Socio-economic determinants of global COVID-19 mortalities: policy lessons for current and future pandemics.

This paper examines the different socio-economic determinants of the fatalities associated with the COVID-19 pandemic globally in social determinants of health frameworks. It adapts the Poisson pseudo-maximum-likelihood (PPML) and the quantile regression techniques to effectively exploit the non-linear estimates of the data in order to derive non-biased point estimates at each quantile and make interquantile comparisons. This is particularly useful in recommending which societal variables become most significant at catastrophic levels of a pandemic like COVID-19 when existing health systems become overwhelmed. These estimators are applied to panel data for 196 countries over days of infection from the first recorded case. The COVID-19-related data is from Our World in Data, and the socio-economic variables are from the World Bank's World Development Indicators. The results establish that an improved adequate health infrastructure for both testing and treatment is necessary, but not sufficient. Health systems ultimately become overwhelmed and ineffective in managing cases and reducing mortality in the face of the rising pandemic. Complementary social, economic, physical and environmental factors are necessary for curbing deaths. These factors relate to improving the health stock of the population through reductions in both communicable and non-communicable comorbidities; enhancing sanitation and hygiene; and improving the nutrition of the population. Socio-economic and environmental measures are the reduction of household and ambient air pollution; reduction of exposure to alcohol and cigarettes; reduction of poverty and ensuring economic inclusion; and learning from the past to fine-tune governments' control measures in order to minimize harm to the population while effectively curbing mortality.

KEY MESSAGES

An improved adequate health infrastructure for both testing and treatment is necessary, but not sufficient for curbing COVID-19 deaths.

Health systems quickly become overwhelmed and ineffective in managing cases and reducing mortality in the face of the rising pandemic.

Complementary social, economic, physical and environmental factors are necessary to minimize mortalities.

The complementary actions found to be significant relate to:

improving the health stock of the population through reduction in both communicable and non-communicable comorbidities; enhancing sanitation and hygiene; improving nutrition of the population.

socio-economic and environmental measures are the reduction of household and ambient air pollution; reduction of exposure to alcohol and cigarettes; reduction of poverty and ensuring economic inclusion.

learning from the past to fine-tune governments’ control measures in order to minimize harm to the population while effectively curbing mortality.

Introduction

Since the announcement of the appearance of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), which causes the coronavirus disease 2019 (COVID-19), the world has witnessed significant stress. On December 30, 2019, the Chinese Wuhan local health authority issued an epidemiological alert following the first documented case on December 8, which led to the closure of the Huanan seafood market two days later (Huang ). The World Health Organization (WHO) declared the disease a global pandemic on March 11, 2020 (Cucinotta and Vanelli 2020).

According to the European Centre for Disease Prevention and Control (ECDC 2020), as of May 5, 2020, the number of laboratory-confirmed cases globally had risen above the 3.5 million mark with over 250 000 recorded deaths, and a spread to over 210 countries globally. However, not all is gloom as the recovery cases have far outweighed the deaths, standing at over a million in the same period. Figure 1 shows the global evolution of fatalities and infections to date (August 31, 2020).

Figure 1

Global trend of cumulative COVID-19 deaths.

Such a rapid rise in fatalities implies that biomedical systems alone will not cope with the challenge. The broader determinants of health, which include social, economic, physical and environmental factors (Horton 2013; de Andrade ) can prove to be more crucial at times like these.

The rapid rise in infections and fatalities have prompted most governments to respond, first by closing their international borders. This has generally been followed by more stringent social distancing measures and subsequent lockdown of the socio-economic machinery. While these measures are aimed at flattening the evolution curve, they are inevitably leading to immediate acute and economic hardship, which may persist long after the pandemic. This problem is much more precarious for most developing countries, with limited resources, high levels of socio-economic vulnerability and poverty coupled with weak governance mechanisms. Some authors such as Anderson have highlighted the impossibility for such governments to be able to minimize COVID-19-related mortality and the associated economic impacts. Whatever happens to mortality relating to the pandemic, and any associated measures to curb it would depend on the broader determinants of health, which would be the social, economic, physical and environmental contexts where the pandemic is evolving. The identification of the key socio-economic factors that affect the fatalities of the disease becomes important in curbing the current mortality due to COVID-19 and possibly future similar pandemics. It is in this manner that this paper contributes to the debate of the determinants of COVID-19 mortality. This analysis will suggest lessons to help countries build resilience through selecting which socio-economic and environmental factors to focus on in order to limit the fatalities of the current and future pandemics.

Methodological framework and econometric strategy

The determinants of deaths due to the COVID-19 pandemic can be assessed within the framework of social determinants of health. The WHO (2020a) classifies the social determinants of health into the conditions in which people are born, grow, work, live and age; and the wider forces and systems that define the conditions of their daily lives. In today’s literature, social aspects of health explain substantially more of health outcome changes than simply medical care does (Horwitz ; Marmot and Wilkinson 2006).

Following the literature (Horton 2013; de Andrade ; Marmot ; Viner ), I specify a formal framework that links disease outcome to various social, economic, physical and environmental contexts of a country in which individuals find themselves. The main problem I face in specifying the model is that while the socio-economic variables are time-invariant, changing only across countries, and measured annually, the dependent variable—COVID-19 mortality—is measured daily, and varies across both countries and time. The only way I can explore this data is to consider the different socio-economic variables as fixed country-specific factors. The specified model is as follows:

Where is an indicator of health, in our case cumulative deaths for country i at day t; X is a vector of time-varying variables and Z a vector of time-invariant country-specific variables; is the intercept and are coefficients to be estimated, day is the number of days from the first case detected in a country and is a classical mean zero disturbance. The nature of our data is such that the time-varying variables are the cumulative tests, cumulative infections and the number of days from first detected case.

Variables and data

The dependent variable for the mortality is the cumulative deaths related to COVID-19 infections. The data is a day by country panel of 196 countries with reported cases from first day of detected case in China (December 31, 2019) to current date at which data was harnessed (August 31, 2020). All the variable acronyms beginning with L denote the log of the variable. I expressed the mortality variable as a thousand of total infections and take the log (LTD). There are concerns around the accuracy of COVID-19 mortality. To allay such concerns, I have undertaken a robustness check by estimating the models using logged excess deaths (LED) as dependent variable in place of COVID-19 deaths. The challenge with excess deaths is that data is available for a limited number of countries. I nevertheless use it because the subsample is able to give useful insight into how sensitive my models are to the measurement of COVID-19 mortality. Due to negative values in excess deaths, I rescaled it so that the lowest value is zero.

It is common knowledge that the testing process is crucial for early detection, and isolation for treatment. Therefore, testing is an important factor in the real-time monitoring of transmission. This is vital for subsequent containment mortality, since early detection provides higher chances of survival (Yuan ). To control for testing, I use total number tested expressed per thousand of the population. The log of the variable (LTT) is used. A higher number of infections also leads to more fatalities hence I also control for infections (LTI).

An important control variable is the Government Response Stringency Index (LST). It is a composite measure based on nine response indicators including school closures, workplace closures and travel bans, rescaled to a value from 0 to 100 (100 = strictest response). This variable is reported together with COVID-19-related variables. All COVID-19-related variables, i.e. infections, deaths and tests are taken from Our World in Data (Ritchie 2020).

The social, economic, physical and environmental variables capture the conditions that prevailed in a country immediately following the disease outbreak. All these indicators are from the World Bank’s (2020) World Development Indicators (WDI) except the obesity prevalence from the WHO (2020b). The selection of these variables for the model is based on the literature (Horwitz ; de Andrade ; Horton 2013; Marmot ; Viner ; Marmot and Wilkinson 2006) and what we know matters the most for mortality. Once the first cases have been recorded in a country, which usually comes through international contacts, what becomes important in fatalities is local population dynamics, health infrastructure endowments, sanitary environment and local population health stock. After importing a few cases, there is the second stage when local inter-personal transmission kicks in. This stage depends on a number of factors. One is population dynamics, mainly density and movements including internal displacements. I use national population density (LNPD), urban population density (LUPD), the population share of over-55s (LP55), log of median age (LMA), and internally displaced people (LIDD). Due to strong collinearity, not all these variables could enter the models at once. However, the index based on the Principal Components Analysis (PCA) of the variables (LPD) was used in a separate model.

can expose individuals to disease and fatality in two ways. First, most substances like alcohol happen in certain social settings that facilitate contact with potential sources of disease. The second channel is the effect of these substances on the health of the individual (GBD 2018; Burton and Sheron 2018; Pryor 1997; Slattery ; Zhang ). I use alcohol consumption per capita (LAC) and smoking prevalence for age 15+ (LSMOKE). The PCA of both variables is used as an index of substance consumption (LSC).

largely determines the rate of mortality from a disease in a country, as it shows the possibility of effectively taking charge of cases and caring for the morbid. The main variable used for health infrastructure is the (logged) number of hospital beds per thousand population (LHB). However, I use the PCA of all the variables plus others that underlie health infrastructure, including various health expenditures, life births attended by skilled personnel, to generate an index of health infrastructure (LHI).

Underlying health conditions of an individual can be a strong determinant of mortality from COVID-19. Key diseases are both communicable (e.g. malaria, HIV and TB) and non-communicable (diabetes, hypertension, etc.). Following a strong suggestion in the literature of the roles of overweight/obesity and diabetes (GBD 2017; Lighter ; Popkin ), I include a diabetes prevalence indicator (LDB) and obesity prevalence (LOB). I have also used the prevalence of various other diseases namely, logs of malaria (LMAL), tuberculosis (LTB) and HIV (LHIV).

The PCA of these, together with other communicable diseases and related causes of death (see Table 2) were used as an index of communicable comorbidity (LCCM). I also used the PCA of cause of death by non-communicable diseases (% of total) (LMNCD) and diabetes prevalence to generate an index of non-communicable comorbidity (LNCM).

Table 2

PCA index variables and description

Acronym	Meaning (log)	Description (PCA of the following variables)
LPD	Population dynamics index	National population density; urban population density; population in the largest city; number of internally displaced persons
LSANI	Sanitation index	Rural and urban populations with basic drinking water services; rural and urban populations with basic sanitation services; population with safely managed drinking water services
LHI	Health infrastructure index	Current health expenditure; current health expenditure per capita; domestic general government health expenditure; domestic private health expenditure per capita; births attended by skilled health staff; external health expenditure; external health expenditure per capita; nurses and midwives per 1000 people; physicians per 1000; hospital beds per 1000
LCCM	Index of communicable	Cause of death, by communicable diseases; incidence of HIV per 1000; incidence of malaria per 1000; incidence of tuberculosis per 100 000; prevalence of HIV (female, male and total % ages 15–24)
LNCM	Index of non-communicable	Cause of death, by non-communicable diseases (% of total); diabetes prevalence (% of population ages 20–79)
LMN	Malnutrition index	Prevalence of severe wasting, weight for height (female, male and total % under 5); prevalence of stunting, height for age (female, male and total % under 5); prevalence of undernourishment (% of population); prevalence of underweight, weight for age (female, male and total % under 5); prevalence of wasting, weight for height (female, male and total % under 5); prevalence of anaemia among children (% under 5); prevalence of anaemia among pregnant women (%); prevalence of anaemia among women of reproductive age (% of women ages 15–49); low-birthweight babies (% of births)
LDP	Deprivation index	Poverty headcount ratio at $1.90 a day (2011 PPP) (% of population); poverty headcount ratio at $3.20 a day (2011 PPP) (% of population); poverty headcount ratio at $5.50 a day (2011 PPP) (% of population); poverty headcount ratio at national poverty lines (% of population)
LAPI	Air pollution index	Mortality rate attributed to household and ambient air pollution, age-standardized (per 100 000 population); PM2.5 air pollution, mean annual exposure (micrograms per cubic meter); PM2.5 air pollution, population exposed to levels exceeding WHO guideline value (% of total); PM2.5 pollution, population exposed to levels exceeding WHO Interim Target-1 value (% of total); PM2.5 pollution, population exposed to levels exceeding WHO Interim Target-2 value (% of total); PM2.5 pollution, population exposed to levels exceeding WHO Interim Target-3 value (% of total)
LMHT	Index of anti-malaria, HIV and TB treatment	Anti-retroviral therapy coverage (% of people living with HIV); children with fever receiving anti-malarial drugs (% of children under age 5 with fever); immunization, DPT (% of children ages 12–23 months); immunization, HepB3 (% of 1-year-old children); immunization, measles (% of children ages 12–23 months)

Note: all the variables used for the PCA indices are from the World Bank’s WDI.

There are some debates around the possible effects of anti-malaria, anti-retroviral and anti-TB treatment in curbing mortality related to COVID-19. I include a PCA of these treatments (LMHT) as a control variable for any possible effect.

is another important factor that can affect the rate of mortality from diseases such as COVID-19, especially as pollution can be associated with upper respiratory disease and other morbidities. I use mean annual exposure to air pollution (LAPEX) and the PCA of the series of WHO related measures of exposure to air pollution and related mortality (LMAP) to generate the index of air pollution (LAPI) as shown in Tables 1 and 2.

Table 1

Individual variables and description

Variables	Description (logs of)	Variables	Description (logs of)
LTI	Total cases per 1000 of population*	LAPEX	Mean annual exposure to air pollution, (μg /m3)
LTD	Total deaths per 1000 infections*	LSAN	% of population using safely managed sanitation services
LED	Excess deaths	LDB	Diabetes prevalence
LTT	Total tests per 1000 of population*	LOB	Age-standardized obesity prevalence (BMI ≥ 30) from WHO
LNPD	National population density	LMAL	Incidence of malaria per 1000
LUPD	Urban population density	LTB	Incidence of tuberculosis per 100 000
LP55	Over-55 population share	LMNCD	Cause of death, by non-communicable diseases
LMA	Median age	LMAP	Mortality due to household and ambient air pollution
LAC	Alcohol consumption per capita	LWSM	Mortality attributed to unsafe water, sanitation and lack of hygiene (per 100 000 population)
LSMOKE	Smoking prevalence for age 15+	LHB	Hospital beds per 1000 of population
LST	Term capturing the stringency of government anti-COVID-19 measures	DAY	Number of days since the first detected case in a country

These variables are sourced from Our World in Data. All other variables are from the World Bank’s WDI. Obesity prevalence is from WHO’s Global Health Observatory.

By mean of PCA, I also generated other indicators that are important in determining the health stock of a society, which are: deprivation (LDP), based on various dimensions of poverty; and malnutrition variable (LMN), which is based on a number of factors that are usually a consequence of malnutrition such as stunting and wasting. The details of variables used in generating the different socio-economic, physical and environmental indices are shown in Table 2.

I fit the variables described earlier into the following models. Two different models are specified. One contains the individual variables (equation 2), while the other is a model of indices generated using PCA as described earlier (equation 3).

Tables 1 and 2 give explanations of the different variables used in the models.

Estimation techniques

There are two main issues with the kind of data used in this work, which require special econometric techniques to deal with them. One is the high number of small values and/or zeros. The second is the strongly non-linear distribution of the mortality observations. Based on this, it is possible that at lower values of mortality, certain variables will behave differently relative to higher values of the dependent variables. On these bases, I choose to estimate the model, first, with Poisson pseudo-maximum-likelihood (PPML), and then by means of quantile regression.

Poisson pseudo-maximum-likelihood (PPML)

The choice of the PPML estimator can be justified on two main grounds as mentioned earlier. Possible heteroscedasticity and accounting for valid zero values in the observations, especially in the earlier periods before the fatalities. The PPML is borrowed from international trade research, which has shown its usefulness in estimating data with a significant number of valid zeros and rapidly (exponentially) rising observations such as with COVID-19 fatalities. The estimator has been shown, in the context of international trade data analyses, to perform well in the presence of the afore-named issues (Santos Silva and Tenreyro 2006; Santos Silva and Tenreyro 2011 ). The associated model for PPML estimation is specified as follows:

The generic model (4) can be applied to equations (2) and (3) for cumulative mortalities.

These estimators are applied to panel data for 196 countries over days of infection from the first recorded case to August 31, 2020. The COVID-19-related data is from Our World in Data, and the socio-economic variables are from the World Bank’s WDI.

Quantile regressions

While the PPML estimates are averages over the distribution, it is possible to exploit the extra information provided by quantile regressions to draw inferences on how the explanatory variables behave at the top of the distribution of the dependent variable relative to the bottom or middle of the distribution. In addition to exploiting the non-linear estimates of the data in order to derive non-biased point estimates at each quantile, the quantile estimator is also able to compare estimates for different quantiles. This is particularly useful in recommending what societal variables become most significant for curbing mortality at catastrophic levels of a pandemic such as COVID-19 when existing health systems become overwhelmed. This quality is important in the context of a rapidly rising rate of COVID-19 mortality within a short span of time. In the current case of cumulative mortalities, the model for linear quantile regression for the qth quantile can be expressed as:

The assumption is that the qth quantile of is zero. The , , , and would be derived from the solution of:

I implement quantitile estimation (equation 6) for 25th, 50th and 75th quantiles.

Key results and discussion

In this section, I interpret, explain and discuss the results of the estimates. I begin with the summary statistics in Table 3 and proceed to the correlation matrix in Table 4. I follow with the report of the results of the PPML point estimates, and finally the quantile regression results.

Table 3

Summary statistics

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Variables	N	mean	sd	min	max	N	mean	sd
	Non-logged					Logged
LTD	34 077	32.79	44.29	0.00	1000.00	27 348	3.24	1.03
LED	34 624	2456.25	950.21	0.00	30113.00	34 617	7.79	0.14
LTI	34 077	1.55	3.43	0.00	42.24	34 077	−1.76	2.88
LTT	12 629	44.20	80.38	0.00	780.13	12 629	2.21	2.40
LST	29 491	63.40	23.93	0.00	100.00	29 182	4.05	0.55
LHB	28 852	3.08	2.52	0.10	13.80	28 852	0.79	0.87
LNPD	34 315	343.78	1541.64	0.14	19196.00	34 315	4.45	1.52
LUPD	28 678	1920.49	2701.80	60.81	20140.81	28 678	6.98	1.06
LP55	34 624	8.01	4.43	0.00	15.74	31 842	2.05	0.50
LMA	31 842	31.03	9.10	15.10	48.20	31 842	3.39	0.31
LAC	30 896	6.43	4.11	0.00	15.20	30 052	1.54	1.04
LSMOKE	34 624	31.03	25.07	0.00	91.90	25 060	3.63	0.58
LAPEX	31 996	87.34	28.48	0.00	100.00	30 871	4.38	0.76
LSAN	34 299	78.19	27.75	7.32	100.00	34 299	4.24	0.57
LHIV	24 499	1.70	4.08	0.10	27.30	24 499	−0.76	1.44
LMAL	34 624	40.25	98.88	0.00	505.57	14 052	2.33	3.06
LTB	34 624	93.49	130.87	0.00	611.00	33 124	3.58	1.57
LDB	33 332	8.05	4.19	0.99	23.36	33 332	1.95	0.55
LOB	9 049	12.77	9.59	0.10	61.00	9 049	2.17	1.01
LMNCD	30 556	18.29	5.60	7.80	30.60	30 556	2.86	0.33
LMAP	30 556	87.39	71.77	7.00	324.10	30 556	4.06	0.99
LWSM	30 556	11.85	20.44	0.10	101.00	30 556	0.41	2.29
LPD	23 672	1.31	1.50	0.00	10.36	23 507	−0.24	1.07
LAPI	11 269	14.31	2.57	0.00	15.87	30 346	0.87	1.23
LSANI	25 253	3.36	2.38	0.00	12.77	11 121	2.66	0.20
LHI	23 924	3.96	1.08	0.00	6.21	25 087	0.93	0.83
LMHT	23 715	1.32	2.20	0.00	13.12	23 745	1.34	0.34
LCCM	30 556	2.36	1.14	0.00	5.59	23 505	−0.71	1.52
LNCM	11 073	3.51	2.36	0.00	11.61	30 397	0.63	0.89
LMN	17 158	2.41	2.12	0.00	7.87	10 854	1.04	0.75
LDP	30 556	3.40	1.96	0.00	7.55	16 943	0.42	1.07
DAY	34 624	88.68	52.32	1.00	240.00	34 624	88.68	52.32

Table 4

Correlation coefficients

	LTD	LED	LTI	LTT	LST	LHB
LED	0.2456*	1
LTI	0.1658*	0.1271*	1
LTT	−0.0957*	−0.007	0.8724*	1
LST	−0.0483*	0.1289*	0.3011*	0.0437*	1
LHB	−0.0267*	−0.0100*	0.2596*	0.3417*	−0.1366*	1
LNPD	0.0520*	0.0017	0.1160*	−0.0345*	0.0075	0.0666*
LUPD	0.1579*	0.0414*	−0.2101*	−0.3650*	0.1000*	−0.4283*
LP55	0.1149*	0.0555*	0.2781*	0.3065*	−0.1615*	0.7318*
LMA	0.0647*	0.0585*	0.3199*	0.4075*	−0.1586*	0.7456*
LAC	0.0905*	0.0443*	0.1279*	0.2223*	−0.1258*	0.4937*
LSMOKE	0.0735*	0.0248*	0.0615*	0.2009*	−0.1486*	0.5580*
LAPEX	0.0697*	−0.1065*	−0.1598*	−0.1980*	0.0914*	−0.1151*
LSAN	0.0083	0.0387*	0.3191*	0.3513*	−0.0800*	0.6490*
LHIV	−0.1020*	−0.0161*	−0.1531*	−0.1871*	0.1154*	−0.2565*
LMAL	0.0106	−0.0723*	−0.1857*	−0.1638*	−0.0423*	−0.4301*
LTB	−0.1000*	−0.0664*	−0.3287*	−0.4639*	0.1075*	−0.4163*
LDB	0.0788*	0.0021	0.1381*	0.1127*	0.0487*	0.1601*
LOB	0.0198	0.0896*	−0.2169*	−0.1621*	−0.1634*	−0.1942*
LMNCD	0.0071	0.0779*	−0.2503*	−0.3042*	0.1244*	−0.3004*
LMAP	0.0851*	0.1001*	−0.3233*	−0.4224*	0.1589*	−0.5539*
LWSM	0.0297*	−0.0571*	−0.3707*	−0.5299*	0.1263*	−0.7041*
LPD	0.0526*	0.0440*	−0.0430*	−0.1624*	−0.0013	−0.004
LAPI	0.0577*	0.0894*	−0.1880*	−0.2781*	0.1827*	−0.2994*
LSANI	0.0391*	0.0432*	0.2050*	0.3699*	−0.1081*	0.4377*
LHI	−0.0021	0.0484*	0.3132*	0.4460*	−0.1227*	0.7305*
LMHT	−0.0929*	−0.0886*	−0.1314*	−0.2408*	0.0643*	−0.3380*
LCCM	−0.1052*	−0.0308*	−0.2986*	−0.3808*	0.1399*	−0.5384*
LNCM	0.0284*	0.0316*	0.2300*	0.2777*	−0.0633*	0.4657*
LMN	0.0024	0.0921*	−0.3597*	−0.3167*	0.0198*	−0.6866*
LDP	−0.0098	−0.0225*	−0.2589*	−0.3401*	−0.0066	−0.6860*
Day	−0.1042*	0.0107*	0.6127*	0.6518*	0.0682*	0.0468*

The summary statistics in Table 3 show that on average, there are 32.8 deaths per thousand recorded infections in the sample. Reported mortality strongly depends on recorded infections, which in turn are conditioned by the testing rate. Also, the more testing that is done, the earlier the cases without symptoms may be discovered, suggesting that while infected cases may imply higher mortality rates, testing may be negatively related to mortality. This necessitates controlling for both infections and testing in the mortality models. On average, 1.55 persons per thousand of the populations of our sample countries are infected by the COVID-19-causing virus, with significant dispersions across countries. There are on average 44.2 persons per thousand of population tested, with some countries achieving as low as zero persons tested and others as high as 128 persons tested on average. The fact that countries that test more tend to have higher reported cases means that countries with low levels of testing have systematic under-reporting of infected cases. This explains why I control for testing in the infections model. The questionable nature of the reported mortality requires comparing it with excess deaths. Due to negative values in excess deaths, I rescaled it so that the lowest value is zero. The average excess deaths for the sample is 2456.25.

Although a number of the variables show the expected correlation signs (Table 4), given the non-linear nature of the key variables, not much useful information would be expected from the correlation coefficients in Table 4. Reliable results can only be inferred from appropriate econometric models that effectively control for statistical issues present in the data.

Econometric results

The econometric results are reported in Tables 5 to 8. The first sets of results in Tables 5 and 6 are point estimates for mortality, using the PPML estimator. The second sets of results are quantile estimates for total deaths, reported in Tables 7 and 8 for individual variables and indices-based regressors, respectively. Each regression table compares reported COVID-19 deaths and excess deaths.

Table 5

Average estimates for cumulative mortality (Model of individual variables)

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	COVID deaths	COVID deaths	COVID deaths	Excess deaths	Excess deaths	Excess deaths
LTI	0.286***	0.378***	0.236***	0.081***	0.087***	0.008***
	(0.009)	(0.009)	(0.013)	(0.007)	(0.008)	(0.001)
LTT	−0.298***	−0.432***	−0.417***	−0.079***	−0.107***	−0.020***
	(0.011)	(0.011)	(0.022)	(0.007)	(0.009)	(0.003)
LST	0.184***	0.274***	0.718***	0.006**	0.062***	0.036***
	(0.026)	(0.033)	(0.084)	(0.003)	(0.008)	(0.005)
LHB	0.002	−0.291***		−0.085***	−0.078***
	(0.017)	(0.021)		(0.008)	(0.009)
LSAN	−0.116***	−0.730***		−0.029***	−0.057***
	(0.030)	(0.033)		(0.004)	(0.009)
LAC	0.280***	0.129***		0.084***	0.034***
	(0.012)	(0.015)		(0.008)	(0.005)
LSMOKE	0.340***	0.327***		0.092***	0.080***
	(0.018)	(0.019)		(0.009)	(0.007)
LAPEX		0.036***			0.064***
		(0.008)			(0.009)
LNPD		0.253***			0.055***
		(0.011)			(0.005)
LUPD		0.359***			0.138***
		(0.025)			(0.016)
LP55		0.812***			0.171***
		(0.073)			(0.028)
LMA		0.767***			0.154***
		(0.112)			(0.052)
LDB			0.496***			0.051***
			(0.054)			(0.007)
LMAL			0.072***			0.002***
			(0.009)			(0.000)
LTB			−0.054**			−0.020***
			(0.121)			(0.003)
LHIV			−0.025**			0.007***
			(0.012)			(0.001)
LMAP			0.229***			0.004**
			(0.032)			(0.001)
Day	0.005***	0.006***	0.005***	−0.000***	0.000**	0.000***
	(0.000)	(0.000)	(0.001)	(0.000)	(0.000)	(0.000)
Constant	1.887***	1.678***	0.683	7.577***	8.991***	7.627***
	(0.172)	(0.402)	(0.424)	(0.024)	(0.236)	(0.022)
Observations	10 702	8 596	3 837	10 702	8 596	3 837
R-squared	0.154	0.397	0.327	0.040	0.087	0.072

Robust standard errors in parentheses.

P < 0.01.

P < 0.05.

P < 0.1.

Table 6

Average estimates for cumulative mortality (Model of indices)

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Variables	COVID deaths	COVID deaths	COVID deaths	COVID deaths	Excess deaths	Excess deaths	Excess deaths	Excess deaths
LTI	0.286***	0.264***	0.290***	0.267***	0.142***	0.026***	0.156***	0.004***
	(0.011)	(0.010)	(0.011)	(0.014)	(0.013)	(0.003)	(0.014)	(0.001)
LTT	−0.152***	−0.143***	−0.142***	−0.299***	−0.165***	−0.026***	−0.183***	−0.009***
	(0.013)	(0.014)	(0.012)	(0.016)	(0.014)	(0.003)	(0.016)	(0.001)
LST	−0.008	0.100**	0.249***	−0.045	−0.011	−0.000	0.000	0.009***
	(0.036)	(0.047)	(0.032)	(0.043)	(0.009)	(0.002)	(0.006)	(0.002)
LHB	−0.063***	−0.405***	−0.242***	0.271***	−0.078***	−0.029***	−0.139***	−0.014***
	(0.020)	(0.028)	(0.032)	(0.026)	(0.008)	(0.004)	(0.015)	(0.002)
LSANI	−1.532***	−1.635***	−1.648***		−1.549***	−1.076***	−1.035***
	(0.086)	(0.114)	(0.123)		(0.134)	(0.019)	(0.094)
LPD	0.289***	0.444***			0.105***
	(0.012)	(0.015)			(0.008)
LCCM		0.378***				0.008***
		(0.013)				(0.001)
LNCM		0.201***				0.177***
		(0.055)				(0.022)
LMHT		−0.126***				−0.009***
		(0.027)				(0.003)
LSUB			0.190***				0.164***
			(0.030)				(0.016)
LAPI			0.111***				0.032***
			(0.008)				(0.006)
LMN				0.074**				0.055***
				(0.037)				(0.007)
LDP				0.328***				0.018***
				(0.026)				(0.002)
Constant	0.285	1.853***	−1.235***	4.303***	4.395***	7.498***	5.600***	7.828***
	(0.272)	(0.378)	(0.358)	(0.206)	(0.296)	(0.050)	(0.211)	(0.013)
Observations	5,588	3,585	6,214	2,778	5,588	4,207	6,214	2,778
R-squared	0.305	0.506	0.179	0.263	0.115	0.100	0.105	0.062

Robust standard errors in parentheses.

P < 0.01.

P < 0.05.

P < 0.1.

Table 7

Quantile estimates for cumulative deaths (individual variables)

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Variables	q25	q50	q75	q25	q50	q75	q25	q50	q75
LTI	0.030*	0.172***	0.243***	0.089***	0.136***	0.220***	0.019	0.133***	0.202***
	(0.016)	(0.009)	(0.011)	(0.012)	(0.014)	(0.009)	(0.017)	(0.008)	(0.018)
LTT	−0.225***	−0.419***	−0.480***	−0.236***	−0.333***	−0.622***	−0.189***	−0.314***	−0.439***
	(0.022)	(0.013)	(0.011)	(0.045)	(0.046)	(0.029)	(0.018)	(0.016)	(0.016)
LST	0.385***	0.072	−0.050**	0.917***	0.334**	0.282***	0.246***	0.034	0.018
	(0.040)	(0.054)	(0.025)	(0.059)	(0.140)	(0.075)	(0.083)	(0.030)	(0.031)
LHB	−0.159***	−0.468**	−0.458
	(0.041)	(0.095)	(0.326)
LNPD	0.243***	0.247***	0.291***	0.200***	0.033*	0.008	0.185***	0.175***	0.301***
	(0.009)	(0.015)	(0.016)	(0.025)	(0.017)	(0.023)	(0.019)	(0.010)	(0.013)
LUPD	0.659***	0.545***	0.492***	0.235***	0.083	0.211***	0.675***	0.502***	0.612***
	(0.049)	(0.029)	(0.050)	(0.038)	(0.070)	(0.053)	(0.052)	(0.039)	(0.049)
LP55	1.322***	1.123***	0.154	1.266***	1.270***	0.399	1.427***	0.990***	0.676***
	(0.201)	(0.088)	(0.171)	(0.221)	(0.396)	(0.295)	(0.163)	(0.060)	(0.068)
LMA	1.224***	0.445**	1.243***	0.935***	0.878	0.333	1.367***	0.114	0.425*
	(0.361)	(0.202)	(0.229)	(0.330)	(0.600)	(0.613)	(0.299)	(0.123)	(0.235)
LAC	0.033**	0.144***	0.237***
	(0.015)	(0.008)	(0.018)
LSMOKE	0.154***	0.016	0.185***
	(0.053)	(0.016)	(0.028)
LAPEX	0.005	0.037***	−0.014
	(0.019)	(0.008)	(0.017)
LMAL				0.077***	0.076***	0.065***
				(0.025)	(0.016)	(0.014)
LTB				0.082**	0.064**	0.211***
				(0.034)	(0.033)	(0.018)
LDB				0.447***	0.191***	−0.062
				(0.059)	(0.046)	(0.073)
LMNCD							0.011	0.204***	0.355***
							(0.038)	(0.043)	(0.096)
LWSM							0.121***	0.139***	0.121***
							(0.020)	(0.018)	(0.022)
Day	0.007***	0.007***	0.006***	0.006***	0.004***	0.008***	0.006***	0.005***	0.007***
	(0.000)	(0.000)	(0.000)	(0.001)	(0.001)	(0.001)	(0.001)	(0.001)	(0.000)
Constant	5.297***	1.991***	1.473*	−3.007***	2.976**	5.669***	7.089***	4.181***	5.095***
	(1.036)	(0.675)	(0.759)	(0.761)	(1.243)	(1.639)	(0.992)	(0.447)	(0.777)
Observations	8,048	8,048	8,048	3,462	3,462	3,462	9,393	9,393	9,393

Standard errors in parentheses.

P < 0.01.

P < 0.05.

P < 0.1.

Table 8

Quantile estimates for cumulative deaths (individual variables)

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	q25	q50	q75	q25	q50	q75
LTI	0.111***	0.142***	0.177***	0.224***	0.180***	0.283***
	(0.012)	(0.010)	(0.014)	(0.032)	(0.020)	(0.005)
LTT	−0.340***	−0.334***	−0.423***	−0.399***	−0.321***	−0.411***
	(0.019)	(0.012)	(0.013)	(0.045)	(0.017)	(0.015)
LST	0.220*	−0.130***	−0.067***	−0.901***	−0.597***	−0.631***
	(0.113)	(0.034)	(0.024)	(0.135)	(0.044)	(0.042)
LHB	−0.168***	−0.033***	−0.122***	−0.368***	−0.061	−0.147***
	(0.041)	(0.012)	(0.022)	(0.069)	(0.044)	(0.029)
LCCM	0.058***	0.130***	0.247***
	(0.010)	(0.009)	(0.010)
LNCM	0.114***	0.030**	0.061**
	(0.016)	(0.012)	(0.025)
LMHT	−0.940***	−0.700***	−0.511***
	(0.037)	(0.024)	(0.021)
LMN				0.368***	0.270***	0.218***
				(0.136)	(0.063)	(0.064)
LDP				0.325***	0.300***	0.329***
				(0.045)	(0.035)	(0.037)
Day	0.007***	0.002***	0.005***	−0.000	0.000	−0.002***
	(0.001)	(0.000)	(0.000)	(0.002)	(0.001)	(0.000)
Constant	2.844***	5.125***	4.882***	7.591***	6.872***	7.720***
	(0.540)	(0.217)	(0.144)	(0.718)	(0.225)	(0.196)
Observations	7,066	7,066	7,066	2,562	2,562	2,562

Standard errors in parentheses.

P < 0.01.

P < 0.05.

P < 0.1.

PPML point estimates

The PPML estimates for cumulative deaths are reported in Tables 5 and 6. These estimates show relatively acceptable goodness of fit as suggested by the R-squares for reported COVID deaths. The reason that these are reliable is because I am dealing with a large sample of countries (196) with relatively smaller time dimensions (about four months of daily observations) for COVID-19 variables, and other variables that are not essentially varying with time, since they are the latest annual averages for each country. However, the values are very low for the excess deaths models. This is likely due to the fact that fewer countries reported excess deaths, and therefore the model estimates only a sub-sample of the data.

The sub-models of individual variables estimates in columns 1 to 6 of Table 5 are mainly a combination of variables in a way as to minimize the effect of collinearity among the variables. Columns 1 to 3 are results for reported deaths, while columns4 to 6 are for excess deaths. It is worth mentioning that the coefficients from the models with reported deaths and excess deaths are generally similar in terms of signs. The magnitudes of the coefficients of the excess deaths models are generally lower than those of the reported COVID deaths. This is reassuring as it suggests that the interpretation of the results, especially based on signs, is robust to errors in COVID death reporting.

The coefficients of the indices generated using the PCA method are in Table 6. The variables are all also in log form and therefore the estimated coefficients can be interpreted as elasticities. The columns are aligned as mentioned earlier. The sub-models in the different columns (1 to 4) for reported deaths and (5 to 8) for excess deaths are PPML estimates for, first of the basic parsimonious model, which consists only of infections (LTI), tests (LTT) and health infrastructure index (LHB), stringency of government measures (LST) and sanitation index (LSANI). The results of the individual variables (Table 5) and indices-based models (Table 6) are interpreted together for complementarity in order to draw useful inferences.

Testing and infections. In the different models, I controlled for both testing rate and infections. While higher rates of infections lead to higher mortality, testing reduces mortality. A 1% increase in cumulative infections results in a 0.38% (Table 5 column 2) increase in reported mortality and a 0.09% (column 5) increase in excess deaths. This is in line with expectation, which is that the higher the infections, the higher the mortality. However, a 1% increase in testing results in 0.43% (0.11%) less reported (excess) mortality per thousands of infections. This result is insightful and not surprising, since higher levels of testing may lead to early detections and early detection carries hopes of recovery. Countries that test more are more likely to save lives from COVID-19 deaths.

Population dynamics. All three variables of population dynamics—national and urban population densities (LNPD and LUPD), over-55s population share and median age (LMA)—matter for COVID-19 mortality. A 1% increase in national population density and urban population density, respectively, leads to 0.25% and 0.36% (0.06% and 0.14%) increases in COVID-19 fatalities (excess deaths). The magnitudes of the age-related demographic variables are even bigger in the COVID-19 model. The results also agree largely with the stylized facts that have been observed, i.e. that the elderly are more vulnerable to dying of COVID-19. A 1% increase in the median age (LMA), over-55s share of the population (LP55) results in 0.77% and 0.81% (0.15% and 0.17%) increases in reported COVID-19 deaths (excess deaths). The elasticity of population dynamics index is 0.44% (column 4 of Table 6) for reported COVID deaths and 0.11% for excess deaths. It is worth noting that the index contains more variables, including various internally displaced groups. The implication is that countries with big and densely populated cities are more at risk of mortality from COVID-19.

Alcohol consumption and smoking prevalence. Both alcohol consumptions per capita (LAC) and the prevalence of cigarette smoking (LSMOKE) significantly contribute to COVID-19 deaths. A percentage increase in alcohol consumption per capita brings about a 0.13% rise in fatalities. The same increase in prevalence of cigarette smoking results in a 0.33% rise in fatalities. The elasticities are 0.03% and 0.08%, respectively, in the excess deaths models. The PCA index of both variables (LSUB) has an elasticity of 0.19% and 0.16% for reported and excess deaths, respectively. It is quite plausible that both alcohol and cigarette consumption contribute to raising mortality through their role in reducing human health. Both alcohol and cigarettes have been associated with various non-communicable diseases such as hypertension, cardiovascular, diabetes for alcohol (GBD 2018; Burton and Sheron 2018), and upper respiratory track diseases and lung cancer for smoking (Pryor 1997; Slattery ; Zhang ). These then serve as underlying conditions that increase mortality due to COVID-19.

Health infrastructure. I use the log of the number of hospital beds per thousand of population (LHB) as proxy for health infrastructure. This variable is highly correlated with numbers of physicians and nurses per thousand of population and also correlates strongly with the PCA of health variables. A 1% increase in beds per thousand of population results in a 0.29% (0.08%) reduction in reported (excess) mortality. The coefficients are relatively higher in the indices-based models in Table 6. It can be inferred that countries with good health infrastructure are relatively better at attenuating mortalities due to the pandemic.

Comorbidities. Of all the individual comorbidities, diabetes prevalence shows the highest positive and significant effect in raising mortality. A 1% increase in diabetes prevalence results in a 0.5% (0.05%) rise in reported (excess deaths). I undertake a more detailed analysis of diabetes together with the related obesity measures later. The others (malaria, TB and HIV) have opposing signs and I suspect it is because these variables are too related and are also associated with certain treatments that may have some attenuating effects on COVID fatality. There are some trials involving TB treatment on COVID-19 (WHO 2020c) and there have been some debates around the effectiveness of hydroxychloroquine and anti-malarial medication in COVID-19 treatment (Ferner and Aronson 2020). In Table 6, the coefficients of the index of anti-malaria, -HIV, -TB treatments (LMHT) appear to support this view, where a 1% increase in this index results in a 0.13% (0.01%) reduction in reported (excess) deaths. In Table 6 where these treatments are controlled for, both communicable and non-communicable comorbidities significantly raise fatalities. I generated two indices, one for communicable (LCCM) and the other for non-communicable diseases (LNCM). A percentage increase in communicable comorbidities raises COVID-19 and excess fatalities by 0.38% and 0.008%, respectively. The same increase in the non-communicable comorbidity index increases the mortalities by 0.20% and 0.18%, respectively. These suggest that both communicable and non-communicable disease burdens weaken health stock and lead to deaths from pandemics like COVID-19. However, this is only visible after controlling for the treatment effects as carried out in this work.

Malnutrition and deprivation. These are also key determinants of mortality. A percentage increase in malnutrition raises both reported COVID and excess deaths by 0.07% and 0.06%, respectively. The same increase in the deprivations index results in a 0.33% (0.02%) increase in reported (excess) mortality. The suggestion is that ensuring good nutrition and reducing deprivation is a way of building the resilience of a population against mortalities due to pandemics like COVID-19.

Air pollution. This has some effects in increasing mortality. A 1% increase in exposure to air pollution (LAPEX) raises COVID reported and excess deaths by 0.04% and 0.06%, respectively. The index of air pollutants (LAPI), which includes other dimensions, has similar elasticities, with a 1% increase resulting in 0.11% and 0.03% higher reported COVID and excess deaths, respectively. As in the case of cigarettes, this effect may equally be due to the impact of air pollution on comorbidities such as upper respiratory diseases.

Sanitation. Sanitation also shows a significant effect in reducing cumulative deaths. A 1% increase in the population using managed basic sanitation (LSAN) results in 0.73% and 0.06% reductions in reported COVID and excess deaths, respectively. The log of sanitation index (LSANI), which includes other sanitation-related variables (see Table 2), has significantly higher effects, possibly due to other variables included in the index that may also significantly impact mortality. A 1% increase in the index leads to 1.64% and 1.08% reductions in both excess and reported COVID mortalities, respectively. I believe that sanitation works through rates of infection and communicable comorbidities. In models where I introduce sanitation (not reported), the negative elasticity of testing is significantly reinforced. Although I have not specifically tested for the interaction of sanitation with infections and testing, due to the behaviour of the other coefficients in the presence of sanitation, I conjecture that high testing in the presence of good sanitation leads to a reduction in the discovery of cases and lower mortality. Good sanitation can also ward off other communicable comorbidities, leading to fewer deaths.

There is much talk about the effect of anti-malarial, anti-TB and anti-retroviral medication on COVID-19. The index that captures these (LMHT) seems to confirm this hypothesis. A 1% increase in the index reduces both excess and reported COVID mortalities by 0.009% and 0.13%, respectively. However, I think that this effect works through lower communicable comorbidity due to treatment, and better health infrastructure. Good health infrastructure implies ready availability of such treatments and effective coverage of the affected population with the treatment. This may indeed be the case as the introduction of the LMNT index reduces the negative effect of health infrastructure in our model.

Finally, it is worth noting that the measures taken by most governments appear to be at best insignificant where the coefficient is negative, or significantly positive. This means that it is likely that the stringent control measures taken by most governments may have exacerbated fatalities in some cases.

While these point estimates are quite informative in recommending policy measures for curbing fatalities, I also take interest in assessing how these factors behave at catastrophic levels of the pandemic, when the health systems become overwhelmed. For this purpose, I estimated quantile models.

Quantile estimates

The quantile regression modelling divides the distribution of cumulative deaths into three quantiles (q25, q50, q75). Table 7 shows the results for the individual variable models. Due to collinearity, I did not introduce all the individual variables in a single model. I had to estimate three separate models to contain all the relevant variables and minimize bias due to collinearity. Also due to the instability of the individual variable models, judging from the low (pseudo) R-squares, most of our inference comes from the index-based model in Table 7 where the goodness of fit is relatively higher, above 0.5.

Both cumulative infections and cumulative tests have stable elasticities across all the quantiles. The effect of total infections remains positive with elasticities of about 0.03% in q25, 0.17% in q50, and 0.24% in q75. This suggests that infections increase mortality more at the later stages of evolution of deaths. That is higher infections at a late stage are likely to lead to higher mortality. This suggests that countries that manage to flatten the infection curves quicker also benefit more in terms of lives saved. Testing has a negative effect on mortality, which increases in magnitude from the lowest to the topmost quantiles. This is simply showing the fact that the more tests increase with infections and fatalities, the more countries may manage to introduce measures that curb the mortalities.

Good health infrastructure can help in reducing mortality. If infrastructure in terms of number of beds per thousand population increases with cases, it will do well to curb mortalities through better care. A 1% increase in the health infrastructure index results in 0.16% and 0.47% reductions in deaths at the 25th and 50th percentile of progression. At the 75th the effect is large, but not significant. Therefore, while health infrastructure carries some potential later on for curbing mortality, it quickly becomes overwhelmed and turns insignificant. This implies that from the 75th percentile of infections and mortality, the public health infrastructure becomes inadequate in handling the pandemic. This effect is compounded by certain risky behavioural factors of the population such as smoking and alcohol prevalence, showing increasing effects in enhancing mortalities across the quantiles. At this stage, it will be a combination of other socio-economic measures that determine a population’s stock of health that would matter more for resilience in the face of mortality due to the pandemic. In searching for such factors, I combine the individual variables results in Table 7 and the results of indices in Table 8.

Throughout the progression of the pandemic, both communicable and non-communicable disease burdens are significant determinants of mortality. While the elasticities of communicable diseases rise across the quantiles, those of non-communicable diseases, though positive, fall from bottom to top quantiles. A 1% increase in communicable comorbidity leads to a 0.06% increase in COVID-19 deaths in q25, 0.13% increase in q50, and 0.25% increase in q75. The elasticities of non-communicable diseases are 0.11% in q25, 0.03% in q50, and 0.06% in q75. The progression of mortality in different countries will therefore be largely determined by their stock of health in terms of both communicable and non-communicable diseases. Countries that take significant measures towards effective treatment of these diseases will gain a great deal in times of such pandemics. This is supported by the mortality curbing effects of anti-malaria, -HIV, and -TB. The treatment index shows a significant negative impact on mortalities with elasticities of 0.94%, 0.70%, and 0.51% for quantiles 25, 50, and 75, respectively.

Another significant factor that relates to stock of a population’s health is nutrition. A percentage increase in the index of malnutrition (LMN) leads to a 0.37% increase in mortality in q25, 0.27% increase in q50, and 0.22% increase in q75. The nutritional health endowment of a population therefore plays a significant role in whether or not there are explosive deaths due to the pandemic, or conversely, recovery rates. Related to malnutrition is deprivation (LDP), and a 1% increase in the index results in a 0.33% increase in mortality for q25, 0.30% in q50, and 0.33% in q75.

The measures taken by most governments appear to show effects in curbing fatalities only in the upper quantiles. This seems to correspond to the segment of the distribution when most of the measures have either been relaxed or fine-tuned. The lesson here is that governments should not see draconian measures as the first order solution to curbing a pandemic. Time is usually required for governments to learn and fine-tune measures to suit the society. Hence governments with the ability to quickly learn and adapt will be able to save their populations much pain while efficiently controlling fatalities.

Highlighting the role of obesity and diabetes

The COVID-19 outbreak is happening in an era of very high obesity prevalence. Popkin have highlighted that all countries in the world today have a prevalence of individuals with overweight/obesity >20%. Obesity is associated with metabolic and inflammatory conditions that play a significant role in the susceptibility to severe lung disease associated with acute respiratory syndrome (ARS), which is the main cause of COVID-19 mortality. Consequently, a strong association has been shown between overweight/obesity and risk of COVID-19 hospitalization (Lighter ). The relationship between obesity and other risk factors such as diabetes has also been highlighted in the literature (GBD 2017). These facts necessitate a separate robust analysis that highlights the roles of diabetes prevalence (LDB) and obesity prevalence (LOB) in separate models.

In Table 9a, the coefficients of diabetes and obesity are presented in columns 1 (obesity only), column 2 (diabetes only), column 3 (both obesity and diabetes) for PPML estimates, and columns 4 to 6 for various quantiles of COVID-19 mortalities in quantile regression. Table 9b presents similar results, but controlling for hospital infrastructure using the number of hospital beds per thousand population. Both obesity and diabetes prevalence show significant effects in raising mortality. In both tables, the first panel of results shows estimates for reported COVID-19 deaths, and those in the second panel are estimates for excess deaths. Without controlling for health infrastructure, both diabetes and obesity prevalence significantly enhance COVID-19 mortality. Diabetes coefficients are generally higher than those of obesity in the models of reported COVID-19 deaths, but vice versa in the models of excess deaths. However, after controlling for health infrastructure, obesity becomes insignificant in both reported COVID-19 deaths and excess deaths models, and for both point and the quantile estimates. This suggests that the negative effect of obesity is attenuated by good health infrastructure, but the same good health infrastructure may not equally reduce the significance of COVID-19 mortality due to diabetes.

Table 9a

The role of obesity and diabetes

Variables	(1)	(2)	(3)	(4)	(5)	(6)
	PPML			Quantile regression
				q25	q50	q75
	Reported COVID-19 deaths
LTI	0.872***	0.867***	0.776***	0.875***	0.561***	0.663***
	(0.047)	(0.047)	(0.050)	(0.108)	(0.129)	(0.076)
LTT	−0.554***	−0.603***	−0.522***	−0.292**	−0.308***	−0.512***
	(0.040)	(0.040)	(0.044)	(0.128)	(0.118)	(0.072)
LST	0.598***	0.663***	0.841***	−0.649**	0.159	0.372
	(0.174)	(0.172)	(0.182)	(0.274)	(0.429)	(0.245)
LOB	0.185***		0.337***	0.167***	0.348***	0.486***
	(0.049)		(0.046)	(0.046)	(0.095)	(0.117)
LDB		0.394***	0.507***	0.023	0.746***	0.381***
		(0.049)	(0.054)	(0.151)	(0.153)	(0.112)
Constant	2.137***	3.218***	1.598*	6.159***	4.289**	3.351***
	(0.813)	(0.777)	(0.846)	(1.215)	(1.892)	(1.159)
Observations	947	947	947	832	832	832
R-squared	0.522	0.592	0.616

Robust standard errors in parentheses.

P < 0.01.

P < 0.05.

P < 0.1.

Table 9b

The role of obesity and diabetes in the presence of health infrastructure

Variables	(1)	(2)	(3)	(4)	(5)	(6)
	PPML			Quantile regression
				q25	q50	q75
	Reported COVID-19 deaths
LTI	0.880***	0.780***	0.749***	1.021***	0.680***	0.573***
	(0.043)	(0.038)	(0.040)	(0.050)	(0.062)	(0.059)
LTT	−0.475***	−0.441***	−0.417***	−0.483***	−0.328***	−0.410***
	(0.035)	(0.032)	(0.033)	(0.042)	(0.030)	(0.040)
LST	0.488**	0.762***	0.829***	−0.276	0.178	0.158
	(0.189)	(0.202)	(0.206)	(0.227)	(0.231)	(0.337)
LHB	−0.609***	−0.945***	−0.907***	−1.871***	−1.427***	−0.721***
	(0.059)	(0.049)	(0.052)	(0.070)	(0.110)	(0.083)
LOB	0.002		0.129***	−0.019	−0.005	0.089
	(0.051)		(0.041)	(0.027)	(0.058)	(0.076)
LDB		0.799***	0.818***	0.975***	0.941***	0.766***
		(0.033)	(0.035)	(0.055)	(0.068)	(0.065)
Constant	3.615***	4.303***	3.611***	9.768***	7.222***	6.576***
	(0.882)	(0.894)	(0.936)	(0.990)	(1.038)	(1.388)
Observations	947	947	947	832	832	832
R-squared	0.562	0.723	0.723

Robust standard errors in parentheses.

 < 0.01.

 < 0.05.

 < 0.1.

It is therefore important to deduce that controlling obesity in normal times will limit strain on the public health systems of countries during times of pandemics such as COVID-19. However, with or without a good health infrastructure, diabetes is likely to lead to an exacerbation of fatalities due to such pandemics. It flows from this that nations should pay attention primarily to curbing diabetes while dealing with lifestyle factors that lead to obesity.

Figure 2 shows the plots of the estimated coefficients of the quantile regressions relative to the Ordinary Least Squares (OLS) average coefficients and their respective confidence interval bands for the index-based mortality model. The plots for all the other variables used in the model are in Figure A1 of the Appendix. The figures justify the use of quantile regressions and also validate the results.

Figure 2

Comparative coefficients of the cumulative deaths models OLS versus quantile.

Socio-economic policies to complement health systems

From the above analyses, a number of findings have come to light. These findings are useful for the design of socio-economic fabrics that can complement health systems in curbing mortality as this pandemic progresses, and in similar future pandemics.

Enhancing capacity for early testing. The first finding is to always keep a well-prepared public health capability for effective testing, early isolation of cases and treatment and contact-tracing. Testing of cases will prevent rapid spread of the disease and reduce mortality through three channels. One being early detection before cases get critically morbid. The second is that early testing and isolation reduce the rate of spread of the pandemic, giving breathing room to the health infrastructure to deal with arising cases. Finally, early testing allows for early identification, isolation of cases and contacts, and prioritization of high-risk cases (Hellewell et al. 2020), i.e. those with significant comorbidities so that the health system can handle before being overwhelmed.

Heath infrastructure. Of course, the first and most important aspect of preparedness that may ensure less fatality is adequate health infrastructure. Higher levels of good health infrastructure have proven to be important in curbing mortality. However, the effect of heath infrastructure, no matter how good, cannot last the span of a pandemic. As the pandemic rises to catastrophic levels, health systems quickly become overwhelmed and ineffective in managing cases and reducing mortality. The effectiveness of health infrastructure is complemented by other non-biomedical factors that can keep the curve of the spread of the pandemic low. This combination of health infrastructure and other (social, economic, physical and environmental) factors can prove very effective in reducing mortality and managing the pandemic in a way that leads to less human cost.

Building the health stock of the population. One of the most important determinants of an explosive rise in COVID-19 mortality arising from the findings of this work is how healthy a population is. The health stock of a population can be captured by the indices of existing comorbidities. Highly morbid societies are more vulnerable to fatalities due to pandemics like COVID-19. Communicable and non-communicable disease burdens are issues that all countries should urgently and effectively be dealing with in normal times in order to prepare the population to minimize the worse effects of pandemics. Societies with higher non-communicable disease burden like obesity and diabetes are at higher risk of mortality; however, curbing diabetes should be an urgent priority for all countries. Promotion of healthy lifestyles and treatment and reversal, where possible, of such disease is paramount for preparedness of a population. Our findings also suggest that the effective treatment of existing communicable disease can reduce the risk of high mortality from the pandemic. This is because the index of anti-malaria, anti-tuberculosis and anti-HIV treatments shows negative impact on mortality due to the pandemic. We think that this impact comes through the reduction of the effects of communicable comorbidities and the strengthening of the effect of public health infrastructure. The availability of these treatments and effective coverage of the affected population also depends on the quality of health infrastructure.

Improving sanitation and hygiene. Our findings reveal that significant adoption of good sanitary behaviour of the population significantly curbs mortality. Societies with good hygiene and sanitation, including access to good clean water will be less exposed not only to the disease, but also to other communicable comorbidities and consequently, less likely to register higher death tolls.

Improving nutrition. Societies with high prevalence of malnutrition would register significantly higher death tolls due to this pandemic, and also possibly future similar pandemics. Malnutrition can significantly reduce the health stock of a population and leave it less able to naturally cope with a pandemic. In such cases, mortality due to pandemic will be higher. At normal times, authorities of countries and international efforts should attend to adequate nutritional needs of citizens.

Reducing household and ambient air pollution. Over the years, due to increasing use of dirty fuels and pollutants, both household and ambient air quality has deteriorated significantly. Most of these pollutants are largely due to the negative externalities of industrialization and economic development. Careful management of these externalities in order to reduce emissions of pollutants and enhance air quality will have a significant effect in curbing mortality due to respiratory-related pandemics. The effect likely passes through other respiratory comorbidities, just like smoking.

Reducing exposure to alcohol and cigarettes. A related factor to population health is risky behavioural elements. This pandemic has raised significant debates in many societies about whether or not alcohol and cigarettes should be banned during the pandemic as a curbing measure. In South Africa for example, major tobacco companies threatened to take the government to court for extending the ban on cigarettes. A similar dissatisfaction has been raised regarding the banning of alcohol. While that debate may be valid in the context of the economic contributions of these industries, a careful cost–benefit analysis needs always to be done when decision-making in these circumstances. While our stance is on neither side of the debate, it is clear that both alcohol consumption and cigarette smoking prevalence enhance mortality due to the pandemic. Countries with high alcohol consumption per capita and high smoking prevalence are also at higher risk of mortality from pandemics like COVID-19. Smoking prevalence has more than doubled the effects of alcohol on COVID-19 mortality. There is certainly need for further research on these factors; however, it is clear that countries that reduce their population’s exposure to these substances will be at less risk of mortality due to the pandemic.

Reducing poverty and economic exclusion. One of the consequences of 21st century economic development is high and persistent inequality. This has resulted in less benefit sharing of economic growth, especially in many developing countries. Consequently, generating high levels of poverty and economic exclusion. Reversing these trends will significantly reduce pandemic-related mortality. There are various avenues through which poverty exposes a population to high pandemic-related mortality. One is less ability to build one’s health stock over time, due to lack of resources to procure individual health inputs. The second relates to the fact that poorer populations are more exposed to hazards and consequently both the pandemic and other diseases. The third channel is the possible association of poverty with low levels of hygiene and sanitation, such as inadequate or no access to clean drinking water. We are also witnessing that poverty reduces the effectiveness of measures aimed at curbing the spread of the disease. Direct access by the poor to economic opportunities, by addressing the constraints to both demand-side and supply-side of economic productive (human, physical, financial, natural and social) capital, is the first point of call. Effective social safety nets and various interventions in times of pandemics can reduce the mortality risk of the poor.

Government stringency measures for pandemic control. Governments have to be cautious with draconian measures. Governments have to learn from experiences like this to be able to fine-tune measures in a way that minimizes harm to the population, while being effective in curbing mortalities. In the current case, the best effect inferred from the point estimates is no effect, and the worst is that the same measures tend to increase mortalities. Quantile-based analysis shows that government measures seem to bear some fruits only later in the evolution of fatalities. It is likely that the fatality-curbing effects are due to governments that have managed to adapt their measures over time. Therefore, a good memory of the basket of effective measures in current times will be quite helpful in the future.

Conclusion

The aim of this paper has been to find out whether or not health systems globally are able to cope with the rapidly rising fatalities due to the COVID-19 pandemic, and if not, to assess the complementary social, economic, physical and environmental determinants that could be strengthened to curb mortality in the current and future similar pandemics. In doing so, I formulate models that link mortalities to social determinants of health. The resulting framework is then estimated using Poisson pseudo-maximum-likelihood (PPML) and quantile regressions. The data for COVID-19 and government stringency index are from Our World in Data (Ritchie 2020). Other variables that capture the social, economic, physical and environmental conditions that prevailed in a country immediately following the disease outbreak are from the World Bank’s (2020) WDI.From these analyses, certain useful findings have come to light, resulting in some health policy implications for the design of socio-economic fundamentals that can complement health systems in curbing mortality as the pandemic progresses, and in similar future pandemics. These polices relate to enhancing capacity for early testing for isolation of cases and treatment and contact-tracing; and improving health infrastructure to ensure adequate care of affected citizens. However, as the pandemic rises to catastrophic levels, health systems quickly become overwhelmed and ineffective in managing cases and reducing mortality. The results suggest that a combination of health infrastructure and other (social, economic, physical and environmental) factors can prove very effective in reducing mortality and managing the pandemic in a way that leads to less human cost.

These factors relate to improving the health stock of the population through reductions in both communicable and non-communicable comorbidities; enhancing sanitation and hygiene; and improving nutrition of the population. Socio-economic and environmental measures are the reduction of household and ambient air pollution; reduction of exposure to alcohol and cigarettes; reduction of poverty and ensuring economic inclusion; and fine-tuning government control measures in order to minimize harm to the population while effectively curbing mortality. While these policy proposals are valid, some of these factors require more probing as avenues for further research.