Resilience of countries to COVID-19 correlated with trust

Up to 1 December 2020, 156 countries had exhibited at least one peak and then decay of cases/capita (of which 36 had experienced a second peak and decay), 151 countries had exhibited at least one peak and then decay of deaths/capita (of which 32 had experienced a second peak and decay), and 93 countries had sufficient testing data to determine at least one peak and then decay of cases/tests (of which 23 had experienced a second peak and decay). Time-series for all countries and the three metrics are shown in Supplementary Fig. 1. For resilience, having filtered cases of reasonably exponential decay for further analysis (r² ≥ 0.8) and included multiple instances of well-fitted recovery occurring in one country in the dataset, we obtain n = 177 decays for cases/capita, n = 159 for deaths/capita, n = 105 for cases/tests. In a few countries a minimum had not yet been reached by 1 December 2020, so the reduction dataset is smaller (cases/capita n = 165, deaths/capita n = 150, cases/tests n = 101).

Comparable resilience and reduction of cases and deaths

The relative measures of resilience (rate of decay) and (proportional) reduction of cases should be more reliably estimated than absolute case numbers but could still be biased by variations in testing intensity across time and space. Encouragingly, we find across countries and waves, resilience of cases/capita and cases/tests are strongly positively rank correlated (n = 100, (rho) =0.86, p < 0.0001) with linear correlation gradient 0.88 (r² = 0.94) indicating that cases/capita tend to decay slightly faster than cases/tests (Fig. 1a). Resilience of cases/capita and deaths/capita are positively rank correlated (n = 150, (rho) =0.61, p < 0.0001) with linear correlation (gradient 0.95, r² = 0.75) indicating cases tend to decay slightly faster than deaths (Fig. 1b). Reduction of cases/capita and cases/tests are also strongly positively rank correlated (n = 94, (rho) =0.83, p < 0.0001) with proportional reductions (linear correlation gradient 1.0, r² = 0.98; Fig. 1c). Reduction of cases/capita and deaths/capita are positively rank correlated (n = 136, (rho) =0.76, p < 0.0001) with linear correlation (gradient 1.06, r² = 0.96) indicating deaths tend to be reduced slightly more effectively than cases (Fig. 1d). We find that variations between countries in the pattern of testing intensity over time can bias resilience results for cases/tests (see Supplementary Discussion). Therefore, as considering cases/tests also restricts the sample size and does not qualitatively alter later correlation results (see Supplementary Table 1 and Supplementary Discussion), we focus on cases/capita and deaths/capita when considering resilience and reduction.

Resilience is only weakly correlated with resistance

One might expect that lower resistance (a higher peak) would lead to lower resilience (slower recovery), e.g. because having a greater peak fraction of a country’s population infected provides more sources of further infection. Conversely, greater peak per capita levels of infection and/or deaths might inspire more effective measures and greater social compliance with those measures to bring down infections and deaths. The problem here is reliably estimating the absolute measure of resistance. Peak deaths/capita data should be more reliable than peak cases/capita (despite some issues with attributing deaths) because many cases have gone undetected particularly during the first wave, yet somewhat surprisingly peak deaths/capita and peak cases/capita are strongly correlated (n = 150, (rho) =0.86, p < 0.0001). There are weak positive correlations between resistance and resilience for deaths/capita, cases/capita, and cases/tests (Supplementary Fig. 2), but high resistance corresponds to a very wide range of resilience, and some countries with low resistance have relatively high resilience, particularly for deaths/capita. Given the weak relationship between resistance and resilience and the problems estimating resistance, we proceed with an independent treatment of resilience from hereon.

A threshold level of resilience is necessary for successful reduction

As would be expected mathematically, reduction (from peak to next minimum) is strongly positively correlated with resilience, following a non-linear relationship (Fig. 2). Really high resilience >  ~ 0.1 d⁻¹ (half-life <  ~ 1 week) tends to end in near complete reduction, but few countries have achieved this level of resilience. Instead several countries still achieve a near complete reduction of cases or deaths if they have a resilience of >  ~ 0.02 d⁻¹ (half-life <  ~ 1 month). Below that threshold level of resilience, reduction inevitably drops. Thus, poor resilience leads to failure to eliminate cases and deaths.

Resilience varies hugely between countries

Resilience of cases/capita, measured as magnitude of decay rate, ranges by a factor of ~ 40, from 0.16 d⁻¹ (Mauritius; most resilient) to 0.0041 d⁻¹ (Costa Rica; least resilient), corresponding to a half-life of ~ 4 to ~ 170 days (Fig. 3a). Resilience of cases/tests also ranges by a factor of ~ 40 (see Supplementary Discussion). Resilience of deaths/capita, ranges by a factor of ~ 25 from 0.10 d⁻¹ (Slovakia; most resilient) to 0.0042 d⁻¹ (Indonesia, Mexico, Romania; least resilient) (half-life ~ 7 to 165 days) (Fig. 3b).

Pairwise correlation results are summarised in Table 1 and are robust to analysing just the first peaks in each country (Supplementary Table 2) or using a more stringent fitting of exponential decay (r² ≥ 0.9; Supplementary Table 3)—both of which reduce the sample size.

Temporal but not spatial correlations

One might expect countries experiencing waves of infections and deaths earlier in the COVID-19 pandemic to have shown less resilience, due to being caught off-guard and having less collective knowledge about how to combat the spread of infections and reduce deaths. However, several of the countries hit earliest were ones with prior experience of the SARS-CoV-1 outbreak. We find negative correlations between timing (day of year) of peak cases/capita and resilience or reduction of cases/capita, and between timing of peak deaths/capita and resilience or reduction of deaths/capita (Table 1, Fig. 4)—i.e. those hit later tended to recover slower and less completely. Potential reasons for this are examined further below. One might also expect countries in closer spatial proximity could negatively influence one another’s resilience, e.g. through cross-border movement of infected individuals. However, long-distance international travel also clearly spread the virus early on, and subsequent restrictions on travel between countries should have reduced causal interactions. Variograms of distance between countries and difference in cases/capita resilience or deaths/capita resilience, show no evidence for spatial autocorrelation of resilience (Supplementary Fig. 3).

Wealth and public health are only weakly correlated with resilience

Demographic and public-health-related factors may be expected to influence country-level resilience, given that some significantly influence spread of the infection in within-country analysis¹⁰. However, no particularly strong controls emerge (Table 1). There are weak negative correlations between population and resilience of cases/capita or deaths/capita, and between country area and resilience of cases/capita or deaths/capita, but these counteract, leaving no significant effect of population density (Table 1). Richer populations (GDP/capita) tend to have greater median age (n = 175, (rho) =0.84, p < 0.0001), life expectancy (n = 179, (rho) =0.85, p < 0.0001), human development index (HDI) (n = 178, (rho) =0.96, p < 0.0001), and hospital beds (per 1000) (n = 161, (rho) =0.61, p < 0.0001). This leads to shared significant weak positive correlations of GDP/capita, median age, life expectancy, HDI, and hospital beds (per 1000) with resilience of cases/capita (Table 1). However, only hospital beds (per 1000) show a significant weak positive correlation with resilience of deaths/capita, and none of these factors significantly correlate with reduction of cases/capita or deaths/capita (Table 1).

Adaptive changes in stringency are positively correlated with resilience

Deliberate government interventions to limit social contact and thus R_e are expected⁹ to result in greater resilience (faster decay of cases and deaths). However, when looking for relationships with the OxCGRT ‘stringency index’^24,25, we only found a weak positive relationship between decay stringency (averaged over the fitted cases/capita decay intervals) and resilience (rate of decay) of cases/capita, no relationship for deaths/capita, and weak negative relationships with reduction of cases/capita and deaths/capita (Table 1). These effects are weak because most countries maintained a similar, near maximum stringency whilst cases and deaths were being brought down, yet they exhibited very differing resilience (recovery rates). Mean stringency (averaged across the whole time series) is significantly negatively correlated with resilience of cases/capita and especially deaths/capita (Table 1). Background stringency (averaged over the intervals when decay is not occurring) is significantly and more strongly negatively correlated with resilience of cases/capita and deaths/capita, and especially with reduction of cases/capita and deaths/capita (Table 1). Only adaptive stringency (the change in stringency from before to during decay intervals) has a significant positive correlation with resilience of cases/capita and deaths/capita, with significant but weaker positive correlations to reduction of cases/capita and deaths/capita (Table 1; Fig. 5a,b). Thus, deploying stringent measures decisively when an epidemic wave erupts is beneficial. However, governments that maintain greater background and overall (mean) stringency tend to have slower recovery and tend to be less effective at reducing cases and deaths.

Trust is positively correlated with resilience

Trust is significantly positively correlated with resilience of cases/capita and deaths/capita and especially with reduction of cases/capita and deaths/capita (Table 1; Fig. 5c,d). There is a clear threshold effect whereby all countries with trust > 40% have sufficient resilience to end in a large or complete reduction of cases and deaths (Fig. 5c,d; Supplementary Fig. 4a, b). Reduction distributions for trust ≤ 40% and trust > 40% are significantly different (cases/capita Mann–Whitney p < 0.001; deaths/capita Mann–Whitney p < 0.0001; Supplementary Fig. 4c,d). Trust and adaptive stringency are not significantly correlated, and in each case when controlling for one of the variables the resilience residuals remain strongly positively correlated with the other variable. (For cases/capita resilience, the adaptive stringency correlation (rho) goes from 0.428 to 0.449 and the trust correlation (rho) goes from 0.432 to 0.472. For deaths/capita resilience, the adaptive stringency correlation (rho) goes from 0.519 to 0.549 and the trust correlation (rho) goes from 0.398 to 0.388.) Trust is negatively correlated with mean stringency (n = 71, (rho) =−0.44, p < 0.001) and background stringency (n = 67, (rho) =−0.47, p < 0.0001), which may help explain why governments with greater background stringency are less effective at reducing COVID-19 cases and deaths—because they tend to reflect less trusting societies. Trust supports economic growth²⁹ and hence has a well-known³⁰ positive correlation with GDP/capita (n = 72, (rho) =0.70, p < 0.0001), which leads to positive correlations of trust with median age (n = 73, (rho) =0.55, p < 0.0001), life expectancy (n = 74, (rho) =0.57, p < 0.0001), HDI (n = 73, (rho) =0.69, p < 0.0001), and hospital beds (per 1000) (n = 71, (rho) =0.45, p < 0.0001). Pairwise correlation results suggest that trust exerts a stronger control than any of these factors on resilience or reduction (Table 1), but this could be influenced by the smaller sample of countries with trust data.

Linear models confirm trust and adaptive stringency both contribute to resilience

To examine this further we built various multiple linear regression models for resilience and reduction, with different mixes of social and demographic factors. Trust and adaptive stringency are consistently retained as the most significant beneficial factors. A model for resilience of cases/capita considering trust, adaptive stringency, GDP/capita, population, and hospital beds, retains adaptive stringency and trust as the most significant beneficial factors, followed by hospital beds, and rejects GDP/capita (Fig. 6a, Supplementary Table 4). A model for resilience of deaths/capita considering the same factors retains adaptive stringency and trust as the most significant beneficial factors, and GDP/capita as detrimental (Fig. 6b, Supplementary Table 5). A model for reduction of cases/capita retains trust and adaptive stringency and trust as the most significant beneficial factors, followed by hospital beds, with GDP/capita as detrimental (Fig. 6c, Supplementary Table 6). A model for reduction of deaths/capita retains adaptive stringency and trust as the most significant beneficial factors (Fig. 6d, Supplementary Table 7). If decay stringency and background stringency are used in place of adaptive stringency, they tend to be retained with significant but opposing effects, but less variance is explained, despite the extra factor (compare Supplementary Tables 8–9 with Supplementary Tables 4–5). These results confirm that trust and adaptive stringency are beneficial to resilience and reduction of both cases/capita and deaths/capita. They also suggest that trust gives rise to the significant pairwise positive correlation of GDP/capita and cases/capita resilience (Table 1) rather than vice versa.

Confidence in politics and government are not correlated with resilience

We examined whether resilience correlates with confidence in organisations pertinent to the social contract. Trust is positively correlated with confidence in politicians (n = 74, (rho) =0.45, p < 0.0001), parliament (n = 74, (rho) =0.49, p < 0.0001), government (n = 72, (rho) =0.34, p < 0.01), and elections (n = 44, (rho) =0.39, p < 0.01). However, there are no significant (p < 0.05) correlations between confidence in any of these organisations and resilience of cases/capita or deaths/capita. Only confidence in parliament has a weak positive correlation with reduction of cases/capita (n = 72, (rho) =0.24, p < 0.05). Hence, we did not consider these factors further in linear models or include them in Table 1.

Trust has more significant effects than any of Hofstede’s cultural dimensions

We also considered whether resilience correlates with any of Hofstede’s six cultural dimensions¹⁴ of power distance, individualism, uncertainty avoidance, masculinity, long-term orientation, and indulgence (defined above). Power distance (expectation from the less powerful that power is distributed unequally) is anti-correlated with trust (n = 49, (rho) =−0.70, p < 0.0001) and, consistent with that, anti-correlated with resilience of cases/capita and deaths/capita (Table 1). Individualism is positively correlated with trust (n = 49, (rho) =0.59, p < 0.0001) and less strongly with resilience of cases/capita, and deaths/capita (Table 1). Long-term orientation (pragmatism and preparation for the future) is positively correlated with trust (n = 63, (rho) =0.34, p < 0.0001) and resilience of cases/capita but not deaths/capita (Table 1). Uncertainty avoidance is negatively correlated with trust (n = 49, (rho) =−0.43, p < 0.01) and resilience of deaths/capita but not cases/capita (Table 1). Masculinity and indulgence do not show significant pairwise correlations with trust or resilience. Including Hofstede’s six cultural dimensions in place of trust in multiple linear regression models allows us to analyse a larger set of countries but explains less of their variance (compare Supplementary Tables 10–13 with Supplementary Tables 4–7). Mixing trust and the Hofstede dimensions in the models always retains trust as more significant than any of the retained Hofstede dimensions (Supplementary Tables 14–17).

Temporal pattern of the pandemic

We now return to interpreting the marked decline in resilience over time as the pandemic spread to new countries (Fig. 4). Adaptive stringency tended to decline with time; it is strongly negatively correlated with day of year of peak cases/capita (Fig. 7a) and day of year of peak deaths/capita (Fig. 7b). This is largely because background stringency increased as the pandemic progressed; it is positively correlated with day of year of peak cases/capita (n = 167, (rho) =0.44, p < 0.0001) and of peak deaths/capita (n = 156, (rho) =0.39, p < 0.0001), and secondarily because decay stringency has a weak anticorrelation with day of year of peak cases/capita (n = 166, (rho) =−0.16, p < 0.05), and day of year of peak deaths/capita (n = 154, (rho) =−0.17, p < 0.05). As the pandemic progressed it also tended to move to less trusting populations, in poorer countries, with worse healthcare. Trust is negatively correlated with day of year of peak cases/capita (Fig. 7c) and peak deaths/capita (Fig. 7d). GDP/capita is anti-correlated with day of year of peak cases/capita (n = 170, (rho) =−0.30, p < 0.0001) and peak deaths/capita (n = 154, (rho) =−0.34, p < 0.0001). Hospital beds (per 1000) are anti-correlated with day of year of peak cases/capita (n = 156, (rho) =−0.35, p < 0.0001) and peak deaths/capita (n = 143, (rho) =−0.30, p < 0.001).

These trends of declining adaptive stringency with time, and moving to less trusting populations, in poorer countries, with worse healthcare over time, could explain the decline in resilience over time (Fig. 4). To examine this further, we added day of year of peak as an extra factor considered in the linear models. Adding day of year of peak cases/capita to our model for resilience of cases/capita (Fig. 6a, Supplementary Table 4), it is retained whereas adaptive stringency is rejected (n = 71, r² = 0.436; Supplementary Table 18, Supplementary Fig. 5a). In contrast, adding day of year of peak deaths/capita to our model for resilience of deaths/capita (Fig. 6b, Supplementary Table 5), retains both with adaptive stringency more significant (n = 67, r² = 0.562; Supplementary Table 19, Supplementary Fig. 5b). Adding day of year of peak cases/capita to our model for reduction of cases/capita (Fig. 6c, Supplementary Table 6), it is retained whereas adaptive stringency is rejected (n = 66, r² = 0.445; Supplementary Table 20, Supplementary Fig. 5c). Adding day of year of peak deaths/capita to our model for reduction of deaths/capita (Fig. 6d, Supplementary Table 7), it is rejected leaving the model unaltered. If we compare the residuals of the original models (Supplementary Tables 4–7) to the relevant day of year of peak we find no significant (p < 0.05) correlations, suggesting that we are not missing a significant additional factor correlated with day of year. Rather, the anti-correlation of day of year with adaptive stringency, trust, and other factors in the models can sometimes replace them or reduce their significance. Hence the marked declines in resilience and reduction over time (Fig. 4) are variously linked to trends of declining adaptive stringency over time (Fig. 7a,b) and the pandemic tending to progress to countries with lower trust (Fig. 7c,d), GDP/capita, and hospital beds.

Source: Ecology - nature.com