key: cord-0988445-oaj73szb
authors: Notari, Alessio; Torrieri, Giorgio
title: COVID-19 transmission risk factors
date: 2021-12-28
journal: Pathogens and global health
DOI: 10.1080/20477724.2021.1993676
sha: 49cef6a9ddc639856640aa41a25e004f6776a663
doc_id: 988445
cord_uid: oaj73szb

We analyze risk factors correlated with the initial transmission growth rate of the recent COVID-19 pandemic in different countries. The number of cases follows in its early stages an almost exponential expansion; we chose as a starting point in each country the first day [Image: see text] with 30 cases and we fitted for 12 days, capturing thus the early exponential growth. We looked then for linear correlations of the exponents [Image: see text] with other variables, for a sample of 126 countries. We find a positive correlation, i.e. faster spread of COVID-19, with high confidence level with the following variables, with respective [Image: see text] -value: low Temperature ([Image: see text] ), high ratio of old vs. working-age people ([Image: see text] ), life expectancy ([Image: see text] ), number of international tourists ([Image: see text] ), earlier epidemic starting date [Image: see text] ([Image: see text] ), high level of physical contact in greeting habits ([Image: see text] ), lung cancer prevalence ([Image: see text] ), obesity in males ([Image: see text] ), share of population in urban areas ([Image: see text] ), cancer prevalence ([Image: see text] ), alcohol consumption ([Image: see text] ), daily smoking prevalence ([Image: see text] ), and UV index ([Image: see text] , 73 countries). We also find a correlation with low Vitamin D serum levels ([Image: see text] ), but on a smaller sample, [Image: see text] countries, to be confirmed on a larger sample. There is highly significant correlation also with blood types: positive correlation with types RH- ([Image: see text] ) and A+ ([Image: see text] ), negative correlation with B+ ([Image: see text] ). We also find positive correlation with moderate confidence level ([Image: see text] -value of [Image: see text] ) with: CO(2)/SO emissions, type-1 diabetes in children, low vaccination coverage for Tuberculosis (BCG). Several of the above variables are correlated with each other, and so they are likely to have common interpretations. We thus performed a Principal Component Analysis, to find the significant independent linear combinations of such variables. The variables with loadings of at least 0.3 on the significant PCA are: greeting habits, urbanization, epidemic starting date, number of international tourists, temperature, lung cancer, smoking, and obesity in males. We also analyzed the possible existence of a bias: countries with low GDP-per capita might have less intense testing, and we discuss correlation with the above variables.

The recent coronavirus pandemic is now spreading essentially everywhere in our planet. The growth rate of the contagion has, however, a very high variability among different countries, even in its very early stages, when government intervention is still almost negligible. Any factor contributing to a faster or slower spread needs to be identified and understood with the highest degree of scrutiny. In [1] the early, growth rate of the contagion has been found to be correlated at high significance with temperature T. In this work, we extend a similar analysis to many other variables. This correlational study could help further investigation to establish causal factors, and it can help policy makers in their decisions. See also [2] [3] [4] for relevant control measures on the epidemic. Some factors are intuitive and have been found in other studies, such as temperature [1, [5] [6] [7] [8] [9] [10] [11] ] (see also [12] for a different conclusion) and air travel [12, 13] ; we aim here at being more exhaustive and at finding also factors which are not 'obvious' and have a potential biological origin, or correlation with one. As discussed in detail in the Conclusions section, a few of the main driving factors that we will discuss have been examined in the literature previously, including by mechanistic models.

The paper is organized as follows. In section II, we explain our methods; in section III, we show our main results; in section IV, we show the detailed results for each individual variable of our analysis; in section V, we discuss correlations among variables; and in section VI, we draw our conclusions.

As in [1] , we use the empirical observation that the number of COVID-19 positive cases follows a common pattern in the majority of countries: once the number of confirmed cases reaches order 10 there is a very rapid growth, which is typically well approximated for a few weeks by an exponential. Subsequently, the exponential growth typically gradually slows down, probably due to other effects, such as: lockdown policies from governments, a higher degree of awareness in the population or the tracking, and isolation of the positive cases. The growth is then typically stopped and reaches a peak in countries with a strong lockdown/tracking policy.

Our aim is to find which factors correlate with the speed of contagion, in its first stage of free propagation. For this purpose we analyzed a datasets of 126 countries taken from [14] on 15 April 2020. We have chosen our sample using the following rules:

• We start analyzing data from the first day d i in which the number of cases in a given country reaches a reference number N i , which we choose to be N i ¼ 30 1 . • We include only countries with at least 12 days of data, after this starting point. • We excluded countries with too small total population (less than 300 thousands inhabitants).

We then fit the data for each country with a simple exponential curve NðtÞ ¼ N 0 e αt , with 2 parameters, N 0 and α; here t is in units of days.

Note that the statistical errors on the exponents α are typically only a few percent of the spread of the values of α among the various countries. For this reason, we disregarded statistical errors on α. The analysis was done using the software Mathematica, from Wolfram Research, Inc.

We first look for correlations with several individual variables. Most variables are taken from [15] , while for a few of them have been collected from other sources, as commented below.

We find no significant correlation of the COVID-19 transmission in our set of countries with many variables, including the following ones:

(1) Number of inhabitants;

(2) Asthma-prevalence;

(3) Participation time in leisure, social and associative life per day; (4) Population density; (5) Average precipitation per year; (6) Vaccinations coverage for: Polio, Diphteria, Tetanus, Pertussis, Hepatitis B; (7) Share of men with high-blood-pressure; (8) Diabetes prevalence (type 1 and 2, together); and (9) Air pollution ('Suspended particulate matter (SPM), in micrograms per cubic metre').

We find strong evidence for correlation with:

(1) Temperature (negative correlation, p-value 4:4 � 10 À 7 ); (2) Old-age dependency ratio: ratio of the number of people older than 64 relative to the number of people in the working-age (15-64 years) (positive correlation, p-value 3:3 � 10 À 6 ); (3) Life expectancy (positive correlation p-value 8:1 � 10 À 6 ); (4) International tourism: number of arrivals (positive correlation p-value 9:6 � 10 À 6 ); (5) Starting day d i of the epidemic (negative correlation, p-value 1:7 � 10 À 5 ); (6) Amount of contact in greeting habits (positive correlation, p-value 5:0 � 10 À 5 ); (7) Lung cancer death rates (positive correlation, p-value 6:3 � 10 À 5 ); (8) Obesity in males (positive correlation, p-value 1:2 � 10 À 4 ); (9) Share of population in urban areas (positive correlation, p-value 1:7 � 10 À 4 ); (10) Share of population with cancer (positive correlation, p-value 2:8 � 10 À 4 ); (11) Alcohol consumption (positive correlation, pvalue 0:0019); (12) Daily smoking prevalence (positive correlation, p-value 0:0036); (13) UV index (negative correlation, p-value 0:004; smaller sample, 73 countries); and (14) Vitamin D serum levels (negative correlation, annual values p-value 0:006, seasonal values 0:002; smaller sample, ,50 countries).

We also find strong evidence for correlation with blood types:

(1) RH + blood group system (negative correlation, p-value 3 � 10 À 5 ); (2) A+ (positive correlation, p-value 3 � 10 À 3 ); (3) B + (negative correlation, p-value 2 � 10 À 4 ); (4) A− (positive correlation, p-value 3 � 10 À 5 ); (5) 0-(positive correlation, p-value 8 � 10 À 4 ); and (6) AB− (positive correlation, p-value 0:028).

We find moderate evidence for correlation with:

(1) B− (positive correlation, p-value 0:013). 1 In practice, we choose, as the first day, the one in which the number of cases N i is closest to 30 . In some countries, such a number N i is repeated for several days; in such cases we choose the last of such days as the starting point. For the particular case of China, we started from January 16th, with 59 cases, since the number before that day was essentially frozen.

We find no significant correlation with:

(1) 0+;

(2) AB+.

We find moderate evidence for correlation with:

(1) CO 2 (and SO) emissions (positive correlation, pvalue 0:015); (2) Type-1 diabetes in children (positive correlation, p-value 0:0008); and (3) Vaccination coverage for Tuberculosis (BCG) (negative correlation, p-value 0:028).

Counterintuitively we also find correlations in a direction opposite to a naive expectation:

(1) Death-rate-from-air-pollution (negative correlation, p-value 3:5 � 10 À 5 ); (2) Prevalence of anemia, adults and children, (negative correlation, p-value 1:4 � 10 À 4 and 7: � 10 À 6 , respectively); (3) Share of women with high-blood-pressure (negative correlation, p-value 1:6 � 10 À 4 ); (4) Incidence of Hepatitis B (negative correlation, pvalue 2:4 � 10 À 4 ); and (5) PM2.5 air pollution (negative correlation, pvalue 0:029).

We performed also non-linear fits with a quadratic function of α for each variable and used the Akaike Information Criterion (AIC) for model comparison, ΔAIC;2Δk À 2Δ lnðLÞ, where Δk is the increase in the number of parameters, compared to the simple linear model and Δ lnðLÞ is the change in the maximum log-likelihood between the two models. Only for a few variables, the ΔAIC is slightly in favor (i.e. ΔAIC less than À 2) of the quadratic model: ΔAIC ¼ À 4; À 2:7; À 2:6; À 3:8; À 2:4; À 3:5; À 2:4 for life expectancy, number of tourist arrivals, starting date of the epidemic, obesity in males, urbanization, death rate from air pollution, Hepatitis B. None of these is large enough to strongly rule out the linear relation. We will therefore omit quadratic fits from now on.

We also find a correlation with GDP per capita, which should be an indicator of lack of testing capabilities. Note, however, that GDP per capita is also quite highly correlated with another important variable, life expectancy, as we will show in section V: high GDP per capita is related to an older population, which is correlated with faster contagion. Note also that correlation of contagion with GDP disappears when excluding very poor countries, approximately below 5 thousand $ GDP per capita: this is likely due to the fact that only below a given threshold the capability of testing becomes insufficient. Figure 1 We performed 2-variables fits, including GDP and each of the above significant variables, to check if they remain still significant. In section IV, we will show the results of such fits, and also the result of individual one variable fits excluding countries below the threshold of 5 thousand $ GDP per capita. We list here below the variables that are still significant even when fitting together with GDP.

In a 2-variable fit, including GDP per capita, we find strong evidence for correlation with: (1) Amount of contact in greeting habits (positive correlation, p-value 1:5 � 10 À 5 ); (2) Temperature (negative correlation, p-value 2:3 � 10 À 5 ); (3) International tourism: number of arrivals (positive correlation, p-value 2:6 � 10 À 4 ); (4) Old-age dependency ratio: ratio of the number of people older than 64 relative to the number of people in the working-age (15-64 years) (positive correlation, p-value 5:5 � 10 À 4 ); (5) Vitamin D serum levels (negative correlation, annual values p-value 0:0032, seasonal values 0:0024; smaller sample, ,50 countries). To be confirmed on a larger sample. (6) Starting day of the epidemic (negative correlation, p-value 0:0037); (7) Lung cancer death rates (positive correlation, pvalue 0:0039); and (8) Life expectancy (positive correlation, p-value 0:0048);

We still find strong evidence for correlation with blood types:

(1) RH + blood group system (negative correlation, p-value 1 � 10 À 3 ); (2) B + (negative correlation, p-value 2 � 10 À 3 ); (3) A− (positive correlation, p-value 1 � 10 À 3 ); and (4) 0-(positive correlation, p-value 3 � 10 À 3 );

We find moderate evidence for correlation with: 

We find moderate evidence for:

1. UV index (negative correlation, p-value 0:01; smaller sample, 73 countries); 2. Vaccination coverage for Tuberculosis (BCG) (negative correlation, p-value 0:023); 3. Obesity in males (positive correlation, p-value 0:02); 4. CO 2 emissions (positive correlation, p-value 0:02); 5. Alcohol consumption (positive correlation, pvalue 0:03); 6. Daily smoking prevalence (positive correlation, pvalue 0:03); and 7. Share of population in urban areas (positive correlation, p-value 0:04).

Counterintuitively we still find correlations with:

(1) Death rate from air pollution (negative correlation, p-value 0:002); (2) Prevalence of anemia, adults and children, (negative correlation, p-value 0:023 and 0:005); (3) Incidence of Hepatitis B (negative correlation, pvalue 0:01); and (4) Share of women with high-blood-pressure (negative correlation, p-value 0:03).

In the next section, we analyze in more detail the significant variables, one by one (except for those which are not significant anymore after taking into account of GDP per capita, but including Type 1 Diabetes). In section V, we will analyze crosscorrelations among such variables, and this will also give a plausible interpretation for the existence of the 'counterintuitive' variables.

We first analyze each individual variable, then we analyze variables that have a 'counterintuitive' correlation, and finally, we separately analyze blood types.

The average temperature T has been collected for the relevant period of time, ranging from January to mid April, weighted among the main cities of a given country, see [1] for details. Results are shown in Figure 2 and Table 2 .

We also found that another variable, the absolute value of latitude, has a similar amount of correlation as for the case of T. However, the two variables have a very high correlation (about 0.91) and we do not show results for latitude here. Another variable which is also very highly correlated is UV index and it is shown later.

This is the ratio of the number of people older than 64 relative to the number of people in the working-age (15-64 years). Data are shown as the proportion of dependents per 100 working-age population, for the year 2017. Results are shown in Figure 3 and Table  3 . This is an interesting finding, since it may suggest that old people are not only subject to higher mortality, but also more likely to be contagious. This could be either because they are more likely to become sick, or because their state of sickness is longer and more contagious, or because many of them live together in nursing homes, or all such reasons together.

Note that a similar variable is life expectancy (which we analyze later); other variables are also highly correlated, such as median age and child dependency ratio, which we do not show here. In analogy to the previous interpretation, data indicate that a younger population, including countries with high percentage of Table 2 . In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with temperature T. We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with temperature T and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the twovariable fit. children, is more immune to COVID-19, or less contagious. See also [16] for a subsequent analysis of the effects of the age structure on the transmission dynamics of the disease.

This dataset is for year 2016. It has high correlation with old-age dependency ratio. It also has high correlations with other datasets in [15] that we do not show here, such as median age and child dependency ratio (the ratio between under 19-yearolds and 20-to 69-year-olds). Results are shown in Figure 4 and Table 4 .

The dataset is for year 2016. Results are shown in Figure 5 and Table 5 . As expected, more tourists correlate with higher speed of contagion. This is in agreement with [12, 13] , that found air travel to be an important factor, which will appear here as the number of tourists as well as a correlation with GDP.

This refers to the day d i chosen as a starting point, counted from 31 December 2019. Results are shown in Figure 6 and Table 6 , which shows that earlier contagion is correlated with faster contagion. One possible Table 3 . In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with old-age dependency ratio, OLD. We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with OLD and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. interpretation is that countries which are affected later are already more aware of the pandemic and therefore have a larger amount of social distancing, which makes the growth rate smaller. Another possible interpretation is that there is some other underlying factor that protects against contagion, and therefore, epidemics spreads both later and slower.

A relevant variable is the level of contact in greeting habits in each country. We have subdivided the countries in groups according to the physical contact in greeting habits; information has been taken from [17]. (1) No or little physical contact, bowing. In this

We have arbitrarily assigned a variable, named GRE, from 0 to 1 to each group, namely GRE ¼ 0; 0:25; 0:5; 0:4; 0:8 and 1, respectively. We have chosen a ratio of 2 between group 2 and 3 and between group 4 and 5, based on the fact that the only difference is that about half of the possible interactions (men-women) are without contact. Results are shown in Figure 7 and Table 7 .

Note also that an outlier is visible in the plot, which corresponds to South Korea. The early outbreak of the disease in this particular case was strongly affected by the Shincheonji Church, which included mass prayer and worship sessions. By excluding South Korea from the dataset one finds an even larger significance, pvalue ¼ 6:4 � 10 À 8 and R 2 � 0:23. In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with vs. starting date of the analysis of the epidemic, DATE, as defined in the text. We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with DATE and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. 

This dataset refers to year 2002. Results are shown in Figure 8 and Table 8 . Such results are interesting and could be interpreted a priori in two ways. A first interpretation is that COVID-19 contagion might correlate to lung cancer, simply due to the fact that lung cancer is more prevalent in countries with more old people. Such a simplistic interpretation is somehow contradicted by the case of generic cancer death rates, discussed in section III, which is indeed less significant than lung cancer. A better interpretation is therefore that lung cancer may be a specific risk factor for COVID-19 contagion. This is supported also by the observation of high rates of lung cancer in COVID-19 patients [18] .

This refers to the prevalence of obesity in adult males, measured in 2014. Results are shown in Figure 9 and Table 9 . Note that this effect is mostly due to the difference between very poor countries and the rest of the world; indeed this becomes non-significant In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with level of contact in greeting habits, GRE, as defined in the text. We also show the pvalue, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with GRE and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. Figure 8 . Exponent α for each country vs. lung cancer death rates. We show the data points and the best-fit for the linear interpolation. Table 8 . In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with lung cancer death rates, LUNG. We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with LUNG and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. when excluding countries below 5 K$ GDP per capita. Note also that obesity in females instead is not correlated with growth rate of COVID-19 contagion in our sample. See also [19] for increased risk of severe COVID-19 symptoms for obese patients.

This is the share of population living in urban areas, collected in year 2017. Results are shown in Figure 10 and Table 10 . This is an expected correlation, in agreement with [20, 21] . 

This is the share of people with type-1 Diabetes in 0-19 years population taken from [22] . Results are shown in Figure 11 and Table 23 . Note, however, that significance decreases a lot when restricting to countries with GDP per capita larger than 5 K$ and goes even beyond 0.05 when combining when GDP in a two-variables fit. Such a correlation, even if not very robust, could be non-trivial and could constitute useful information for clinical and genetic research. See also [23] .

This dataset refers to year 2016. Results are shown in Figure 12 and Table 12 . Note that this variable is highly correlated with old-age dependency ratio, as discussed in section V. While the correlation with alcohol consumption may be simply due to correlation with other variables, such as old-age dependency ratio, this finding deserves anyway more research, to assess whether it may be at least partially due to the deleterious effects of alcohol on the immune system. In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with share of population in urban areas, URB, as defined in the text. We also show the pvalue, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with URB and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. Figure 11 . Exponent α for each country vs. prevalence of type-1 Diabetes (0-19 years). We show the data points and the best-fit for the linear interpolation.

This dataset refers to year 2012. Results are shown in Figure 13 and Table 13 . As expected this variable is highly correlated with lung cancer, as discussed in section V. We find that COVID-19 spreads more rapidly in countries with higher daily smoking prevalence. Note, however, that this becomes nonsignificant when excluding countries below 5 K$ GDP per capita. Correlation of α with smoking thus could be simply due to correlation with other variables or to a bias due to lack of testing in very poor countries. Alternative interpretations are that smoking has In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with alcohol consumption (ALCO). We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with ALCO and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. negative effects on conditions of lungs that facilitates contagion or that it contributes to increased transmission of virus from hand to mouth [24] . Interestingly, note that our finding is in contrast with claims of a possible protective effect of nicotine and smoking against COVID-19 [25, 26] .

This is the UV index for the relevant period of time of the epidemic. In particular, the UV index has been collected from [27] , as a monthly average, and then with a linear interpolation we have used the average value during the 12 days of the epidemic growth, for each country. Results are shown in Figure 14 and Table  14 . Not surprisingly in section V, we will see that such quantity is very highly correlated with T (correlation coefficient 0:93). Note also that here the sample size is smaller (73) than in the case of other variables, so it is not strange that the significance of a correlation with α here is not as high as in the case of α with Temperature. More research is required to answer more specific questions, for instance whether the virus survives less in an environment with high UV index, or whether a high UV index stimulates vitamin D production that may help the immune system, or In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with daily smoking prevalence (SMOK). We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with SMOK and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. both. Very recent results [28] consistently find that indeed SARS-CoV-2 is deactivated efficiently under simulated sunlight in few minutes, due to UVB.

Another relevant variable is the amount of serum Vitamin D. We collected data in the literature for the average annual level of serum Vitamin D and for the seasonal level (D s ). The seasonal level is defined as: the amount during the month of March or during winter for northern hemisphere, or during summer for southern hemisphere or the annual level for countries with little seasonal variation. The dataset for the annual D was built with the available literature, which is unfortunately quite inhomogeneous as discussed in Appendix A. For many countries, several studies with quite different values were found and in this case we have collected the mean and the standard error and a weighted average has been performed. The countries included in this dataset are 50, as specified in Appendix. The dataset for the seasonal levels is more restricted, since the relative literature is less complete, and we have included 42 countries.

Results are shown in Figure 15 and Table 34 for the annual levels and in Figure 15 and Table 34 for the seasonal levels. Note, however, that such results are preliminary and based on inhomogeneous data, and have to be confirmed on a larger sample.

Interestingly, in section V, we will see that D is not highly correlated with T or UV index, as one naively could expect, due to different food consumption in different countries. A slightly higher correlation, as it should be, is present between T and D s . Note that our results are in agreement with the fact that increased vitamin D levels have been proposed to have a protective effect against COVID-19 [29] [30] [31] . Figure 16 

This is the data for year 2017. We have also checked that this has very high correlation with SO emissions (about 0.9 correlation coefficient). We show here only the case for CO 2 , but the reader should keep in mind that a very similar result applies also to SO emissions. Note also that this is expected to have a high correlation with the number of international tourist arrivals, as we will show in section V. Results are shown in Figure 17 and Table 17 . In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with mean annual levels of vitamin D (variable name: D). We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with D and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. 

This dataset is the vaccination coverage for tuberculosis for year 2015. 2 Results are shown in Figure 18 and Table 18 . We find a negative correlation. Note that this depends mostly on the three countries with low coverage present in the plot: by excluding them the correlation becomes non-significant. Note also that several countries with no compulsory vaccination (such as USA or Italy) do not have an estimate for BCG coverage Figure 19 . Exponent α for each country vs. death rate from air pollution per 100,000. We show the data points and the best-fit for the linear interpolation.

and were not included in the analysis. Including them, with very low coverage, would probably affect a lot the significance. This correlation, even if not highly significant and to be confirmed by more data, is also quite non-trivial and could be useful information for clinical and genetic research (see also [32] [33] [34] ) and even for vaccine development [35-37]. Figure 23 

We show here other correlations that are somehow counterintuitive, since they go in the opposite direction than from a naive expectation. We will try to interpret these results in section V.

This dataset is for year 2015. Results are shown in Figure 19 and Table 19 . Contrary to naive expectations and to claims in the opposite direction [38] , we find that countries with larger death rate from air pollution actually have slower COVID-19 contagion.

This dataset is for year 2015. Results are shown in Figure 20 and Table 20 . Countries with larger share of high blood pressure in females have slower COVID-19 contagion. Note that we do not find a significant correlation instead with high blood pressure in males. In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with death rate from air pollution per 100,000 (POLL). We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with POLL and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. 

This is the incidence of hepatitis B, measured as the number of new cases of hepatitis B per 100,000 individuals in a given population, for the year 2015. Results are shown in Figure 21 and Table 21 . Countries with higher incidence of hepatitis B have slower contagion of COVID-19.

Prevalence of anemia in children in 2016, measured as the share of children under the age of five with hemoglobin levels less than 110 grams per liter at sea level. A similar but less significant correlation is found also with anemia in adults, which we do not report here. Results are shown in Figure 22 and Table 22 . The significance is quite high, but could be interpreted as due to a high correlation with life expectancy, as we explain in section V. A different hypothesis is that this might be related to genetic factors, which might affect the immune response to COVID-19.

Blood types are not equally distributed in the world and thus we have correlated them with α. Data were taken from [39] . Very interestingly, we find significant correlations, especially for blood types B+ (slower COVID-19 contagion) and A− (faster COVID-19 contagion). In general, also all RH-negative blood types correlate with faster COVID-19 contagion. It is interesting to compare with findings in clinical data: (i) our finding that blood type A is associated with a higher risk for acquiring COVID-19 is in good agreement with [40] , (ii) we find higher risk for group 0-and no correlation for group 0+ (while [40] finds lower risk for groups 0), and (iii) we have a strong significance for lower risk for RH+ In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with prevalence of anemia in children (ANE). We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with ANE and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. Figure 23 . Exponent α for each country vs. percentage of population with blood type A+. We show the data points and the best-fit for the linear interpolation.

In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with percentage of population with blood type A+. We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with A+ blood and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. types and in particular lower risk for group B+, which is probably a new finding, to our knowledge. These are also non-trivial findings which should stimulate further medical research on the immune response of different blood-types against COVID-19.

In this section, we first perform linear fits of α with each possible pair of variables (for blood types, we considered only RH+ and B+). We show the correlation coefficients between the two variables, for each pair, in In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with percentage of population with blood type 0 À . We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with 0 À and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. . We also show the p-value of the t-statistic of each variable in a pair and the total R 2 of such fits in Figure 31 . We give here below possible interpretations of the redundancy among our variables, and we perform multiple variable fits in the following subsections.

The set of most significant variables, i.e. with smallest p-value, which correlate with faster propagation of COVID-19 are the following: low temperature, high percentage of old vs. working people and life expectancy, number of international tourist arrivals, high In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with percentage of population with blood type AB−. We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with AB− and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. Table 29 . In the left top panel: best-estimate, standard error (σ), t-statistic, and p-value for the parameters of the linear interpolation, for correlation of α with percentage of population with RH-positive blood (RH+). We also show the p-value, excluding countries below 5 thousand $ GDP per capita. In the left bottom panel: same quantities for correlation of α with RH+ and GDP per capita. In the right panels: R 2 for the best-estimate and number of countries N. We also show the correlation coefficient between the 2 variables in the two-variable fit. percentage of RH-blood types, earlier starting date of the epidemic, high physical contact in greeting habits, and prevalence of lung cancer. Such variables are, however, correlated with each other and we analyze them together below. Most other variables are mildly/ strongly correlated with the previous set of variables, except for: prevalence of Type-I diabetes, BCG vaccination, Vitamin D levels (note, however, that the latter is based on more inhomogeneous data and should be confirmed on a larger sample), which might indeed be considered as almost independent factors. From the above table one can verify that all the 'counterintuitive' variables (death-rate due to pollution, POLL, prevalence of anemia, ANE, prevalence of hepatitis B, HEP, high blood pressure in women, PRE) have a strong negative correlation with life expectancy, LIFE. This offers a neat possible interpretation: since countries with high deaths by pollution or high prevalence of anemia or hepatitis B or high blood pressure in women have a younger population, then the virus spread is slower. Therefore, one expects that when performing a fit with any of such variable and LIFE together, one of the variables will turn out to be non-significant, i.e. redundant. Indeed, one may verify from Table 31 that this happens for all of the four above variables.

Redundancy is also present when one of the following variables is used together with life expectancy in a 2 variables fit: smoking, urbanization, obesity in males. Also, old age dependency and life expectancy are obviously quite highly correlated.

Other variables instead do not have such an interpretation: BCG vaccination, type-1 diabetes in children and vitamin D levels. In this case other interpretations have to be looked for. Regarding the vaccination Figure 30 . Correlation coefficients between each variable in a pair. Such coefficient corresponds to the off-diagonal entry of the (normalized) covariance matrix, multiplied by À 1. In the last column and row we show the p-value of each variable when performing a one-variable linear fit for the growth rate α. Note also that the fits that include vitamin D variables (D and D s ) and UV index are based on smaller samples than for the other fits and were collected with rather with inhomogeneous data, and so have to be confirmed on a larger sample, as explained in the text. a promising interpretation is indeed that BCGvaccinated people could be more protected against COVID-19 [32] [33] [34] .

Lung cancer and alcohol consumption remain rather significant, close to a p-value of 0.05, even after taking into account life expectancy, but they become non-significant when combining with old age dependency. In this case, it is also difficult to disentangle them from the fact that old people are more subject to COVID-19 infection.

Blood type RH+ is also quite correlated with T; however, it remains moderately significant when combined with it.

Finally vitamin D, which is measured on a smaller sample, also has little correlations with the main factors. This is quite interesting, since it may open avenues for research on protective factors and health policies. It is quite possible that high Vitamin D helps the immune response to COVID-19 [29] [30] [31] . Note, however, that this finding 

is based on more inhomogeneous data and should be confirmed on a larger and more homogeneous sample.

It is not too difficult to identify redundant variables, looking at very strongly correlated pairs in Table 30 . It is generically harder, instead, to extract useful information when combining more than 2 or 3 variables, since we have many variables with comparable predictive power (individual R 2 are at most around 0.2) and several of them exhibit mild/ strong correlations. In the following we perform examples of fits with some of the most predictive variables, trying to keep small correlation between them.

Here we show an example of a fit with three parameters.

Here we show an example of a fit with four parameters. As we combine more than three parameters, typically at least one of them becomes less significant.

One may think of combining all our variables. This is, however, not a straightforward task, because we do not have data on the same number N of countries for all variables. In both cases we fit with

where v i are linear combinations of the above variables.

In case (1), we find a fit with total R 2 ¼ 0:46, with N ¼ 103. There are, however, only six significant independent orthogonal linear combinations. Such six combinations, which have R 2 ¼ 0:42, are: 

where we omitted variables whose coefficients ('loadings') on v 1 ; v 2 ; v 3 ; v 4 ; v 6 are less than 0.1. The variables with loadings of at least 0.3 are: GRE, URB, DATE, ARR, T, LUNG, SMOK, and OBE. The significance of the principal components is given in Table 32 .

In case (2), we find a fit with total R 2 ¼ 0:48, with N ¼ 103. There are again only six significant independent orthogonal linear combinations. Such 6 combinations, which have R 2 ¼ 0:42, are: where, again, we omitted variables whose loadings are less than 0.1. The variables with loadings of at least 0.3 are now: HEP, GRE, ARR, ANE, OLD, T, CO 2 , and URB. The significance of the principal components is given in Table 33 . 

We have collected data for countries that had at least 12 days of data after a starting point, which we fixed to be at the threshold of 30 confirmed cases. We considered a dataset of 126 countries, collected on April 15th. We have fit the data for each country with an exponential and extracted the exponents α, for each country. Then we have correlated such exponents with several variables, one by one. We first performed a bivariate analysis, where we correlated the exponents with many different variables. However, many such factors are intercorrelated, and so we also performed a Principal Component Analysis (PCA), which leads to a number of few independent orthogonal combinations. From the bivariate analysis, we found a positive correlation with high confidence level with the following variables, with respective p-value: low temperature (negative correlation, p-value 4 � 10 À 7 ), high ratio of old people vs. people in the working-age (15-64 years) (p-value 3 � 10 À 6 ), life expectancy (p-value 8 � 10 À 6 ), international tourism: number of arrivals (p-value 1 � 10 À 5 ), earlier start of the epidemic (p-value 2 � 10 À 5 ), high amount of contact in greeting habits (positive correlation, p-value 5 � 10 À 5 ), lung cancer death rates (p-value 6 � 10 À 5 ), obesity in males (p-value 1 � 10 À 4 ), share of population in urban areas (p-value 2 � 10 À 4 ), share of population with cancer (p-value 2:8 � 10 À 4 ), alcohol consumption (p-value 0:0019), daily smoking prevalence (p-value 0:0036), low UV index (p-value 0:004; smaller sample, 73 countries), and low vitamin D serum levels (annual values p-value 0:006, seasonal values 0:002; smaller sample, ,50 countries).

We also find strong evidence for correlation with blood types: RH + blood group system (negative correlation, p-value 3 � 10 À 5 ); A+ (positive correlation, p-value 3 � 10 À 3 ); B + (negative correlation, p-value 2 � 10 À 4 ); A− (positive correlation, p-value 3 � 10 À 5 ); 0-(positive correlation, p-value 8 � 10 À 4 ); and AB− (positive correlation, pvalue 0:028). We find moderate evidence for correlation with: B− (positive correlation, p-value 0:013).

We find moderate evidence for positive correlation with: CO 2 (and SO) emissions (p-value 0:015), type-1 diabetes in children (p-value 0:023), vaccination coverage for Tuberculosis (BCG) (p-value 0:028).

Counterintuitively we also find negative correlations, in a direction opposite to a naive expectation, with: death rate from air pollution (p-value 3·10 −5 ), prevalence of anemia, adults and children, (p-value 1 � 10 À 4 and 7 � 10 À 6 , respectively), share of women with high-bloodpressure (p-value 2 � 10 À 4 ), incidence of Hepatitis B (pvalue 2 � 10 À 4 ), and PM2.5 air pollution (p-value 0:029).

As is clear from the figures, the data present a high amount of dispersion, for all fits that we have performed. This is of course unavoidable, given the existence of many systematic effects. One obvious factor is that the data are collected at country level, whereas many of the factors considered are regional. This is obvious from empirical data (see for instance the difference between the epidemic development in Lombardy vs. other regions in Italy, or New York vs. more rural regions), and also sometimes has obvious explanations (climate, health factors vary a lot region by region) as well as not so obvious ones. Because of this, we consider R 2 values as at least as important as pvalues and correlation coefficients: an increase of the R 2 after a parameter is included means that the parameter has a systematic effect in reducing the dispersion ('more data points are explained').

Several of the above variables are correlated with each other and so they are likely to have a common interpretation and it is not easy to disentangle them. The correlation structure is quite rich and non-trivial, and we encourage interested readers to study the tables in detail, giving both R 2 , p-values and correlation estimates. Note that some correlations are 'obvious', for example between temperature and UV radiation. Others are accidental, historical and sociological. For instance, social habits like alcohol consumption and smoking are correlated with climatic variables. In a similar vein correlation of smoking and lung cancer is very high, and this is likely to contribute to the correlation of the latter with climate. Historical reasons also correlate climate with GDP per capita.

Other variables are found to have a counterintuitive negative correlation, which can be explained due their strong negative correlation with life expectancy: death-rate due to pollution, prevalence of anemia, Hepatitis B and high blood pressure for women.

We also analyzed the possible existence of a bias: countries with low GDP-per capita, typically located in warm regions, might have less intense testing and we discussed the correlation with the above variables, showing that most of them remain significant, even after taking GDP into account. In this respect, note that in countries where testing is not prevalent, registration of the illness is dependent on the development of severe symptoms. Hence, while this study is about infection rates rather than mortality, in quite a few countries, we are actually measuring a proxy of mortality rather than infection rate. Hence, effects affecting mortality will be more relevant. Preexisting lung conditions, diabetes, smoking and health indicators in general as well as pollution are likely to be important in this respect, perhaps not affecting α per se but the detected amount of α. These are in turn generally correlated with GDP and temperature for historical reasons. Other interpretations, which may be complementary, are that comorbidities and old age affect immune response and thus may directly increase the growth rate of the contagion. Similarly it is likely that individuals with comorbidities and old age, developing a more severe form of the disease, are also more contagious than younger or asymptomatic individuals, producing thus an increase in α. In this regard, we wish to point the reader's attention to the relevant differences in correlations once we apply a threshold on GDP per capita. It has long been known that human wellness (we refer to a psychological happiness study [41] , but the point is more general) depends non-linearly on material resources, being strongly correlated when resources are low and reaching a plateau after a critical limit. The biases described above (weather comorbidity, testing facilities, preexisting conditions and environmental factors) seem to reflect this, changing considerably in the case our sample has a threshold w.r.t. a more general analysis without a threshold.

About pollution our findings are mixed. We find no correlation with generic air pollution ('Suspended particulate matter (SPM), in micrograms per cubic metre'). We find higher contagion to be moderately correlated only with and CO 2 /SO emissions. Instead, we find a negative correlation with death rates due to air pollution and PM2.5 concentration (in contrast with [38] , see also [42] ). Note, however, that correlation with PM2.5 becomes non significant when combined with GDP per capita, while CO 2 /SO becomes non significant when combined with tourist arrivals. Finally, death rates due to air pollution is also redundant when correlating with life expectancy.

Given the existence of intercorrelations among our factors, we performed a Principal Component Analysis, to disentangle the minimal number of significant nonredundant factors. We selected a sample of N = 103 countries, where: (1) we omitted all variables with a counterintuitive behavior, (2) we omitted Vitamin D, BCG, UV and Blood variables, as we did not have data for so many countries, (3) we also omitted LIFE, since its meaning is already described by the similar variable OLD. Therefore we used the following variables: 1, T, OLD, ARR, GRE, OBE, URB, SMOK, CO2, DATE, LUNG, GDP, and ALCO. As a result we found that only 6 linear orthogonal combinations are significant, and the variables with a loading of at least 0.3 on any of such 6 combinations are: GRE, URB, DATE, ARR, T, LUNG, SMOK, and OBE.

Including also the counterintuitive variables we find again only 6 significant linear orthogonal combinations, but the variables with a loading of at least 0.3 are now: HEP, GRE, ARR, ANE, OLD, T, CO 2 , and URB. Some of the above significant factors have been examined in the literature, with the status of a causal relationship explanation varying greatly. Some factors, such as GRE and ARR, quite obviously are expected to have some effect. To our knowledge we established for the first time at a global level. Some previous studies at a local level were made in [43, 44] .

URB is naively 'obvious' but a bit more subtle, since it does not refer to population density (which we found to be among the non-significant variables) but specifically to Urbanization. Previous studies have already stressed that urban areas are more subject to generic outbreaks [45] and also specifically to COVID-19 (see [21] for a study involving US cities). There are two main factors at work here: a 'trivial' one is that urban societies, both because of the nature of their economic activities and social factors are much more interconnected. The power-law scaling seen in comparative city studies [21] is exactly what one expects from the hypothesis that urban contagion patterns are determined by interconnectedness. However, the greater health-care density in urban areas [21] will also lead to a quicker identification and reporting of cases [20] , which would also explain part of the dependence we observe. Reporting bias described in [20] , together with the growth of the awareness of the pandemic (more intense use of social distances and masks in countries where the epidemic started later) over the period described in our work, is also at work for DATE (see also [46] for a study, subsequent to our work, that confirms similar conclusions).

LUNG, SMOK, and OBE are probably dominated by the bias toward symptomatic cases, and that has been extensively examined in the medical literature. To our knowledge, we established for the first time their correlation with COVID-19 spread at a global level. In particular, OBE is well known to be linked to a compromised immune system [47] , and it has been linked to the number of hospitalizations due to COVID-19 [48] . Similar but more general reasons (poor lung function and link to poor hygiene habits) could account for SMOKE [49] and the link to LUNG has been studied in detail, including by detailed modeling [50] . Which brings us to T, the variable which more clearly has an 'environmental' rather than 'social' or 'medical' origin. As shown in [51] , if the chemical reactions behind the viral activation obey usual kinetic equations, a strong positive correlation of temperature with virus growth is expected.

Some of the variables that we have studied cannot be arbitrarily changed but can be taken into account by public health policies, such as temperature, amount of old people and life expectancy, by implementing stronger testing and tracking policies, and possibly lockdowns, both with the arrival of the cold seasons and for the old aged population.

Other variables instead can be controlled by governments: testing and isolating international travelers and reducing number of flights in more affected regions; promoting social distancing habits as long as the virus is spreading, such as campaigns for reducing physical contact in greeting habits; campaigns against vitamin D deficiency, decrease smoking and obesity.

We also emphasize that some variables are useful to inspire and support medical research, such as correlation of contagion with: lung cancer, obesity, low vitamin D levels, blood types (higher risk for all RH-types, A types, lower risk for B+ type), and type 1 diabetes. This definitely deserves further study, also of correlational type using data from patients.

In conclusion, our findings could thus be very useful both for policy makers and for further experimental research.

Temperature dependence of COVID-19 transmission

An investigation of transmission control measures during the first 50 days of the COVID-19 epidemic in China

Transmission dynamics of COVID-19 in Wuhan, China: effects of lockdown and medical resources

Analysis of COVID-19 transmission in Shanxi Province with discrete time imported cases

Temperature decreases spread parameters of the new COVID-19 case dynamics

Temperature significant change COVID-19 Transmission in 429 cities

The role of absolute humidity on transmission rates of the COVID-19 outbreak

Spread of SARS-CoV-2 coronavirus likely to be constrained by climate

Will coronavirus pandemic dimin

High temperature and high humidity reduce the transmission of COVID-19

Temperature, humidity and latitude analysis to predict potential spread and seasonality for COVID-19

Ecological Study on COVID-19: associations between the early growth rate and historical environmental and socio-economic factors in 96 countries using GAM (Generalized Additive models

Effects of latency and age structure on the dynamics and containment of COVID-19

Clinical characteristics of COVID-19-infected cancer patients: a retrospective case study in three hospitals within Wuhan

Risk of COVID-19 for patients with obesity

A novel methodology for epidemic risk assessment of COVID-19 outbreak

COVID-19 Attack Rate Increases with City Size March 30, 2020. Mansueto Institute for Urban Innovation Research Paper

A nicotinic hypothesis for COVID-19 with preventive and therapeutic implications

Low incidence of daily active tobacco smoking in patients with symptomatic

UV radiation monitoring archive

Simulated sunlight rapidly inactivates SARS-CoV-2 on surfaces

Evidence that vitamin D supplementation could reduce risk of influenza and COVID-19 Infections and deaths

The role of vitamin D in the prevention of coronavirus disease 2019 infection and mortality

Connecting BCG vaccination and COVID-19: additional data

Correlation between universal BCG vaccination policy and reduced morbidity and mortality for COVID-19: an epidemiological study

Exposure to air pollution and COVID-19 mortality in the United States

Due to a mistake for Bangladesh (data with RH+ sum to more than 100%). we took data from

Happiness, income satiation and turning points around the world

Higher virulence of COVID-19 in the air-polluted regions of eight severely affected countries medRxiv

Spatial transmission of COVID-19 via public and private transportation in China

Impact of international travel and border control measures on the global spread of the novel 2019 coronavirus outbreak

Urbanisation and infectious diseases in a globalised world

Inferring the main drivers of SARS-CoV-2 global transmissibility by feature selection methods

Mechanisms and manifestations of obesity in lung disease

COVID-19 and obesity

Tobacco smoking a potential risk factor in transmission of COVID-19 infection

Modeling lung perfusion abnormalities to explain early COVID-19 hypoxemia

Mechanistic theory predicts the effects of temperature and humidity on inactivation of SARS-CoV-2 and other enveloped viruses

Vitamin D map developed by D.A. Wahl et al. on behalf of International Osteoporosis Foundation (IOF)

A global representation of vitamin D status in healthy populations

Vitamin D: an overview of vitamin D status and intake in Europe

Current vitamin D status in European and Middle East countries and strategies to prevent vitamin D deficiency: a position statement of the European Calcified Tissue Society

Vitamin d status in central Europe

Vitamin D status, parathyroid function, bone turnover, and BMD in postmenopausal women with osteoporosis: global perspective

the one in which the number of cases N i is closest to 30. In some countries, such a number N i is repeated for several days; in such cases we choose the last of such days as the starting point. For the particular case of China

Canadian journal of public health. Revue canadienne de sante publique

Sunlight and Vitamin D

Relationship of 25-hydroxyvitamin D with all-cause and cardiovascular disease mortality in older community-dwelling adults

A global representation of vitamin D status in healthy populations

Distribution Of Phenotypic And Genotypic Abo And Rhesus Blood Groups Among Bangladeshi Population

Tokyo-172 BCG vaccination complications

GT acknowledges support from FAPESP proc. 2017/06508-7, participation in FAPESP tematico 2017/05685-2 and CNPQ bolsa de produtividade 301432/2017-1. We would like to acknowledge Alberto Belloni, Jordi Miralda and Miguel Quartin for useful discussions and comments.

No potential conflict of interest was reported by the author(s).

This work was supported by the CNPQ Bolsa de produtividade [306152/2020-7]; FAPESP Bolsa de pesquisa [2021/ 01700-2].

http://orcid.org/0000-0002-8854-3987

We collected most data on vitamin D from [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] and from references therein. For a first dataset of 50 countries, we have collected annual averages. For many countries, several studies with different values were found, and in this case, we have collected the mean and the standard error (when available) and a weighted average has been performed.The resulting values that we have used are listed in Table  34 . The sample size here is smaller and such dataset is based on quite inhomogeneous research and thus should be confirmed by a more complete dataset.