Only a minority of all COVID-19 cases are getting recorded through testing and are reported. The second national seroprevalence survey in August–September 2020 suggested that the case detection ratio was around 6–7:100 (19). For the purpose of this exercise, as on July 31, 2020, it was assumed that the case detection ratio was at least 5:100. The extent to which COVID-19 deaths get reported among the large subset (95% or above) of unreported COVID-19 in the community depends on (i) the prevailing coverage of routine death surveillance; and (ii) extent of errors in MCCD (Figure 1). Within the reported COVID-19 cases also, the COVID-19 deaths could be missed due to the poor coverage of routine death surveillance and errors in MCCD. Despite WHO and ICMR guidelines (5, 6), the suspected COVID-19 deaths are not getting reported in India (20). The COVID-19 illness also presents a myriad of clinical presentations post-recovery that may lead to mortality which may not be captured by the present system in India (21).

Therefore, before we compare and interpret the differences in the reported COVID-19 deaths and the corresponding indicators of India with countries having near 100% routine death surveillance (Supplementary Table 1), we should adjust for the prevailing coverage and quality of routine death surveillance. This also applies to interpretation of differences across states within India.

At the national and the state level, the correction (multiplication) factor was inverse of the coverage of routine death surveillance. This factor varied from state to state. It was 5.5 nationally, (100/18.1) and ranged from 40 (100/2.5) for Jharkhand, to none (100/100) for Goa (Table 1). We calculated the estimated COVID-19 deaths, nationally and state wise, by multiplying the correction factor with the reported COVID-19 deaths. As on July 31, 2020, the estimated COVID-19 deaths in India were 196 997; an estimated 144 DPM (Supplementary Table 2).

As on July 31, 2020, Uttar Pradesh, Madhya Pradesh, and Jharkhand were not in the top 10 states based on reported DPM. When we adjusted for routine death surveillance and calculated the estimated DPM (Supplementary Table 2), they were in the top 10. The ranks of Bihar (25 to 17) and West Bengal (10 to 6) jumped upward (Supplementary Figure 1). Incidentally, these were the only five states that had <10 000 TPM. These five states constitute 43% of the national population. Based on reported COVID-19 deaths, these five states constituted 12% of total COVID-19 deaths but based on estimated COVID-19 deaths, they constituted 43% of the total COVID-19 deaths. The correlation coefficient (r) between TPM and coverage of routine death surveillance was 0.673 (p < 0.001).

As we used routine death surveillance data from 2018 and applied it to the year 2020, we also did sensitivity analysis for the correction factor by assuming different percentage increases in coverage of routine death surveillance between 2018 and 2020 (Supplementary Table 3). However, we do not expect any significant improvement as at the national level over the past 9 years, death registration has only increased from 67 to 86%, while the MCCD among registered deaths has been stagnant at approximately 20% (11, 12).

To adjust for under-reporting due to errors in MCCD among the deaths captured by routine death surveillance, we did not have an objective state-wise and national estimate of the error. There were state- and city-specific modeling exercises and media reports where the reported COVID-19 deaths (due to errors in MCCD) were as low as 50% (correction factor of two) of the actual deaths reported from hospitals (22, 23). We assumed two correction factors for errors in MCCD (1.5 and 2). We arrived at the estimated COVID-19 deaths after adjusting for errors in MCCD (Supplementary Table 2). The errors in MCCD explain why for a state like Goa with 100% routine death surveillance, under-reporting of COVID-19 deaths is still possible.

By multiplying the correction factor for error in MCCD and the correction factor for the prevailing routine death surveillance, we calculated a combined correction factor for the poor coverage and the quality of the routine death surveillance (Table 1). At the national level, the overall under-reporting could therefore possibly be by a factor as low as 5.5 (5.5^*1) or as high as 11 (5.5^*2) (Table 1). It could be higher than 11 if the upper limit for correction factor for errors in MCCD is more than two. Using a correction factor of 11, as on July 31, 2020, the estimated deaths could be as high as 393 932 (288 DPM) (Supplementary Table 2). Similarly, we calculated the range of combined correction factor for states (Table 1).

We created a matrix to calculate the estimated IFR for India as on July 31, 2020 taking a range of scenarios for the estimated COVID-19 deaths and the case detection ratio (Supplementary Table 4). With a correction factor ranging from 5.5–11 and case detection ratio of 5:100, the estimated IFR ranged from 0.58:100–1.16:100.

We arrived at an estimate of the COVID-19 under-reporting in India using publicly available indicators regarding the coverage of routine death surveillance and some assumptions in errors in MCCD. Hence, there are two levels of under-reporting. First, COVID-19 death not being captured by routine death surveillance. Second, though the death is captured, the COVID-19 death is missed due to error in MCCD (Figure 1). This is also relevant for other countries with poor routine death surveillance (countries in south-east Asia and Africa). A strong positive correlation between TPM and the prevailing coverage of routine death surveillance suggests that states with weak mortality surveillance were also the ones who had conducted fewer tests.

To put the estimated COVID-19 deaths in perspective (around 200 000–400 000 as on July 31, 2020), every year there are an estimated 127 000 (95% CI: 64 046, 190 139) influenza- associated deaths in India. With respect to the under-reporting of deaths, an estimated 450 000 TB- related deaths occur annually in India which is 5.6 times the reported TB deaths (24–26).

As on July 31, 2020, of the top 10 countries (based on reported cases), eight had 90%−100% coverage of routine death surveillance (India: 18%, Peru: 57%) and all had higher reported COVID-19 DPM than India (3, 27). Small countries like the UK and Belgium with high DPM tested almost 25% of their population; have a robust routine death surveillance system and they reconcile COVID-19 deaths from their routine death surveillance system (3, 27). Other countries like Brazil, Mexico, Chile, and Peru that have a similar age distribution and socioeconomic status also have higher reported DPM when compared to India (28). The difference could be explained by their better routine death surveillance (Supplementary Table 1). As on May 13, 2021, when we compared with countries in south-east Asia, where most of the countries have a poor routine death surveillance (27), India had the highest reported COVID-19 DPM (3). Additional factors to consider before comparing DPM are the estimated cases based on seroprevalence surveys, death definitions used, testing strategies adopted, response to the epidemic, prevalence of comorbidities, quality of medical care, and population density.

Countries with high coverage of routine death surveillance may also miss COVID-19 deaths due to errors during MCCD or if they do not reconcile deaths from routine death surveillance. But the overall extent of under-reporting is expected to be lower when compared to countries with poor death surveillance as chances of missing deaths (irrespective of the cause) and cause-specific deaths are comparatively less due to robust mortality surveillance systems.

Using reported deaths in the numerator, the ICMR study (at around 1:100 case detection ratio in May–June) suggested the IFR to be 0.15:100 in high stratum districts and 0.01:100 in low stratum districts (29). The IFR estimated by us for India (0.58:100–1.16:100) is within the range estimated in countries with near perfect routine death surveillance (0.5:100 in Switzerland to 1.4:100 in Italy) and similar to the point estimate (0.68:100) based on a systematic review and meta-analysis (30, 31).

The combined correction factor provides a bird's-eye view. The best way to estimate the COVID-19 under-reporting using publicly available data would be to factor in excess deaths during the COVID-19 epidemic as compared to previous years with a sub-group analysis of age groups and cause of death (if available). The difference between excess deaths and reported COVID-19 deaths will give an idea about the possible extent of under-reporting (32). Subject to availability of granular data, exact number of excess deaths attributed to COVID-19 may be calculated by adjusting for rate of change in the number of registered deaths over the past few years, decrease in accident- and pollution-related deaths during COVID-19, and a possible increase in deaths due to disruption in routine health services. Another option is to look for excess home deaths (33). In India, the place of death (home or health facility) is a variable collected during death registration (11). The “Excess deaths” analysis is a practical option for cities, districts and states with high coverage of death registration. However, with few exceptions, the excess deaths data has not been made public like elsewhere (34, 35).

In countries with high coverage of routine death surveillance, missed suspected deaths (due to error in MCCD) can also be identified based on an increase in deaths due to pneumonia, respiratory failure, sepsis or ill-defined causes higher than the maximum limit for the number of weekly occurrences of each cause (36).

In addition to excess deaths analysis, another feasible option is post-mortem surveillance. This would mean testing all registered deaths in a surveillance area for a specific surveillance period for COVID-19 at frequent intervals. This is feasible and not resource intensive. A study from Zambia (2020) suggested that of the 364 deaths around the tertiary health facility of capital city Lusaka, 70 were COVID positive (51 in community and 19 in facility) based on post- mortem polymerase chain reaction testing. Of 70, only six were tested pre-mortem for COVID giving an under-reporting by a factor of 10 (37). Zambia has not reported the routine death surveillance coverage data to WHO indicating its routine death surveillance is poor (27).

The coverage of routine death surveillance cannot be improved overnight but errors in MCCD can be prevented among the captured deaths. We can ensure that all suspected COVID-19 deaths are tested and we do not wrongly assign co-morbidity as the cause of death in confirmed COVID-19 patients.

In India, by using reported deaths to infer mortality, we are indirectly encouraging states with poor surveillance and discouraging states with relatively good surveillance. This also does not create an environment of data transparency. This false optimism of “low mortality” based on reported deaths could result in laxity among the administrators where appropriate mitigation measures may not be implemented. The health infrastructure may not be appropriately strengthened. The public health messaging based on this narrative may result in laxity among the public in following COVID19 appropriate behavior (10). The second wave that we witnessed in April–May 2021 in India could be a result of this “false optimism.” This could prove to be disastrous in states with poor routine death surveillance.

The administrators (local/state/national) should be encouraged to monitor and track the estimated COVID-19 deaths and act accordingly. If the administrators act only based on the reported deaths or only when the hospital infrastructure is over-stretched (only 34% deaths happen in hospitals in India) (11), it will be too little and too late.