		6.517	7.390		7.921	7.920
		(11.121)	(10.801)		(9.269)	(8.995)		(7.961)	(7.897)
Mask mandate			2.884***			1.419			−0.036
			(0.927)			(1.049)			(1.072)
Observations	6336	6336	6336	17,072	17,072	17,072	27,808	27,808	27,808
R-squared	0.885	0.886	0.890	0.874	0.874	0.875	0.834	0.834	0.834
Workplaces
Post x Political Trust	12.623*	14.751*	12.692*	8.176	9.442	11.012*	9.225	10.094*	11.023*
	(7.574)	(7.634)	(7.092)	(5.988)	(6.066)	(5.854)	(5.727)	(5.800)	(5.646)
Post x Wealth index	−10.892**	−11.801**	−12.372**	−11.641***	−12.183***	−12.424***	−9.687***	−10.060***	−10.231***
	(5.159)	(5.101)	(4.830)	(3.766)	(3.749)	(3.684)	(3.568)	(3.568)	(3.530)
Controls X:
# daily deaths (t-1)	0.035	0.037	0.041	−0.003	−0.003	−0.005	−0.008***	−0.008***	−0.010***
	(0.032)	(0.033)	(0.032)	(0.003)	(0.003)	(0.003)	(0.003)	(0.003)	(0.003)
Population density	−11.444***	−10.392**	−10.572**	−12.911***	−12.303***	−11.454***	−12.332***	−11.910***	−11.382***
	(4.308)	(4.497)	(4.441)	(2.819)	(2.873)	(2.740)	(3.109)	(3.171)	(3.087)
Age		−36.193**	−37.643***		−21.363*	−22.764**		−14.622	−15.591
		(15.102)	(14.091)		(11.522)	(11.153)		(10.522)	(10.283)
Mask mandate			−7.078***			−2.824*			−1.863
			(1.351)			(1.492)			(1.608)
Observations	10,145	10,145	10,145	27,346	27,346	27,346	44,544	44,544	44,544
R-squared	0.884	0.885	0.892	0.853	0.853	0.853	0.806	0.806	0.806

Note: Difference-in-differences (DD) estimation of Google mobility index (for different types of activity as indicated) or index of time spent in private residence on trust data (LAPOP) using daily regional variation for the period from March 1 to April 5, June 5, and August 5, respectively, with continuous trust (regional trust measure, calculated as the proportion of people with trust scores above national average). We report the coefficient on Post x Trust and Post x Wealth, with Post a dummy indicating the average lockdown date (March 16, 2020). Estimations include the lagged daily number of COVID-19 fatalities, day dummies, region fixed effects and Post interacted with regional control variables. Robust standard errors in parentheses, cluster-bootstrapped at region level (1000 replications). Significance level: *** p < 0.01, ** p < 0.05, *p < 0.1.

During the first two weeks after the lockdown, as the proportion of people in the region with political trust above the national average moves from 0 to 1, mobility increases on the 100-point mobility scale between 7.9 and 12.2 for residence (negative coefficient means less time spent in home) and between 8.2 and 14.8 for workplace mobility. Despite similar magnitudes, there is greater significance for residence, while workplace mobility is marginally significant. Wealth predicts decreased mobility across all time periods. However, for residential mobility the significance and magnitudes fade, while for workplace mobility the effect sizes are larger, remain consistent across time periods, and maintain marginal significance throughout. Results for deaths, population density, age, and mask mandates replicate those found in the previous model.

Table 3 shares results for all mobility types for the three time periods using the political trust and wealth region fixed effects model (columns C, F, and I in Table 2, since this estimation was robust to other specifications,see Supplementary Material.). On one hand, political trust results for grocery and pharmacy, which we consider as primary consumption, are consistent with residential and workplaces (labor) mobility, though wealth is not predictive. On the other hand, political trust is not predictive of retail and recreation, which we consider a confounded measure of secondary consumption (alongside parks), while wealth is predictive.

Table 3.

Effect of trust and wealth on alternative mobility types.

Panel difference in difference estimates of Post Trust	Residence	Workplaces	Grocery and pharmacy	Retail and recreation	Parks	Transit stations
	(i)	(ii)	(iii)	(iv)	(v)	(vi)
01 march - 05 april
Political trust	−11.881**	12.691*	19.411*	4.315	1.044	12.222
	(4.919)	(7.092)	(10.821)	(8.643)	(7.621)	(10.351)
Wealth index	7.400**	−12.372**	−3.834	−11.763**	−9.348*	−9.723
	(3.489)	(4.830)	(7.856)	(5.648)	(5.323)	(6.694)
Observations	6336	10,145	8351	8964	9540	7488
R-Squared	0.890	0.892	0.827	0.911	0.884	0.906
01 march - 05 june
Political trust	−8.948**	11.012*	23.282**	1.579	−4.220	2.554
	(4.242)	(5.854)	(10.232)	(7.537)	(7.217)	(8.105)
Wealth index	6.606**	−12.421***	−4.408	−10.131**	−6.908	−7.242
	(0.002)	(3.684)	(7.425)	(4.650)	(4.983)	(5.104)
Observations	17,072	27,346	22,503	24,153	25,705	20,176
R-Squared	0.875	0.853	0.786	0.877	0.854	0.896
01 march - 05 august
Political trust	−7.879**	11.023*	22.723**	2.172	−4.936	2.425
	(3.893)	(5.646)	(9.959)	(7.728)	(7.264)	(8.113)
Wealth index	5.011*	−10.232***	−4.882	−9.798**	−5.240	−6.411
	(2.606)	(3.530)	(7.177)	(4.817)	(4.939)	(5.098)
Observations	27,808	44,544	36,655	39,342	41,870	32,864
R-Squared	0.834	0.806	0.752	0.833	0.805	0.861

Note: Difference-in-differences (DD) estimation of Google mobility index (for different types of activity as indicated) or index of time spent in private residence on trust data (LAPOP) using daily regional variation for the period from March 1 to April 5 and August 5, respectively, with continuous trust (regional trust measure, calculated as the proportion of people with trust scores above national average). We report the coefficient on Post x Trust and Post x Wealth, with Post a dummy indicating the average lockdown date (March 16, 2020). Estimations include the lagged daily number of COVID-19 fatalities, day dummies, region fixed effects, and Post interacted with regional control variables. Robust standard errors in parentheses, cluster-bootstrapped at region level (1000 replications). Significance level: *** p < 0.01, ** p < 0.05, *p < 0.1.

In sum, wealth more consistently predicts increased compliance across time, as observed in both developed (Wright et al., 2020) and developing (Bargain and Aminjonov, 2020b) countries. Furthermore, these effects are sustained over several months in Latin America. However, even when controlling for weatlh, the effects of political trust on decreased compliance contradicts findings from high-income countries (Bargain and Aminjonov, 2020a; Barrios et al., 2020; Durante et al., 2021). We seek to explain these results by focusing on experiences with social protection and corruption and their relation to political trust.

6. Mechanisms

While most studies of political trust and compliance, conducted largely in high-income countries, indicate that trust increases compliance, some evidence suggests otherwise. An early study in Singapore posited that the high levels of political trust may have created a moral hazard whereby citizens were more risk tolerant given faith in government management of the crisis (Wong and Jensen, 2020). High political trust has also been linked to less protective actions in the face of natural disasters (Wachinger et al., 2012).

Other research has examined the evolving relationship between trust, risk perception, and compliance during the COVID-19 pandemic. A series of cross-sectional surveys in the United Kingdom find that trust in government was a consistent predictor of decreased perception of COVID-19 risk, while the heterogeneous effect of risk perception on protective health behaviors increased over time (Schneider et al., 2021). Thus, even in high-income countries, the dynamic relationship between political trust and compliance is not straightforward. In comparison, experiments administered to cross-sectional waves in Wuhan before and at the onset of the COVID-19 crisis revealed decreased trust, increased risk aversion, and greater sensitivity to risk effects in gain and loss domains (i.e., a linearization of the Prospect Theory value function) immediately after the lockdown, with transitory effects following the death of a famed whistleblower (Shachat et al., 2021). In line with prior studies of natural disasters, recessions, and wars as well as increasing evidence from the pandemic, our main results highlight how the relation between trust and risk are contingent on context.

The counterintuitive relationship between political trust and compliance in Latin America during the pandemic suggests either that political trust creates a moral hazard or political trust proxies other variables or underlying mechanisms that may affect compliance decisions in the region. We explore these possibilities by decomposing underlying predictors of political trust and wealth and testing the effects of predictors on compliance.

6.1. Decomposing political trust and wealth

We initially identified via a stepwise regression predictors of political trust and wealth, respectively. For the political trust stepwise model, we systematically introduced socio-demographic variables, including density, formality, wealth, and education. Next, we included measures of experiences with social protection, corruption, and crime, all of which are related in the literature to public goods provision (or lack thereof). The final regressions added interpersonal trust and associativity as controls.11 For the stepwise regression with wealth as the dependent variable, we followed the same sequence with the independent variables, except we included political trust as a predictor.

Since our main results were conducted at the regional level, all variables were aggregated at this level. Formality, wealth, and density were constructed as in the main results estimation. Education captured the number of years of schooling. Assistance was the response to whether the person or someone in the household had received any regular or periodic government benefits in the form of money, food, or products, not counting pensions. Corruption measured whether in the last 12 months a police officer had asked the respondent for a bribe. Security captures the level of safety people perceive in their neighborhood because of experience as the victim of an assault or robbery. To calculate associativity, we used the frequency of religious, school, and community association activity. Interpersonal trust was the degree of trust in people in the local community. Scores were standardized.

Results from Table 8.4.1 (see Supplemental Material) finds that the most important predictors of political trust are government assistance (positive) and police bribes (negative). Table 8.4.2 (see Supplemental Material) identifies the same two predictors as the most important for wealth, except with reverse signs. Unsurprisingly, density and education were also predictive. Government assistance was only marginally predictive of wealth, once controlling for the other variables.12 Given that the DD model used for the main results controlled for density and wealth, one may hypothesize that the predictive relationship between political trust and mobility could be driven by having either benefited from government assistance or suffered from police corruption. Next, we removed political trust and wealth from the DD model and replaced them with the three variables predictive of both - social protection, police bribes, and education.13

6.2. Social protection, corruption, and education

The main results indicated that political trust predicted more mobility (less compliance) immediately after lockdown, while wealth predicted less mobility (more compliance) in all time periods. The results for wealth suggest that people with greater economic necessity and less ability to smooth consumption may be more compelled to leave the home to work (Ravallion, 2020; Robalino, 2020) or engage in occupations they cannot perform in the home (Dingel and Neiman, 2020; Garrote et al., 2020). However, contrary to high-income countries, results for political trust indicate that it does not increase compliance (Bargain and Aminjonov, 2020a; Brodeur et al., 2020; Brzezinski et al., 2020), leaving two options. Either political trust creates a moral hazard (Wong and Jensen, 2020) or it proxies for different mechanisms underlying compliance in Latin America - or both.

Government assistance, which is highly correlated with political trust, may increase the risk tolerance of populations. For example, in the two weeks after the lockdowns, all 15 countries in the sample announced emergency social protection mechanisms, yet logistically had not begun mass distribution. These announcements may have generated a short-term public health moral hazard for those who had benefited previously from the government's social protection schemes, regardless of wealth or education. As for corruption experiences, which also relate to less political trust, experiences with bribes may lead people to disregard government mandates or, because of inconsistent application of the orders, may encourage them to comply more because of the risk of bribes for non-compliance if caught breaking mobility restrictions.

Table 4 reports results for social protection, police bribes, and education on a monthly basis between April and August 2020. For the first two weeks of the lockdown, as the proportion of the regional population benefiting previously from social protection moves from 0 to 1, time spent in the home decreases by 16.3, while workplace mobility increases 21.2 on the 0 to 100 scale. In subsequent months, the magnitudes and significance of the effects waned. Meanwhile, police bribes related to a 13.3 decrease in time spent in home and 33.6 increase in workplace mobility in the first two weeks after the lockdowns, though this effect disappears by early May. Interestingly, for bribes the magnitudes of the effects increase in subsequent months, reaching marginal significance. Finally, education as a wealth proxy predicts decreased mobility consistently across all time periods. These results replicate in part the main findings, while offering more insight on what may drive the short-term, dynamic results of political trust.

Table 4.

Education, social protection, and corruption.

	Panel DD (using daily regional mobility)
	01 march - 05 april	01 march - 05 may	01 march - 05 jun	01 march - 05 jul	01 march - 05 aug
Residential
Post x Assistance	−16.301***	−10.622***	−7.160*	−5.370	−4.081
	(4.226)	(3.822)	(3.791)	(3.603)	(3.444)
Post x Police bribe	−13.271**	2.961	6.629	5.182	4.332
	(5.660)	(4.930)	(4.776)	(4.515)	(4.248)
Post x Education	1.788***	1.684***	1.708***	1.721***	1.625***
	(0.329)	(0.298)	(0.295)	(0.287)	(0.276)
Controls X:
# daily deaths (t-1)	−0.126***	−0.011***	0.000	0.003**	0.003**
	(0.018)	(0.002)	(0.001)	(0.001)	(0.001)
Population density	4.290**	4.246***	4.060***	3.749***	3.541***
	(1.690)	(1.526)	(1.383)	(1.357)	(1.352)
Age	12.861	12.093	11.052	11.694	11.411
	(10.122)	(8.649)	(8.240)	(7.792)	(7.208)
Mask mandate	5.538***	3.143***	1.882*	0.944	0.414
	(1.121)	(1.043)	(1.049)	(1.031)	(1.016)
Observations	6336	11,616	17,072	22,352	27,808
R-squared	0.905	0.903	0.881	0.857	0.839
Workplaces
Post x Assistance	21.221***	8.343	7.070	7.216	5.732
	(5.878)	(5.344)	(5.317)	(5.170)	(4.964)
Post x Police bribe	33.622***	−4.987	−14.252*	−14.040*	−15.291*
	(9.353)	(8.619)	(8.609)	(8.489)	(8.232)
Post x Education	−2.370***	−2.056***	−1.547***	−1.278***	−0.956***
	(0.319)	(0.334)	(0.321)	(0.325)	(0.327)
Controls X:
# daily deaths (t-1)	0.058*	0.004	−0.002	−0.009***	−0.009***
	(0.034)	(0.004)	(0.003)	(0.003)	(0.003)
Population density	−11.541***	−11.162***	−11.961***	−12.351***	−12.443***
	(3.539)	(2.893)	(2.710)	(2.935)	(3.002)
Age	−35.061***	−32.092***	−23.852**	−19.661*	−15.764
	(13.122)	(11.331)	(10.422)	(10.143)	(9.765)
Mask mandate	−10.261***	−2.254*	−1.967	−1.381	−0.975
	(1.531)	(1.337)	(1.500)	(1.568)	(1.623)
Observations	10,145	18,605	27,346	35,802	44,544
R-squared	0.898	0.893	0.854	0.825	0.806

Note: Difference-in-differences (DD) estimation of Google mobility index (for different types of activity as indicated) or index of time spent in private residence on trust data (LAPOP) using daily regional variation for the period from March 1 to April 5, June 5, and August 5, respectively, with continuous trust (regional trust measure, calculated as the proportion of people with trust scores above national average). We report the coefficient on Post x Assistance, Post x Police bribe and Post x Education, with Post a dummy indicating the average lockdown date (March 16, 2020). Estimations include the lagged daily number of COVID-19 fatalities (cf. Our World in Data), day dummies, region fixed effects and Post interacted with regional control variables (population density). Robust standard errors in parentheses, cluster-bootstrapped at region level (1000 replications).

Significance level: *** p < 0.01, ** p < 0.05, *p < 0.1.

We interpret these results as evidence for a public health moral hazard. Social protection is only predictive at the beginning of the lockdown. If the variable were to capture fully economic vulnerability with which it correlates, then one would expect it to remain predictive when the economic constraints were greater for staying at home. Instead, both education and wealth capture across time the consistent effects of the economic ability to adhere to mobility restrictions. One alternative explanation is that the emergency cash transfers activated by the region's governments may have encouraged mobility among poor households early in the lockdown; however, mass transfers were not distributed so quickly (Gentilini et al., 2020). A second alternative is that the transfers, as designed, encouraged less mobility. Yet this would likely have affected the impact of the wealth and education variables, whose magnitudes and significance remained consistent throughout. Furthermore, some evidence from impact evaluations of emergency transfers suggest that transfers may have increased mobility by increasing the economic activity of recipients (Bird et al., 2023; Brooks et al., 2022). Unfortunately, data limitations do not allow further testing of these hypotheses.

7. Conclusion

At the beginning of the pandemic, the implementation of government measures was believed to require a level of citizen trust in government. Not much was known nor understood about the virus, how it spread, and the risks it created. Lower-income households in both developed and developing countries suffered from irregular cash flows and minimum savings. With lower-income employment concentrated in commerce and service sectors, more vulnerable citizens had less opportunities to work from home. People from these households simply could not afford to restrict mobility for extended periods, spurring the creation of emergency social protection (Gentilini et al., 2020).

While evidence suggests that political trust and wealth explain increased compliance in high-income countries, the mechanisms remain unclear and less evidence exists for LMICs, where the relationship between wealth and trust are often reversed. Our results provide evidence for how trust operates differently in Latin America, and perhaps other LMICs, compared to high-income countries. Political trust does not act in the same way as envisioned for high-income settings and may even serve as a proxy in Latin America for other mechanisms. Unlike in Europe, lower-income segments in Latin America exhibit higher levels of political trust. While wealth predicts compliance in the region, as seen in high-income countries, political trust counterintuitively predicts more mobility, suggesting the possibility that political trust increases risk tolerance. By identifying social protection as a main driver for political trust in Latin America, we offer evidence for the existence of a public health moral hazard during the immediate lockdown period in the region.

While further exploration is needed to trace the relationship between social protection and compliance during the pandemic, the results for political trust highlight the importance of understanding the distinct relationship between political trust and compliance in LMICs. This requires moving beyond between-country studies of trust and examining determinants of within-country trust based on government-citizen relationships. In other words, future studies of political trust and compliance should extend the empirical base beyond high-income countries because they likely lack external validity for LMICs, where the relationship between trust and compliance may differ because of distinct relationships across socioeconomic gradients with government services. With this knowledge, policymakers, especially in Latin America, could better craft social protection policies in the face of future public health or natural disaster emergencies.

Supplementary material

Supplementary material associated with this article can be found, in the online version, at 10.1016/j.jebo.2022.12.010.

Appendix A. Supplementary materials

Supplementary Data S1

Supplementary Raw Research Data. This is open data under the CC BY license http://creativecommons.org/licenses/by/4.0/

Data availability

• Data will be made available on request.