Estimating COVID-19 mortality in Italy early in the COVID-19 pandemic

Abstract

Estimating rates of COVID-19 infection and associated mortality is challenging due to uncertainties in case ascertainment. We perform a counterfactual time series analysis on overall mortality data from towns in Italy, comparing the population mortality in 2020 with previous years, to estimate mortality from COVID-19. We find that the number of COVID-19 deaths in Italy in 2020 until September 9 was 59,000-62,000, compared to the official number of 36,000. The proportion of the population that died was 0.29% in the most affected region, Lombardia, and 0.57% in the most affected province, Bergamo. Combining reported test positive rates from Italy with estimates of infection fatality rates from the Diamond Princess cruise ship, we estimate the infection rate as 29% (95% confidence interval 15-52%) in Lombardy, and 72% (95% confidence interval 36-100%) in Bergamo.

Fig. 1. Validating counterfactuals for the pre-pandemic data.

we show the observed weekly mortality due to all causes for the period of 1 January to June 27 (black) in all 20 regions in Italy, and our prediction for the expected mortality in the absence of COVID-19 (conditional Gaussian Process (CGP) with its 1 and 2−σ error from the variance of the Gaussian model, i.e., 68% and 95% confidence interval, respectively, in green and synthetic controls method (SCM) in orange). The first reported COVID-19 mortality occurred in the week ending on February 22 (thin red vertical line). The historical data from 2015 to 2019 (blue) and corresponding historical mean (gray) is shown for comparison and are not a good fit to the observed pre-pandemic data. In the dashed-black line, we also show the observed mortality after removing reported COVID-19 deaths.

Fig. 2. Excess mortality compared with reported COVID-19 deaths in regions of Northern Italy and the province of Bergamo.

(a) Excess weekly deaths, and (b) cumulative excess deaths, over the predicted counterfactual in comparison to the reported COVID-19 deaths (in pink) for the period since February 23rd (available COVID-19 data). Estimates from both the synthetic controls method (SCM, orange) and conditional Gaussian Process (CGP, green) counterfactuals agree. We show 1 and 2−σ error (68% and 95% confidence interval) from the variance of the Gaussian model. We find that COVID-19 deaths are under-reported by multiple factors for every period and every region. We extrapolate the data excess beyond June 27th, which is the last week with available total mortality data (dashed-black line), with dashed-lines. To do this, we make the conservative assumption after June 27 that the reported COVID-19 deaths are accurate and account for the excess mortality over predicted trends.

Fig. 3. Age distribution of excess mortalities.

Same as Fig. 2b but for different age groups. We find a statistically significant excess over the reported COVID-19 deaths that is increasing with age.

Fig. 4. Fraction of missed deaths over time.

For the period of the pandemic, we show per week the fraction of missed deaths (green) with corresponding 1−σ (68% CI) estimated from the variance of the Gaussian model, the number of hospitalizations (normalized with the maximum weekly hospitalization up to 18 April 2020, in orange) and the number of tests conducted (normalized as a fraction of tests conducted in the week of 11–18 April 2020, in blue). We find that the missed fraction goes down as the number of tests increases while the hospitalizations have remained consistently high in the last 4 weeks.

Fig. 5. Fatality rates for different age groups and regions.

(Left) Population fatality rate (PFR) from the cumulative estimates divided by the regional population. (Center) Lower bounds on infection fatality rate (IFR) using the maximum test positive rate (TPR) as an upper bound on infection fraction. (Right) Estimates of the true IFR when normalizing the age 70–89 group to the Diamond Princess IFR (in shaded blue, with the corresponding Poisson error estimate). We also show estimates from Verity et al. with corresponding (68% CI). in magenta, which gives less steep age dependence. In the center and right panel, the gray lines are weighted mean estimates for IFR with 1 sigma weighted standard deviation bands. The horizontal lines are the age-averaged IFR for the entire population. Error bars for all the regions and age groups are 1−σ (68% CI) error from the variance of the Gaussian model combined with Poisson errors based on number of deaths that differs for every region and age group. In all panels, we have staggered the points horizontally for every age group for better visibility.