We present, to our knowledge, the largest study of COVID-19 hospital survival in Brazil. We show that survivors are younger¹⁸ with female preponderance¹⁹ and tend to have fewer comorbidities,²⁰ replicating worldwide findings. However, we also report a number of other important sociodemographic trends which are specific to Brazil.

There is significant regional variation in both case-mix and outcome. The high number of cases (both in absolute and per-capita terms) seen in SP, RJ and AM (Fig. 2) are interesting. Of note, these regions are important ports of entrance to Brazil: AM hosts the Free Economic Zone of Manaus and most of the international flights route through SP and RJ: in 2019 7.7 million international passengers landed in SP and 2.2 million in RJ²¹ (further details in the supplementary materials). Therefore, these states are characterized by a higher international circulation of people. Furthermore, both SP and RJ are characterized by a particularly high population density and the COVID-19 outbreak coincided with the rainy season in AM, which is usually associated with a peak in respiratory infections.

In order to explore regional differences we split the Brazilian states into socioeconomically comparable North and Central-South regions. The finding of a higher comorbidity burden in hospitalised patients in the North region is concordant with a lower life expectancy²² and is also borne out by differences in the average age of survivors/non-survivors between the North and Central-South regions, as well as the significantly larger percentage of non-survivors in the North. In addition, we found that, in both regions, survivors are more likely to be branco and that branco Brazilians are more likely to be admitted to ICU than pardo Brazilians. In other words, the increased death rate of pardo Brazilians seems to be due in part to non-ICU admission, raising concerns regarding the organization of public and private medical resources.

It is also noteworthy that the distribution of hospitalized COVID-19 patients between the Central-South (81%) and North (19%) regions in our data is discordant with the actual population sizes of these two regions (64% and 36% respectively).¹⁰ This highlights the diversity in the country and may, at least in part, be due to either intrinsically lower hospitalization rates in the North region and/or disproportionate impact of COVID-19 in populous areas such as São Paulo and Rio de Janeiro (both in the Central-South region), the populations of which total about 20 million people.

The disparity in comorbidity among hospitalized patients between the North and Central-South regions is also striking (Table I), with non-survivors displaying a greater prevalence of almost all comorbidities in the North: Only asthma and immunosuppression are below hospital prevalence 70%. By contrast, for the Central-South region, the only comorbidities affecting more than 70% of non-survivors are renal and neurological diseases. Whilst a number of structural explanations are of course possible, this is likely to reflect poorer health status overall in the North region which, again, has a substantially larger percentage of non-survivors.

The disproportionately large percentage of comorbidity-free survivors in the Central-South region (Fig. 3) is remarkable. We may speculate that this may be due to differences in comorbidity ascertainment either due to structural differences in the way data is collected - perhaps comorbidity data was less available from patients who were sicker at the time of presentation - or because less severe patients, perhaps with concerns regarding their comorbid health, presented to hospital preferentially in the Central-South region. It is interesting to observe that branco and pardo Brazilians have a similar number of comorbidities in these populations. Therefore, it seems that comorbidities are not associated with ethnicity in the group studied, but are correlated with regional socioeconomic development (education, income and health).

However, an interplay between ethnic and regional socioeconomic factors is apparent in Fig. 5 with younger pardo and preto Brazilians seemingly less likely to survive the disease compared to the branco population; the difference being more pronounced in the Central-South region. It is worth placing this in the context of the typical life-expectancy in Brazil of 76 years (as of 2017).²³ By way of comparison the life expectancy in the European Union is 80.9 years.²⁴ Note also that the average life expectancy varies significantly by region, being higher in the Central-South than in the North: Santa Catarina has a life expectancy of 79.4 years while that of Maranhao is 70.9. This provides a baseline for the trend shown in Fig. 5.

From the ethnic distribution of patients admitted to ICU shown in Table II it appears that branco Brazilians are more likely than pardo Brazilians to be admitted to ICU once hospitalised. Also, by comparing percentages of hospitalization with deaths, one sees that branco Brazilians are more likely to survive than pardo Brazilians. However, by comparing total hospitalization with deaths after ICU admission, one sees more similar proportions between both ethnicities.

It appears that there are substantial ethnic differences in the proportion of patients admitted to ICU, and this also varies between North and Central-South regions. The greater proportion of deaths without admission to ICU for pardo is noteworth. This is likely to reflect higher levels of access to private healthcare for branco as compared to pardo Brazilians as ICU admission policies are known to differ between public and private hospital set-tings.²⁵ This may suggest that the lower rate of admission to ICU may be a factor for the increased proportion of deaths among pardo Brazilians. Note that the distribution of comorbidities, symptoms and age does not show strong ethnic variations, especially between pardo and branco. That the proportions of the different ethnicities admitted to ICU with COVID-19 are similar to those in the full 2019 SIVEP-Gripe dataset¹⁴ in Table III suggests that this is not a specific feature of COVID-19.

At an individual level survival analysis showed that, after age, the most important factor for hospital mortality was being of pardo or, to a lesser extent, preto ethnicity with hazard ratios of 1.47(1.33-1.58) and 1.32(1.15-1.52) (95% CI) compared to the branco baseline. The other risk factors largely replicate worldwide findings although we see that male sex is perhaps slightly less of a risk factor compared to what has been found in other series.¹⁹ This ethnic inequity has important social roots and implications: pardo and preto Brazilians have, on average, less economic security, are less likely to be able to stay at home and work remotely, and comprise a significant proportion of health and care workers, making them disproportionately the most vulnerable to COVID-19.²⁶ We also performed the Cox regression separately on the patients in the North and South regions, on patients in São Paulo and also excluding the São Paulo and Amazonas states (the ones with the highest number of cases per capita, see Fig. 2) to test robustness of our findings to outliers: The results were qualitatively consistent and did not change our conclusions that ethnicity is a key risk factor (details in the supplementary materials).

Examining the magnitude of the random effect estimates, we see substantial variation in hazard by region. The states that belong to the North region tend to have higher hazard ratios than the ones belonging to the Central-South region, which provides additional justification for splitting Brazil into two socioeconomically similar sets. It is also in agreement with the significantly larger percentage of non-survivors in the North as shown in Table I. Incorporating number of ICU beds/ventilators and nurses per 100 million inhabitants for each region as proxies for physical availability of healthcare resources did not qualitatively change our result (supplementary materials) suggesting a more fundamental difference in healthcare access and trajectory.

Rio de Janeiro, despite high standards of education, income and health, has one of the highest hazard ratios, very similar to those seen for the less developed Pernambuco and Amazonas regions. It is possible that this represents the effect of a highly populous area on viral spread. At the same time, Table IV shows the ethnic distribution for the seven states with highest number of cases, together with their hazard ratios. States in the North region have a hazard ratio greater than unity and a high proportion of pardo Brazilians, while states in the Central-South region have a hazard ratio less than one and a higher proportion of branco Brazilians. The notable exception is Rio de Janeiro, which exhibits a much higher hazard ratio than the neighboring state of São Paulo and an ethnicity similar to the states in the North region.

Many preto Brazilians may identify themselves as pardo Brazilians.²⁷ For this reason, it is reasonable to consider the preto and pardo populations together. Indeed, as seen from our analysis, both ethnic groups share higher percentages of non-survivors and higher hazard ratios.

The results of our analysis can then be interpreted according to the interplay of two independent, but correlated, effects:

• mortality by COVID-19 increases going North: ‘vertical’ effect,
• mortality increases for the pardo and preto population: ‘horizontal’ effect.

We may speculate that the vertical effect is due to expected increasing levels of comorbidities (or poorly controlled comorbidities) and general healthcare access, which we might expect in regions such as the North where socioeconomic levels are lower. We similarly postulate that the horizontal effect is likely driven by the greater reliance on publicly-funded healthcare in pardo and preto communities whereas branco Brazilians on average have greater access to private healthcare. This may explain the lower ICU admission for pardo as compared to branco Brazilians as ICU admission policy and availability vary substantially between public and private hospitals.²⁵

For most states the vertical and horizontal effects are correlated, giving a larger cumulative mortality. Indeed, lower socioeconomic development correlates with a larger pardo and preto population. Again, Rio de Janeiro is an outlier with an ethnic composition (horizontal effect) that is similar to the states in the North region, but high levels of development (vertical effect) more akin to those of Central-South states. In other words, Rio de Janeiro violates the typical correlation between vertical and horizontal effects seen elsewhere.

Whilst we believe our work is the most comprehensive of its kind to date in Brazil, there are a number of limitations which are worthy of discussion. Limitations and possible biases in case ascertainment cannot be ruled out, in common with all observational / database research. Ethnicity is missing in 39% of our data. This is not specific to COVID-19 (32% for the full SIVEP-Gripe dataset) but we cannot be sure that this is not subject to bias. We have limited our analysis to patients who were hospitalised since testing in the community is more likely to be biased according to local factors. However, again here, we cannot be sure that the availability of testing practice is homogeneous even in this population. Indeed, the fact that a large fraction of patients that have tested positive are admitted to the hospital clearly shows that testing, at least as far as this dataset is concerned, is performed only when symptoms are severe, indicating in turn that the number of COVID-19 cases in Brazil is likely to be much higher than suggested by available data.²⁸²⁹

It is possible that health-seeking behaviour varies with both ethnicity and region and late presentation may be an important determinant of ultimate hospital outcome. We are not able to consider this in our analysis as physiological severity at hospital presentation / admission data is not available in the SIVEP-Gripe dataset. However, a recent UK study did not demonstrate an important effect of physiological severity¹⁶ at least for ICU mortality, suggesting a high degree of homogeneity at admission, at least for this group. We cannot comment on whether this applies to Brazil or for the hospital population as a whole, and this would be an interesting area for future research.

We conducted our study during the Brazilian outbreak in the hope of generating actionable results. However, the analysis of early data introduces the possibility of leadtime and outcome ascertainment bias, although since our data encompasses a period substantially longer than the typical hospital stay, we hope that this does not affect our results qualitatively in any substantial way.

Whilst we have focused on hospital mortality, it is important to appreciate that we do not have data on out-ofhospital mortality (which may be appreciable) and neither can we robustly address the question of access to hospital services both by region or ethnicity / socioeconomic status. As such, a consideration of hospital mortality is likely to substantially underestimate the true effects of COVID-19 and it is plausible to assume that healthcare availability inequities would be further amplified in patients who are not hospitalized.

In conclusion, we present evidence suggesting a higher risk of death among pardo and preto Brazilians and in the North region. We hope that this analysis assists the authorities in better directing and aligning their response to COVID-19 in order to protect pardo and preto Brazilians from their higher death risk from SARS-CoV-2. Our results also indicate that the states in the North and Northeast macroregions are more vulnerable to the COVID-19 pandemic, an issue that merits further urgent attention by the federal government.