An Evidence-Based Estimation Of Covid Deaths That Is Far Better Than NYT’s

In India’s milder Covid-19 wave-1, we did way better than the West on reported deaths per million. That India didn’t see a perceptible jump in all-cause mortality, as the West did, suggests that our Covid deaths weren’t materially undercounted.

This is strongly corroborated by 2020 mortality data across states, cities and insurers. We likely ended wave-1 with less than 300 million infected, less than two lakh true deaths and an infection fatality rate between 5-10 per 10,000.

Wave-2 was way worse in India, just as in the West. Its 12x steepness vs wave-1 temporarily overwhelmed systems. This led to a month-long ‘spike-phase’, when we scrambled to ramp up testing, oxygen, beds and infrastructure.

Elevated deaths and undercounting were more likely during this phase.

However, elevated doesn’t automatically equal massive, as implied by the 5-10x claims backed by unreliable anecdotes. We need a reliable answer to the question: “Elevated, but by how much”?

I attempt a better approach, by analysing exhaustive death statistics during the spike-phase across multiple Indian locations. I calculate unexplained excess deaths per capita during this phase and (aggressively) extrapolate this across India.

Evidence indicates that, even during spike-phase, undercounting was under 3x, likely 2-2.5x. Outside of the spike-phase, wave-2 is no different than wave-1, on both disease intensity and data reliability.

Wave-2 will likely end with around three lakh reported deaths, six or seven lakh true deaths, 700-800 million infected and an infection fatality rate (IFR) closer to upper end of 5-10 per 10,000.

To put India’s fatalities in perspective, even after the above adjustments for undercounting, Covid will account for less than 5 per cent of all-cause mortality over 2020 and 2021. The equivalent figure for the US is 15-20 per cent. Much as introspection is necessary to better handle future spikes, self-flagellation is unnecessary.

Background

A recent New York Times article guessed that India’s true Covid death toll could be as high as 4 million (at the upper end of its guessestimates). Their flawed approach doesn’t separately analyse two waves and uses an oversimplified formula with absurdly pessimistic assumptions.

While ignoring is a fine option, an incessant stream of baseless doomsday narratives impacts discourse, and more importantly, collective mood. As we tame Covid and resume livelihood, it is important not to be nudged into collective self-doubt by biased peddlers of flawed stories.

So, here’s a better approach to the same topic. I study both waves separately, using appropriately different approaches. Being in the rear-view mirror, wave-1 can be reliably analysed, as independent statistical confirmation of our Covid reporting is available.

In wave-2, we have literally just come up for breath. To fight recency and partial evidence, I have to be more creative in methodology without compromising rigour. While wave-2 inferences are less reliable, it’s possible to do better than sweeping anecdotes or sensationalist guesswork.

Unlike in West, Covid wave-1 did not dent India’s all-cause mortality.

In 2020, Covid deaths per million in Western countries were an order of magnitude greater than in India. As a consequence, all-cause mortality showed a step-jump from near-zero growth to double-digit growth. While all-India mortality data isn’t available yet, multiple states have independently confirmed that India saw no such trend.

Kerala and Gujarat witnessed fewer deaths in 2020 over 2019 (read here, here).

Rajasthan showed a 6 per cent increase over 2019. Delhi, which had worst deaths per capita in wave-1, publicly stated that 2020 deaths were lower than 2019.

Mumbai, another wave-1 hotspot, was an exception, with 24 per cent higher deaths in 2020 against the previous five-year average.

With reported Covid deaths accounting for half of this increase, residual blip is modest after adjusting for population growth.

Kolkata showed 2 per cent growth in total deaths and decline if we exclude Covid deaths. While not a perfect measure, an analysis of pan-India insurer death-claims also showed 2020 growth in line with prior trend.

India’s Wave-1 Death Count Is Generally Reliable

Had our Covid death count been way off, India would have seen a perceptible blip in all-cause mortality. That we didn’t suggest that our true Covid death count was not far from reported death counts. This is confirmed by the aforementioned negative-to-modest growth in deaths in 2020 over 2019 across cities and states.

However, reliable does not mean error-free. The US, with better death registration systems, is estimated to have true Covid deaths at 1.5x of reported deaths.

Our weaker systems, especially in poorer states and rural areas, make it possible for our undercount factor to be slightly higher (although majority of deaths occurred in richer states with near 100 per cent death registration).

All things considered, India’s true death count could have been 1.5-2x of reported death count (considered ‘normal’ undercounting, all the world over), but no higher. For what it’s worth, our Covid testing was superior to the US in the first wave, with India’s test positivity rate at 5 per cent versus 8 per cent for the US.

Lest We Forget, Wave-1 Too Had Its Share Of Fearmongering Anecdotes

Having experienced tempestuous wave-2, wave-1 feels like a breeze. To refresh your memory, we had nearly as many alarmist death-undercounting accusations in 2020.

Professors suggesting 5x deaths across India, volunteer armies finding 2x deaths, queues outside crematoriums, graveyard counts not tallying with reported deaths are stories from 2020, not 2021. With 2020 hindsight (pun intended), these anecdotes seem unfounded if not absurd.

In real-time though, they resonated as well as today’s anecdotes do. Before falling prey to unreliable 2021 stories, it’s worth reminding ourselves that this movie has played out before (read more from the 2020 playlist here, here, here).

India’s Wave-1 Parameters Can Be Reliably Established

Since a majority of Covid cases are asymptomatic and undetected, it’s better to understand Covid parameters in terms of true infections, not reported cases.

India’s national sero-survey in late-December 2020 showed 21.4 per cent pan-India antibody prevalence. That corresponds to 300 million infected people.

Since, at that point, we had 10 million reported cases, our multiplier of true to detected infections is 30x. With 0.15 million deaths in 2020, India’s infection fatality rate was 0.05 per cent, or 5 per 10,000.

While the NYT uses this approach, their worst-case scenario assumes an absurd infection fatality rate of 0.6 per cent, or 12x what was observed in wave-1.

In summary, our wave-1 data is generally reliable. Undercounting was ‘normal’. Severity was low enough to not dent all-cause mortality in 2020, unlike in the West.

If we mark the end-point of the first wave as February 2021 (since wave-2 started in March), India saw 0.16 million reported Covid deaths and 11 million reported cases. True deaths could have been 1.5-2x higher, like elsewhere, and true infections 30x higher.

India’s Second Wave Is Way Different, Due To Sheer Steepness Of Spike

Apart from recency making for unreliable analysis, the main factor to incorporate into any study of the second wave is its sheer steepness. While we reached one lakh cases in six months in wave-1, we reached four lakh cases in two months in wave-2.

Wave-2 was an unmanageable 12x steeper than wave-1. Worse, we went from 80 deaths per day in February to 4,000 deaths per day in April.

That’s a 50x increase in strain on our administrative and healthcare systems in two months. No system can cope with this.

Temporary loss of control is inevitable, as we experienced in April, before hauling down disease in May. I refer to this part of wave-2, roughly lasting a month, as the ‘spike-phase’.

While the B.1.617 variant clearly spreads faster, experts don’t believe it to be extra virulent (ie higher fatalities per infection).

However, fatalities aren’t merely a function of virus and our vulnerability to it. It’s also a function of healthcare response in averting preventable deaths. Dramatically higher steepness and an overwhelmed system can result in extra fatalities as well as abnormal undercounting.

Non-Linear Nature Of Covid Spike Makes Extreme Anecdotes Inevitable

Covid is a seasonally concentrated disease. The worst month of wave-1 saw 2.5x more fatalities than a typical month. System stress and denial of healthcare for other ailments exacerbates peak intensity.

If Covid contributed to 2-3 per cent of India’s all-cause mortality over 2020, the worst few weeks in a badly hit city could be 10x worse, with a perceptible rise in all-cause mortality. The relationship between peak (localised) damage and average pan-India damage is highly non-linear.

Even on reported numbers, the wave-2 peak saw 4x deaths as the wave-1 peak. During wave-2’s spike-phase, Covid alone could drive all-cause mortality to 1.5x normal levels in a hotspot.

With healthcare system scrambling to fighting one disease, other ailments, accounting for over 80 per cent of India’s deaths, are underserved.

All things considered, total deaths running at double of normal is possible in badly-hit spots during a spike phase. Administrative measures can further worsen perception.

Covid protocols for most funerals worsens crowding and inflates superficial body counts. As does patients flocking to medical hubs and designated crematoriums.

As we saw in wave-1, anecdotes are always unreliable inputs into analysis and public policy. They turn positively dangerous in spike phase, when they are more numerous and seem more believable.

While it feels heartless to ignore poignant photos and tearful stories, we have to do so to reliably estimate true damage during the spike phase.

While I seek data from the spike phase, I limit it to systematically gathered data, ignoring cameras, interviews and opinions.

Wave-2 Is Best Analysed In Two Parts: Spike Phase And Rest Of Wave

It’s clear that wave-2 has two distinct parts. A spike phase where all bets seem off. A ‘normal’ phase where the system caught up with Covid and tamed it.

The former is an extraordinary period. The latter is no different from wave-1. Representativeness is key to sound analysis. No data from outside the spike phase can represent its chaos and elevated mortality.

So, we have to extrapolate from whatever reliable data is available within it. Likewise, spike-phase data is singularly unrepresentative of the rest of the wave. We can’t extrapolate from worst month to entire wave or year. Ergo, a two-part analysis is needed for wave-2.

Before starting analysis, let me define spike phase more precisely. The best measure is India’s test positivity rate going above 15 per cent, reflecting the system’s inability to even track disease.

This lasted a month, from mid-April to mid-May (averaging 21 per cent over this month). Other metrics were elevated over this phase. Medical oxygen demand spiked from 1,500 tonnes to 9,000 tonnes.

Reported cases stayed above 3.5 lakh a day and deaths around 4,000 a day for a month.

While there’s some lead/lag in these metrics, and not all states simultaneously hit their peaks, I’ll treat the spike phase as a one-month period over which India’s system lagged disease. Simply think of it is as our worst month.

Even During Spike-Phase, Evidence Suggests That Undercounting Was Below 3x

An estimate is as good as the data that goes into it and the method used. Fortunately, in four places (Gujarat, Delhi, Kanpur, Kolkata), media reports (read here, here, here) used comprehensive death records maintained by their respective governments, covering both the spike phase and an appropriate baseline period (with biases common to both).

Analogous statistics for the peak period of wave-1 were also available from Chennai, Mumbai and Nagpur (read here, here, here), allowing additional validation.

Now, we need a method that facilitates comparison between locations and extrapolations across India. My approach is to extract ‘unexplained excess-deaths’ by subtracting Covid deaths from excess deaths over the baseline period.

Then, I normalise it to ‘per month per million population’ to make different periods and places comparable.

Also, numbers thrown out in isolation are meaningless (40,000 here, 11,000 there) without an appropriate denominator. If this metric falls within a reasonable band across independent data points, it can be treated as representative of the spike phase.

A few points before extrapolating. First, Covid is unlikely to be sole cause of 100 per cent of the unexplained excess deaths.

Second, while statistically robust, there’s a selection bias. Media chose to highlight these stories over others. Knowing that negativity sells, these are likely to be at the adverse end of what they found.

Third, wave-1 showed subsequent mean-reversion in such excess deaths (ie they turned negative in later months). Even in wave-2, Delhi is already seeing funerals lag reported deaths.

Fourth, the above reports perform non-existent or inadequate population adjustment in baseline (the Gujarat study has other serious inconsistencies that I’ll not digress into).

While this data is broadly representative of the spike phase, these four factors point to strong overestimation. As I extrapolate to pan-India, this yields an inflated upper bound for undercounting. While still useful, it’s way higher than the most likely outcome. Now, here’s the data.

Unexplained excess deaths per month per million (see G in table below) are clustered around an average of around 200 per million monthly across a range of locations.

Wave-1 and wave-2 aren’t that different on this metric, perhaps indicating that, in already recording 4,000 Covid deaths daily across India, overcounting wasn’t as bad as expected.

This already hints that wave-2 evidence isn’t extraordinary enough to support extraordinary 5-10x death claims.

Total excess deaths (E) are 50-100 per cent of India’s all-cause mortality of 600 per month per million.

This is consistent with the above reasoning that deaths can run at 50-100 per cent above trend in hotspots during the spike phase. Column (F) suggests that our Covid death counting isn’t half bad.

To get to an upper bound for pan-India Covid death undercounting, I am going to make an outrageous extrapolation.

Attribute 100 per cent of the unexplained excess deaths of 200 per million per month to Covid and blindly extrapolate it across India’s 1.37 billion people for one whole month.

This gives us 2.7 lakh ‘undercounted’ deaths. Since this is over and above the over 1.2 lakh reported Covid deaths that month, this implies that true Covid deaths could be a maximum of 3x of reported deaths, even during the abnormal spike phase.

Above 3x is clearly an overestimate. It’s conceptually flawed to extend a spike localised in time and space to 1.37 billion people. India didn’t do as bad as a a few hotspots that media chose to highlight.

Many parts had a way milder spike phase (eg, North East, Odisha). States that were hit by Covid later were better prepared. While a precise estimate is impossible, my realistic estimate is that pan-India undercounting was no more than 2-2.5x during the spike phase.

This is still higher than the 1.5-2x ‘normal’ undercounting. But way lower than the baseless 5-10x claims pulled out of thin air and small minds.

Rajasthan’s Recent Data Confirms Our Inference Of Modest Wave-2 Undercounting

By way of independent validation of above inference, the Rajasthan government announced that their 2021 deaths (till 28 May) grew 5.3 per cent over 2020, comparing favourably to a 5.9 per cent rise in 2020.

This was in response to anecdotal allegations of undercounting. With reported Covid deaths accounting for around 60 per cent of the 5.3 per cent increase, there’s little undercounting.

Rajasthan’s test positivity ratio (good indicator of unreliable Covid stats) in wave-2 is at 15 per cent versus 13 per cent for India (and way worse than 5 per cent for Gujarat, accused of ‘massive’ undercounting).

Rajasthan also lags both India and Gujarat in tests per million people. If a state with below par testing shows little undercounting, how can India possibly do worse (unless you blasphemously accuse a Congress CM of fudging official death stats).

Rest Of Wave-2 Can Be Studied Similar To Wave-1

Assuming no surprises to the current trend, wave-2 is likely to end with a total of 25 million reported cases and three lakh reported deaths. Of these, 10 million cases and 1,2 lakh deaths occurred during the spike phase.

Wave-2’s cumulative test positivity ratio is likely to end up at 10 per cent (21 per cent in spike, 7 per cent otherwise). Since a 7 per cent TPR outside the spike phase is close to wave-1's 5 per cent, we can view it through a normal lens.

‘Normal’ undercounting of 1.5-2x is likely to prevail. So, 1.8 lakh reported deaths outside the spike phase could correspond to around three lakh actual deaths.

Combining spike-phase and the rest, we are looking at six lakh plus true deaths (a shade over 2x of reported deaths) across wave-2 and 700-800 million true infections.

Wave-2’s cumulative damage is likely to be three times that of wave-1.

Wave-2 Covid could contribute to less than 5 per cent of India’s all cause mortality. That 3x damage happened in a third of the time is the prime reason why wave-2 felt nightmarishly worse than wave-1.

Amidst gloom, I hope to have conveyed evidence-based positivity. Our death undercounting is way below the alarmist 5-10x estimates. Cumulative undercounting over wave-2 is only modestly worse than wave-1, mainly due to the steepness of the spike phase briefly overwhelming our systems.

With a majority of India now carrying antibodies, there’s hope wave-3 is both distant and milder. Having survived a hellish spike phase of wave-2, our systems will also be better prepared if (or when) it arrives. Increasing the vaccination pace will add to resilience. On that positive note, let’s get back to work.