India’s Covid-19 case fatality rate is impressively low at 1.7%, compared to the world average of 4%. This apparently superior performance could, however, simply be a product of three factors: the country's relatively younger population, the lag between Covid infection and death, and the under-reporting of deaths. In this article, Philip, Ray, and Subramanian give precise quantitative expression to the first two factors. Based on the analysis, they contend that India’s Covid-19 experience does not indicate unqualified success relative to several other countries - with or without the undercounting of deaths.
On 10 September 2020, India’s case fatality rate (CFR) from Covid-19 stood at an impressively low figure of 1.7%. This is far lower than the world average. The Indian government has not been slow to tout this as proof that India is coping resiliently and well with the pandemic. The reaction from other quarters has been just as quick but not as positive: India has a particularly young population, which would tend to bring down the CFR. The lag between infection and death, coupled with increased testing, could also bring down the fatality rate by expanding the denominator (positive cases) relative to the numerator (fatalities). And of course, there could be severe undercounting of Covid-19 deaths.
Age distribution, and lag between infection and death
In this day and age, with echo chambers cranked up to maximum volume, it is easy to imagine these as largely academic arguments. But of course, they are not. The quantitative effect of these various propositions has implications for the assessment of both intrinsic and comparative performance of countries in combating the pandemic. That said, the emphasised phrase in the previous sentence is important. We need to know how strong these effects are, over and above the mere assertion that these are possible effects. In Philip et al. (2020), we do just this for two of the three counterarguments listed above. We give precise quantitative expression to the relevance of the age distribution, and to the relevance of the lag between infection and death. In this article, we describe our approach, and provide some updated results.
We first describe the two effects of interest in slightly greater detail. First, CFR does not take account of the precise age-specific distribution of Covid cases and deaths. It is well-known that ‘younger countries’ will exhibit lower CFRs simply because the young are more resilient to the virus. In principle, it is possible for a country like India to have a higher age-specific mortality rate, while still enjoying lower mortality on the whole.1
Second, there is a time lag that occurs between the onset of infection and the date of death (when infection leads to death, that is). Verity et al. (2020) report a mean (average) duration of around 18 days from infection to death, while Wilson et al. (2020) conclude that “... a median of 13 days passed from pneumonia confirmation to death...”. That suggests that the contemporaneous CFR – deaths on a given date divided by the number of cases at that same date – might be an inaccurate reflection of the true case-mortality rate. We would need to divide by the number of Covid cases at some relevant anterior date; call this the lagged CFR. It is, of course, no surprise to learn that lagged CFR for India is higher (Mohanty et al. 2020). But that would be true for all countries in the growing phase of the disease or as testing expands – not just India. What is more interesting are comparisons across countries, for it is only then that we can note the relative deflation in case fatality brought on by the growth in reported cases. In particular, it is easy to see (and risky to miss) the fact that all else equal, a country with a faster growth rate of confirmed cases exhibits a higher ratio of lagged CFR to contemporaneous CFR.
Empirical analysis: Prediction and decomposition
In what follows, then, we expand on these qualitative observations to get some hard numbers. We ask first: what would happen to India’s CFR if age-specific mortality rates were similar to those in other countries, including ‘older’ countries? That is, what would ‘predictions’ for India’s CFR look like, using age-specific fatality rates from comparison countries coupled with India’s distribution of Covid-19 cases? If these predictions are lower than India’s actual performance, it would lend serious credence to the age distribution counterargument. And of course, we could do this for both contemporaneous and lagged CFRs.
A second route to understanding age distribution is via the technique of ‘decomposition’. This is a sharper method, but more demanding on data requirements relative to the prediction exercise, as we need information on age-specific incidence and mortality for both India and the comparison country. One could then decompose the difference in CFRs, and thereby isolate two factors that combine to generate an overall statistic: one corresponding to the difference in CFRs based on differences in case distribution (an ‘incidence effect’), and the other to the difference in CFRs based on differences in age-specific fatality rates (the ‘fatality effect’). Like the prediction exercise, the decomposition method can also accommodate lagged CFRs with ease. And it is entirely possible that two countries have similar CFRs as a serendipitous consequence of the incidence and fatality effects operating in opposite directions and largely cancelling each other out. The CFR would then generate similar end-state descriptions while concealing vastly differing underlying processes leading up to that end state, a feature shared with many other similarly doubtful summary statistics.
In implementing these exercises, we compare India to a set of 17 other countries for which age-distribution information is either available or can be constructed.2 Each prediction and decomposition exercise comes in three variants: one with contemporaneous CFRS, and two more with lagged CFRs that employ 14- and 21-day lags.3
Results
The results of our empirical exercises are summarised in Tables 1 and 2. Table 1 contains the prediction exercise. The very first row contains the Indian CFR - contemporaneous and lagged - at three dates: 20 August, 30 August, and 10 September. The rows below contain the corresponding predictions for India that arise from a set of 17 comparison countries. Let us take one of them; say Spain. Spain had an average CFR of 6.6% on 30 August, over three and a half times as high as India’s. But the numbers in blue along Spain’s row reveal the hypothetical CFRs (both contemporaneous and lagged) that would have resulted were Spain to have India’s age distribution. Of course, those numbers are lower than 6% - Spain is an older country. But it is more than that - the numbers are lower than India’s rates, a much stronger statement. On 10 September, Spain’s contemporaneous prediction for India was just 68% of the Indian number. The corresponding prediction for South Korea was 45%. Lagged (by either 14 or 21 days), the predictions fall even more: to about 64% of the Indian number for Spain, and to under 40% for South Korea. These numbers speak for themselves.
More generally, all the underpredictions are highlighted in blue. Alas for India, these are the vast majority of the numbers.
Table 1. Predicted Indian lagged CFRs (LCFRs)
All data-lag combinations |
||||||||||
20 Aug |
30 Aug |
10 Sept |
||||||||
Country |
CFR 30 Aug |
0 |
-14 |
-21 |
0 |
-14 |
-21 |
0 |
-14 |
-21 |
India |
1.79 |
1.90 |
2.74 |
3.40 |
1.79 |
2.45 |
2.95 |
1.68 |
2.27 |
2.65 |
China |
5.25 |
2.73 |
2.77 |
2.80 |
2.73 |
2.75 |
2.77 |
2.73 |
2.74 |
2.75 |
South Korea |
1.64 |
0.90 |
1.01 |
1.03 |
0.78 |
1.00 |
1.05 |
0.76 |
0.88 |
1.01 |
Japan |
1.88 |
0.84 |
1.16 |
1.51 |
0.80 |
0.99 |
1.16 |
0.82 |
0.93 |
1.03 |
Philippines |
1.60 |
1.65 |
2.47 |
3.35 |
1.64 |
2.22 |
2.76 |
1.67 |
2.02 |
2.35 |
Netherlands |
8.93 |
1.91 |
2.19 |
2.30 |
1.79 |
1.99 |
2.14 |
1.61 |
1.83 |
1.93 |
Italy |
13.29 |
3.03 |
3.11 |
3.14 |
2.91 |
3.06 |
3.10 |
2.76 |
2.96 |
3.05 |
Spain |
6.60 |
1.62 |
1.97 |
2.14 |
1.40 |
1.79 |
1.96 |
1.14 |
1.47 |
1.67 |
Bavaria* |
4.62 |
1.76 |
1.84 |
1.87 |
1.67 |
1.80 |
1.84 |
1.57 |
1.70 |
1.77 |
Sweden |
6.93 |
1.90 |
2.00 |
2.05 |
1.86 |
1.93 |
1.98 |
1.83 |
1.88 |
1.91 |
Switzerland |
4.14 |
1.11 |
1.20 |
1.23 |
1.03 |
1.14 |
1.18 |
0.96 |
1.07 |
1.12 |
South Africa |
2.25 |
2.46 |
2.77 |
3.12 |
2.66 |
2.83 |
2.99 |
2.79 |
2.91 |
3.01 |
Chile |
2.74 |
2.08 |
2.22 |
2.30 |
2.10 |
2.23 |
2.31 |
2.10 |
2.23 |
2.30 |
Colombia |
3.18 |
2.84 |
4.13 |
5.17 |
2.84 |
3.73 |
4.52 |
2.87 |
3.44 |
3.93 |
Argentina |
2.12 |
1.63 |
2.34 |
2.88 |
1.73 |
2.40 |
2.88 |
1.71 |
2.38 |
2.79 |
Turkey |
2.35 |
1.90 |
2.04 |
2.10 |
1.87 |
2.02 |
2.09 |
1.91 |
2.08 |
2.15 |
Portugal |
3.16 |
1.00 |
1.06 |
1.08 |
0.97 |
1.03 |
1.06 |
0.92 |
1.01 |
1.04 |
California* |
1.83 |
1.37 |
1.65 |
1.81 |
1.38 |
1.56 |
1.74 |
1.42 |
1.54 |
1.63 |
Notes: Numbers in the first row report Indian LCFR for different dates and lags (0-, 14-, 21-day). Subsequent rows record counterfactual LCFR for India predicted using age-specific LCFRs of the respective comparison country (*state) and India’s case distribution. Underpredicting countries are highlighted in blue. Country-specific CFRs are reported in red for comparison.
Sources: Case and death distributions with sources listed in the Appendix of Philip et al. (2020), and LCFRs based on Roser et al. (2020) (published online at OurWorldInData.org retrieved from https://ourworldindata.org/coronavirus; online resource)
Table 2 summarises the decomposition exercise for a 21-day lagged CFR. India’s rates for three dates are shown on the top row, as before. Now, for any of the comparison countries, and for each of the three dates, the corresponding row lists, first, the lagged CFR, the raw difference with India (‘Diff’), the incidence effect (‘IE’), and the fatality effect (‘FE’). Not surprisingly, the raw differences are generally positive, which is the reason that the Indian government has declared satisfaction with India’s response to Covid-19. The next column, again generally positive, shows that the bulk of the raw difference is accounted for by the incidence effect in favour of India - the fact that the Indian age distribution is skewed in favour of the young. The last and most interesting column is that of the fatality effect. It is related to the prediction exercise in Table 1, but is not quite the same, as it estimates the difference using an average of the Indian and comparison age distributions - the prediction effect uses only the Indian age distribution. In line with the prediction effect, the majority of these terms - highlighted in blue – are negative, suggesting that age-adjusted, India is not doing as well as we might think. In fact, it is safe to say that compared to these 17 countries, India has been performing quite poorly.
Table 2. 21-day LCFR: Difference decomposition for India and comparison countries (*states)
20 Aug |
30 Aug |
10 Sept |
||||||||||
Country |
LCFR |
Diff |
IE |
FE |
LCFR |
Diff |
IE |
FE |
LCFR |
Diff |
IE |
FE |
India |
3.40 |
0.00 |
- |
- |
2.95 |
0.00 |
- |
- |
2.65 |
0.00 |
- |
- |
China |
5.40 |
2.00 |
2.55 |
-0.55 |
5.32 |
2.37 |
2.36 |
0.01 |
5.29 |
2.64 |
2.24 |
0.40 |
S. Korea |
2.15 |
-1.25 |
1.24 |
-2.49 |
2.21 |
-0.74 |
1.17 |
-1.90 |
2.12 |
-0.53 |
1.08 |
-1.61 |
Japan |
3.52 |
0.12 |
1.64 |
-1.52 |
2.70 |
-0.25 |
1.32 |
-1.57 |
2.40 |
-0.25 |
1.18 |
-1.42 |
Philippines |
3.27 |
-0.13 |
-0.08 |
-0.05 |
2.69 |
-0.26 |
-0.06 |
-0.19 |
2.29 |
-0.36 |
-0.06 |
-0.30 |
Netherlands |
11.51 |
8.11 |
7.63 |
0.48 |
10.72 |
7.77 |
6.91 |
0.86 |
9.68 |
7.03 |
6.22 |
0.81 |
Italy |
14.35 |
10.95 |
8.83 |
2.12 |
14.18 |
11.23 |
8.34 |
2.90 |
13.94 |
11.29 |
7.95 |
3.34 |
Spain |
10.09 |
6.69 |
7.09 |
-0.40 |
9.23 |
6.28 |
6.34 |
-0.06 |
7.86 |
5.21 |
5.52 |
-0.31 |
Bavaria* |
5.18 |
1.78 |
2.80 |
-1.03 |
5.09 |
2.14 |
2.62 |
-0.48 |
4.89 |
2.24 |
2.45 |
-0.21 |
Sweden |
7.61 |
4.21 |
4.25 |
-0.04 |
7.36 |
4.41 |
3.96 |
0.45 |
7.11 |
4.46 |
3.74 |
0.72 |
Switzerland |
4.95 |
1.55 |
3.42 |
-1.87 |
4.74 |
1.79 |
3.14 |
-1.34 |
4.49 |
1.84 |
2.90 |
-1.06 |
S. Africa |
2.64 |
-0.76 |
-0.41 |
-0.35 |
2.53 |
-0.42 |
-0.38 |
-0.05 |
2.54 |
-0.11 |
-0.36 |
0.26 |
Chile |
3.01 |
-0.39 |
0.56 |
-0.95 |
3.01 |
0.06 |
0.53 |
-0.47 |
3.00 |
0.35 |
0.51 |
-0.16 |
Colombia |
5.79 |
2.39 |
0.36 |
2.03 |
5.06 |
2.11 |
0.32 |
1.79 |
4.39 |
1.74 |
0.28 |
1.47 |
Argentina |
3.53 |
0.13 |
0.41 |
-0.29 |
3.52 |
0.57 |
0.40 |
0.18 |
3.42 |
0.77 |
0.38 |
0.39 |
Turkey |
2.64 |
-0.76 |
0.48 |
-1.24 |
2.62 |
-0.33 |
0.45 |
-0.78 |
2.70 |
0.05 |
0.44 |
-0.39 |
Portugal |
3.53 |
0.13 |
2.47 |
-2.34 |
3.46 |
0.51 |
2.28 |
-1.77 |
3.38 |
0.73 |
2.14 |
-1.40 |
California* |
2.39 |
-1.01 |
0.43 |
-1.44 |
2.30 |
-0.65 |
0.40 |
-1.05 |
2.15 |
-0.50 |
0.37 |
-0.86 |
Sources: Case and death distributions from sources listed in the Appendix of Philip et al. (2020), and overall LCFRs in Roser et al. (2020) (published online at OurWorldInData.org retrieved from https://ourworldindata.org/coronavirus; online resource) are applied to the decomposition formula in Philip et al. (2020) to obtain incidence effects (IE) and fatality effects (FE).
As an example of how aggregates could hide patterns, consider Japan on 30 August. It had a lagged CFR of 3.5%, just a shade higher than India’s number of 3.4% recorded on the same date. But the incidence and fatality effects are both large, each of these differences accounting for close to half the baseline death rates. The incidence effect works in India’s favour – Japan has the older population. But the fatality effect wipes out this advantage almost entirely – what India has in youth, it gives up in age-specific death rates.
In conclusion
Recall that there are other factors that could perhaps explain India’s low CFR, including the possibility of dramatic undercounting of Covid-19 deaths. After viewing these tables, the reader must judge the extent to which the undercounting of deaths is even needed to explain India’s exceptional numbers. Undercounting may well be present (Thapar 2020, Chatterjee 2020). We have little to say directly about this possibility, as we do not have the data, but Tables 1 and 2 together suggest that an undercounting argument need not even be invoked to understand the large discrepancies. India’s Covid-19 experience does not indicate unqualified success, relative to several comparison countries, with or without undercounting. And at the very least, there is reason to believe that an undiscriminating employment of the raw CFR as an indicator of success deserves to be treated with some scepticism. It seems to be important to assert this when both objective appraisal and fair accountability are threatened by summary indicators of performance that are inadequate or misleading, and when such summary measures are employed to their advantage by politicians and policymakers.
This article is published in collaboration with Scroll.in: https://scroll.in/article/973616/covid-19-why-indias-low-case-fatality-rate-should-be-treated-with-scepticism
Notes:
- See, for instance, Ray and Subramanian (2020) and Mukhopadhyay (2020) for India, and Dudel et al. (2020) for other countries.
- For description of sources and procedures, please refer to our paper (Philip et al. 2020), but the latest versions of the tables are in this I4I article.
- The decomposition procedure we use is the Shorrocks-Shapley exact decomposition procedure; the outcome of this decomposition coincides with a procedure advanced by Kitagawa (1955) for use in specifically demographic contexts. We are certainly not the first to carry out such decomposition exercises – for similar efforts, the reader is referred to Dudel et al. (2020), among others. The precise arithmetic of decomposition is dealt with in our paper, and is beyond the scope of this article.
Further Reading
- Chatterjee, Patralekha (2020), “Is India Missing COVID-19 Deaths?”, The Lancet, 5 December 2020.
- Dudel, Christian, Tim Riffe, Enrique Acosta, Alyson A. van Raalte, Cosmo Strozza, and Mikko Myrskyla (2020), “Monitoring trends and differences in COVID-19 case fatality rates using decomposition methods: Contributions of age structure and age-specific fatality”, medRxiv, 18 May 2020.
- Kitagawa, Evelyn M (1955), “Components of a Difference Between Two Rates”, Journal of the American Statistical Association, 50(272):1168-1194.
- Mohanty, Sanjay Kumar, Umakanta Sahoo, Udaya Shankar Mishra and Manisha Dubey (2020), “Age Pattern of Premature Mortality under varying scenarios of COVID- 19 Infection in India”, medRxiv, 11 June 2020.
- Mukhopadhyay, P (2020), ‘Is India’s Covid-19 death rate higher than Italy’s?’, Hindustan Times, 11 June 2020.
- Philip, M, D Ray and S Subramanian (2020), ‘Decoding India’s Low Covid-19 Case Fatality rate’, National Bureau of Economic Research (NBER) Working Paper No. 27696.
- Ray, D and S Subramanian (2020), ‘India’s Lockdown: An Interim Report’, CAGE Working Paper No. 476, 20 May 2020.
- Roser, M, H Ritchie, E Ortiz-Ospina and J Hasell (2020), ‘Coronavirus Pandemic (COVID-19)', OurWorldInData.org.
- Thapar, K (2020), ‘Could COVID-19 Deaths in India Actually Be 5 Times the Official Figure?’, The Wire, 25 August 2020.
- Verity, Robert, Lucy C Okell, Ilaria Dorigatti, Peter Winskill, Charles Whittaker, Natsuko Imai, Gina Cuomo-Dannenburg, Hayley Thompson, Patrick G T Walker, Han Fu, Amy Dighe, Jamie T Griffin, Marc Baguelin, Sangeeta Bhatia, Adhiratha Boonyasiri, Anne Cori, Zulma Cucunubá, Rich FitzJohn, Katy Gaythorpe, Will Green, Arran Hamlet, Wes Hinsley, Daniel Laydon, Gemma Nedjati-Gilani, Steven Riley, Sabine van Elsland, Erik Volz, Haowei Wang, Yuanrong Wang, Xiaoyue Xi, Christl A Donnelly, Azra C Ghani and Neil M Ferguson (2020), “Estimates of the severity of coronavirus disease 2019: a model-based analysis”, The Lancet Infectious Diseases, 20(6):669-677.
- Wilson, Nick, Amanda Kvalsvig, Lucy Telfar Barnard and Michael G Baker (2020), “Case-Fatality Risk Estimates for COVID-19 Calculated by Using a Lag Time for Fatality”, Emerging Infectious Diseases, 26(6):1339-1441.
By: Dr Susanta Kumar Padhy 27 September, 2020
Well written paper