Inequality in healthcare: saturation of hospitals and death by COVID19




S. Falcone and E. Navarra
Authors: Stefano Falcone (ECARES; Université Libre de Bruxelles; email: Stefano.Falcone@ulb.be) and Elisa Navarra (ECARES; Université Libre de Bruxelles; email: Elisa.Navarra@ulb.be).^{1}
The spread of COVID19 pandemic has put healthcare systems around the world under strain. Is there any relationship between the burden on hospital facilities and the number of deaths by COVID19?
In this study, we exploit regional variation in the Italian healthcare system and the severity of the contagion to provide suggestive evidence that wellequipped intensive care units (ICUs) can save human lives from deaths by COVID19. We show that this result holds even when controlling for possible confounding factors and that the relationship is not linear: one additional intensive care bed (IC) is more strongly associated with fewer deaths by COVID19 when ICUs are more extremely saturated.
To grasp the intuition of this relationship, in Figure (a) we plot the number of beds in ICUs (per 100 infected patients) against the number of total deaths by COVID19 (per 100 infected patients) as of April 30, 2020. By dividing both the number of deaths and the number of IC beds by the number of total infected, we are conditioning on the severity of the contagion which spread faster in Northern regions. A spurious correlation between number of deaths and number of IC beds would in fact be positive, since Northern regions were more severely hit by COVID19 and their ICUs already better equipped before the start of the crisis than those in the Centre and South of Italy. The association between ICUs’ potential saturation rate and COVID19’s mortality rate is negative and statistically significant at the 1% level (Table 1, column 1): 1 additional IC bed (per 50 infected patients) is associated with 1 saved life (per 100 infected patients). However, this linear correlation masks a strong heterogeneity across Italian regions, with a tighter association in regions where the saturation of ICUs hits more severely.
The association between the strain on ICUs and deaths by COVID19 is confirmed when looking at the effective burden on ICUs (namely, IC beds per infected in ICU). Weighting the number of IC beds by the number of infected patients in ICUs allows us to rely on a more comparable measure of ICUs’ saturation across regions. In fact, the total infected statistics might vary also due to differences in unobservable characteristics of tests and testing procedures. In Figure (b), we plot total deaths (per infected in ICU) against IC beds (per infected in ICU). We use the number of infected in ICUs at its regional peak, so that the selected date may be different across regions. As in Figure (a), the correlation is only significant at the 10% significance level. However, this correlation masks a strong heterogeneity across regions, with IC beds more strongly associated with deaths by COVID19 when intensive care facilities are more extremely saturated.
While highlighting a significant positive correlation between saturation of IC beds and COVID19 mortality, Figures (a) and (b) also suggest a geographical clustering of regions in mildly and extremely saturated ones. We define as mild saturation, regions with 6 or more IC beds per 100 infected and extreme saturation those with less than 6 IC beds per 100 infected patients (for Figure (b), these are, respectively, 2 or more IC beds per infected in ICU, and less than 2). Mildly and extremely saturated regions are the same in per100infected and perinfectedinICU terms, with the exception of FriuliVenezia Giulia (extreme only in the first measure) and Puglia (extreme only in the second measure). The severity of ICU saturation almost overlaps with the geographical clustering in Northern versus Central and Southern regions: in fact, healthcare facilities are under a heavier strain in the North due to the more dramatic spread of COVID19 in that part of Italy, despite the preexisting higher number of IC beds per capita. Given this geographical clustering, we considered appropriate to produce heteroscedasticityconsistent standard error and control for socioeconomic indicators that could confound the associations shown in the Figure. Outcome variables, regressors of interest and selected controls are listed in Table 1, showing mean differences between extreme and mild saturation regions, defined in per100infected terms. Only the difference in life expectancy is statistically significant at 5% level, which makes it a good candidate as control.^{2} The inclusion of controls in our regressions confirms that ICUs’ saturation is a statistically significant predictor of COVID19 mortality (Table 1, columns 2 and 6). In addition, the magnitude of the relationship is higher with controls, pointing to a downward confounding due to extremely saturated regions being also characterized by a longer life expectancy. For a given public health expenditure, availability of physicians and life expectancy of the population, one additional bed (per 60 infected) can save 1 extra life (per 100 infected), while 3 additional beds (per infected in ICU) can save 2 human lives (per infected in ICU).
To further investigate the heterogeneity of the correlation, in columns (3)(4) and (7)(8) of Table 2, we show separately results for mild and extreme saturation regions. In both cases, we find that the correlation between saturation of ICUs and COVID19 mortality is strongly significant for extremely saturated regions and not significant for less saturated regions. When using per100infected variables, the magnitude of the correlation is around 5 times stronger in regions extreme than for the whole sample, while it is 6 times higher for perinfectedinICUs variables. In extremely saturated regions, 1 additional IC bed (per 60 infected in ICU) is correlated with 5 saved lives (per 100 infected patients), while 3 additional beds (per infected in ICU) can save 12 lives (per infected in ICU). This confirms that the burden on ICUs is more detrimental in regions where the saturation of ICUs hits more severely.
Although these correlations do not necessarily unveil causal relationships, our results provide preliminary evidence of a nonlinear positive link between ICUs saturation and deaths by COVID19. Our findings hold even when considering the new IC beds built with great effort in the last weeks (result not shown here). If the causal nature of these associations is confirmed and in view of a potential second wave of COVID19, then investing in ICUs and hospital equipment could contribute to reduce the lethality of the disease. Given the lower number of IC beds per capita in the South than in the North of Italy, due also to new equipment having being built in the last weeks in more severely hit regions, our results suggest that the cost in terms of human lives in the South would be dramatically higher than in the North shall the second wave of COVID19 hit with equal severity across the national territory. Moreover, our study provides evidence in favor of policies directed to the relocation of patients from extremely to mildly saturated regions.
Linear Model
\(\text{Deaths (per 100 infected)} = \alpha + \beta\ \text{IC beds (per 100 infected)} + \epsilon\)
\(\beta = 0.51\)
\(SE(\beta) = 0.11\)
\(\text{Average Deaths (per 100 infected)} = 10.07\)
Linear Model
\(\text{Deaths (per infected in ICU)} = \alpha + \beta \text{IC beds (per infected in ICU)} + \epsilon\)
\(\beta = 0.64\)
\(SE(\beta) = 0.31\)
\(\text{Average Deaths (per infected in ICU)} = 4.24\)
Note: Deaths is the total number of deaths by COVID19 as of April 30, 2020. ICU beds is the total number of beds in ICUs in 2018. Variables in per100infected terms are divided by the total number of infected times 100.
Data source: Protezione Civile (Deaths and infected by COVID19); Ministry of Health (ICU beds).
Table 1: Balance Test
Mean Extreme  Mean Mild  Difference  pvalue  

Deaths per 100 infected  11.533  7.872  3.661  0.012 
Deaths per infected in ICU  5.323  2.604  2.719  0.009 
ICU beds per 100 infected  2.657  10.840  8.183  0.000 
ICU beds per infected in ICU  1.100  3.623  2.523  0.000 
Life expectancy  83.235  82.530  0.705  0.018 
Public health exp. per capita  1,914.156  1,865.410  48.746  0.405 
Physicians per 1,000 inh.  3.950  4.096  0.146  0.472 
Observations  20 
Note: Mean Extreme reports the mean of listed variables in regions with less than 6 IC beds per 100 infected patients. Mean Mild reports the mean of listed variables regions with 6 or more IC beds per 100 infected patients. Difference calculates the difference between Mean Extreme and Mean Mild. pvalue reports the pvalue for the ttest on the difference of the means. Deaths is the total number of deaths by COVID19 as of April 30, 2020. ICU beds is the total number of beds in ICUs in 2018. Variables in per100infected terms are divided by the total number of infected times 100. All variables in perinfectedinICU are divided by the number of infected patients in ICU at its regional peak. Life expectancy is the regional life expectancy at 0 to 4 years old in 2018. Public health exp. per capita is the regional public health current expenditure in euros divided by regional population in 2016. Physicians per 1,000 inh. is the regional total number of physicians (generalists and specialists) per 1,000 inhabitants in 2017.
Data source: Protezione Civile (Deaths and infected by COVID19); Ministry of Health (ICU beds); Italian National Institute of Statistics (Physicians per 1,000 inh., Life expectancy, Public health exp. per capita).
Table 2: Table
Deaths per 100 infected 
Deaths per infected in ICU 


(1)  (2)  (3)  (4)  (5)  (6)  (7)  (8)  
Beds (infected)  0.512^{***} (0.107) 
0.599^{***} (0.128) 
0.334 (0.238) 
2.882^{**} (0.832) 

Beds (ICU)  0.635^{*} (0.314) 
0.712^{*} (0.388) 
0.195 (0.369) 
4.272^{**} (1.648) 

Life expectancy  0.814 (0.872) 
0.190 (0.849) 
1.618 (1.245) 
0.193 (0.735) 
1.122 (1.407) 
0.0584 (1.110) 

Health exp.  0.00419 (0.00471) 
0.00268 (0.00516) 
0.00132 (0.00565) 
0.00383 (0.00374) 
0.00274 (0.00540) 
0.00836 (0.00509) 

Physicians  0.247 (1.157) 
0.401 (1.607) 
3.336 (1.847) 
0.497 (1.260) 
0.657 (1.451) 
1.549 (1.725) 

Observations  20  20  8  12  20  20  8  12 
Mean Dep. Var.  10.07  10.07  7.872  11.53  4.234  4.234  2.870  5.142 
Sample  All  All  Mild  Extreme  All  All  Mild  Extreme 
Standard errors in parentheses. ^{*}p < 0.10, ^{**}p < 0.05, ^{***}p < 0.01.
Note: In regressions (1) to (4), Deaths per 100 infected is the total number of deaths by COVID19 (per 100 infected patients) as of April 30, 2020. In regressions (5) to (8), Deaths per infected in ICU is the total number of deaths divided by the number of infected patients in ICUs at its regional peak. Beds (infected) is the total number of beds in ICUs in 2018 per 100 infected patients. Beds (ICU) is the total number of beds in ICUs in 2018 per infected patient in ICU. Life expectancy is regional life expectancy at 0 to 4 years old in 2018. Health exp. is the regional public health current expenditure in euro divided by regional population in 2016. Physicians is the regional total number of physicians (generalists and specialists) per 1,000 inhabitants in 2017. Mean Dep. Var. reports the average number of deaths per 100 infected and the average number of deaths per infected in ICUs for the regression samples. Regressions (1) and (5) do not include controls. Regressions (1), (2), (5), (6) are run over the entire sample. Regressions (3) and (7) are run over the Mild subsample defined, respectively, as regions with 6 or more IC beds per 100 infected, and regions with 2 or more IC beds per infected in ICU. Regressions (4) and (8) are run over the Extreme subsample defined, respectively, as regions with less than 6 IC beds per 100 infected, and regions with less than 2 IC beds per infected in ICU. Robust standard errors in parenthesis.
Data source: Protezione Civile (Deaths and infected by COVID19); Ministry of Health (ICU beds); Italian National Institute of Statistics (Physicians per 1,000 inh., Life expectancy, Public health exp. per capita).

We thank Glenn Magerman, Federico Gallina and other members of Learning from the curve ↩

The unshown balance test using mild and extreme defined over infectedinICU terms yields qualitatively identical results. ↩