A Second Wave imminent in Mumbai ? Implications for Covid 19 from L Ward

By: Harshita Magroria, Sukhada Gole, Sonam Maske, Viraj Shende, Yashashree Rane¹

This article shows why we are very concerned about a large scale second wave in Mumbai non-slums triggered by the Festive Season celebrations in the next few weeks. We also offer a forward view of into 2021. See our videos in Hindi and English. We have written it to share with our fellow citizens, healthcare workers and leaders in our community our work since August 2020 as we completed our Fellowship with The Praja Foundation. We were guided by Directors of Praja and by The Covid 19 Localisation Modelling Group²

Summary

We are a group of University Students in Mumbai studying Chemistry, Economics, Finance, Political Science and Public Policy; simultaneously involved in the Elected Representative Fellowship programme at Praja Foundation. We are also concerned citizens of Mumbai. In August we set out build on the earlier work³ to understand: what is happening now and what may happen next in Mumbai with respect to COVID-19?

We found out these are not easy questions to answer.

Mumbai is a patchwork of very different living conditions, population density, access to medical treatment, general levels of health and wealth and vast differences in the means to cope with COVID-19 restrictions. The photo in Exhibit 1 shows a slum area in Mumbai right next to better off Middle Income residential neighbourhoods. This is typical of Mumbai with very different living conditions within very short distances.

Almost every Ward in Mumbai is unique and even within Wards conditions vary significantly. All of these differences have led to vastly different outcomes and have resulted in a need to analyse the strategy for each Ward almost street by street.

We chose to analyse L Ward in Mumbai. Exhibit 2 shows the location of L Ward at the Center of Mumbai and in terms of Impact of Covid 19 also in the Middle rank of Covid 19 Cases. L Ward is itself a Patchwork Tapestry and in some ways a kind of representative sample of Greater Mumbai. It has around 70% slum population living in some larger slums but also in smaller slums in many locations across L Ward. The remaining 30% of the residential population live in middle-income areas with good housing infrastructure and lower density of the population and industrial areas. And interwoven with these slum and non-slum areas are the main railway station, commercial districts, markets, public spaces and industrial areas.

By contrasting the residential slum and non-slum areas in our analysis, we have been able to foresee:

• The slum areas will most likely leave COVID-19 behind them in the near term. Reinfection risk remains but our analysis provides a lead indicator of the likely path. Because the Slums were impacted most severely back in April they are now further along and in some ways a bellwether what could happen next in the non-slums
• The non-slum areas might be at risk of a surge of infections 7x the level experienced in April. Reinfection risk remains but the analysis in the slum can help narrow the range.

The COVID-19 Localisation Model⁴ helped us to quantify the regional infection trajectory of the virus. The results are directional in nature and our modelling work using the Localisation Model and the data we have access to⁵ implies almost all slum dwellers could have been infected by the virus. If such is the case, it is highly unlikely anyone currently infected by COVID-19 will come into contact with a susceptible person in the slums and pass on the virus. By contrast the non-slums have many susceptible people remaining so there is a risk of a significant outbreak that can happen at any time because the slum and non-slum populations are inevitably mixing.

Our confidence about this conclusion is reinforced by a rigorous seroprevalence survey⁶ conducted in July and benchmarks against other similar surveys globally⁷. Exhibit 3 shows a summary of the survey result where there is significant variation but on average the slums measured 57% with antibodies compared to 16% in the non-slums. On average this means Covid 19 has spread 3–4x faster through the slums compared to the non-slums.

This is the story of how we learned to use the Covid 19 Localisation Model to arrive at these conclusions, how we built our confidence in the validity and how the modelling work has changed our thinking about how to use data, modelling and analysis to help manage the Covid 19 crisis.

In the process we have changed too: we now all have greater confidence in how we will benefit from learning these skills more generally and we all hope to continue to grow by working further with the model and coaching others. We are excited about this and optimistic that this work can make a real difference with Covid 19 and in our development as we move from learning at University to actively participating in the workforce.

Getting started: Finding, checking data and Setting up the Covid-19 Localisation Model

Our first step was to take the course offered by The Covid 19 Localisation Modelling Group (TC-19LMG) and then follow the steps in the course to set up the model for L Ward. That process revealed many differences in our assumptions as a group and pushed us to reconcile our understanding through systematic data collection and checking. By following that process and with some coaching from TC-19LMG we were able to set up the model to mimic the historical data on cases.

Exhibit 4 shows a screen capture from our L Ward model base case setup showing what the model shows has happened and a comparison to historical data. The logic of the model and all of our inputs mimics the data closely. The model uses this logic and requires explicit assumptions from us about key behavioural parameters to produce a forward view of what we would expect next if L Ward were a homogenous slum community:

The two graphs on the left compare the data (red coloured lines) with the simulation (the blue lines) of new daily cases and the accumulation of those new cases. The two middle graphs are the data on daily deaths and the accumulated deaths; the green lines show our assumption about how many deaths we believe happened including unreported. On the right is a graph showing the resulting number of people who have been infected (blue) and have been infected and recovered (green) versus the data from in the seroprevalence survey (red) uprated⁸ by 10% as a single spike in mid-July. The simulation runs for two years from 16 January 2020 so the red lines stop when the data runs out but the simulation keeps going because the logic that allows it to mimic the data in the past is driving the evolution of the COVID-19 pandemic going forward.

The performance of the simulation mimics all of the data with good accuracy historically. This historical comparison combined with many other tests of validity including a detailed investigation of root causes for significant changes in the speed of spread COVID-19 and extreme experiments to test the impact of uncertain data and parameters; it gives us good confidence that the underlying logic is accurate and therefore the forward view is a good guide to what may happen next.

A few comments about the history and observations about the forward view:

1. The most important check in all the graphs is the simulation passes through the uprated seroprevalence data point all the way at the right. This tells us that the speed of the spread in the model is in line with what actually has happened in slum areas of L Ward in Mumbai.
2. Next most important is the model is closely mimicking the daily and accumulated deaths. Why is this next? Because deaths are clear and tangible and backed up by a medical examiner investigation and certificate. It is possible some deaths may not be recorded but that is not something we can judge. We can only track the data we have.
3. Last in importance is the new cases. Why? Between 70–85% of COVID cases are asymptomatic: we can’t see them. But we can work backwards from Seroprevalence and deaths and the epidemiological evidence from across the world embedded in the Localisation model to make a very accurate estimate of how many people must have been infected. We have estimated using the model that the testing programmes finding new cases have discovered between 2%-4% of the actual infections. The new case data is a very inaccurate measure of what is really happening.
4. The resulting forward view is very sensitive to reinfection risk and this is currently a major uncertainty. Assuming low reinfection risk there is a low level of new infections in the slums, then a modest increase next year because our simulation has low reinfection risk embedded. Tracking the rate of reinfections is a critical early waning metric to quantify the most likely path forward.

The critical importance of Invisibles

The detailed look at the simulation explains why this is happening. Exhibit 5 is an annotated screen shot from the L Ward base case. It shows a more complete picture including many key ‘Invisibles’ and it fills in gaps in the data where we only have intermittent reporting: Notably Remaining Susceptible people at the left, total new cases/day and Total People Currently Infected in the middle and Total People Resistant at the Right. None of these are reported by any authority, but the simulation has these quantified due to the logic of the epidemiology embedded. These Invisibles explain what we believe is very likely happening in L Ward.

The key graphs are the Remaining Susceptible people and Total Currently Infected. If you look closely you can see a small pause in the number of Susceptible corresponding to early October (as we write this). This happens because to match the data we needed an assumption (corroborated by interviews) about the number of shielded people beginning to recirculate in early October; this also shows up as a pause in the decline of total new cases and an uptick in the total die/day in October. But the Remaining Susceptible people is about 20,000. Simultaneously, the Total Infected is about 60,000 out of a population of 900,000. So, the simple math if everyone mixing uniformly the probability of an infected individual meeting an uninfected individual is 20,000/900,000 = 2.2%. That may seem like a high percentage but when the outbreak started it was effectively 100%. Very clearly it is much harder for the virus to spread with few susceptible and the likelihood drops as infected people recover. So with only 60,000 people infected in early October the chances of a Second Wave in the slums are much lower than they were in April. The chance of encounter at any time during the day with an infected person is 60,000/900,000 = 6.6% x the number of encounters. The chance of one encounter resulting in an infection is 6.6% x 2.2% = 0.15%. that may seem small, but remember it because that is the benchmark for the next scenario. By contrast, in the non-slum areas there is a very different situation.

Exhibit 6 shows with the orange lines shows what to expect if Seroprevalence was only 16% in L Ward in July:

This is a hypothetical experiment because L Ward has only around 30% people living in non-slums but for the purposes of analysing the impact we have assumed everyone in L Ward is living in non-slum conditions and reset the pandemic progress to match the seroprevalence result and the available historical data. This experiment also assumes the same age demographics: (different age demographics would impact the death rates). With these assumptions the scale of cases is approximately 3x what we believe is likely to be happening in the non-slum areas of L Ward. (and because L Ward has 70% slums the blue line scale is exaggerated by 42%). The scenarios show the dynamics of the outbreak in the non-slum are radically different from the slum⁹. Once again, the key Graphs are the Remaining Susceptible people and Total Currently Infected. The blue lines corresponding to the slum are unchanged but in orange scenario corresponding to the non-slums we have slowed down the spread of COVID-19 to approximately 1/3 of the pace in the slums. We had to make corresponding changes to other parameters to match the historical data and these give an idea of how different the conditions in the non-slums are: both the % of infections discovered and the death rates must be. Exhibit 7 shows key parameter changes implied to match the historical data for all of L Ward. The results show clear differences in behaviour and response to the outbreak:

Why these differences? The pandemic is impacting people in the slums and non-slums differently and the demographics are different. An earlier reaction and later recovery of contacts/day makes logical sense: the non-slums have the resources to respond and sustain their vigilance and the implied sustained vigilance in the simulation is a type of substitute for shielding. The other adjustments are in principle straightforward: if you slow the infection down by 1/3 everything else has to be almost 3x more to discover the same infections and have the same deaths. This puts a much larger percentage of the population still at risk in early October. Remaining Susceptible is almost 600,000 and 40,000 infected. There are 24x more people at risk and a 4.4% chance of meeting an infection for any encounter during every day if everyone is mixing. So now the chance of an infection for any encounter is 600,000/900,000 x 40,000/900,000 = 3%.

Compared to our base case that is 3% divided by 0.15% = 20 x the risk of an outbreak in the non-slums versus the slums and the scale of the outbreak in the non-slums is likely to be 7x the levels of infections in May. This was a shocking result, we needed to check it and see to what extent uncertainties about our assumptions might impact the result. We found ways to help do that using the Localisation Model.

Localisation helps to prevent the ‘muddle’ of averages

We conducted one additional seroprevalence matching scenario experiment to see what would happen if we ‘transformed’ the model to the L Ward weighted average of 45% seroprevalence implied by 70% slum and 30% non-slum. The green lines in Exhibit 8 show a very important result:

Using the weighted average seroprevalence for L Ward almost entirely masks the radical differences in conditions and risks for the slum vs non-slum population.

There is almost no difference at all comparing the green 45% vs the blue 57% scenarios: Clearly the localisation that explicitly contrasts the situation for the slums vs. the non-slums is extremely important. Assuming the averages gives a very misleading understanding which could leave the leaders and healthcare services entirely unprepared for the scale and severity of what is likely to happen next.

Where does that leave us?

The situation in the slums and non-slums is very different, but one factor the model does not account for is they are living almost together and invariably mixing. Consequently both areas are susceptible to the virus. However they demand differing strategies and policies:

1. Unless reinfection risk is determined to be significant (we examine this later), in the slums the danger is mostly past although not eliminated. Many people have been shielded; caution is still important. The remainder of the population at risk has dropped to much lower levels and it perhaps is time to try to return to a normal life.
2. The residents of the non-slums have a different challenge but with high levels of resistance in the slums the risk of encountering an infection is much lower now than at the beginning of the outbreak. It is not straightforward what to do next and it is highly dependent on the answer to another question.

What is the impact of pre-existing resistance?

In many locations around the world certain groups have appeared to be naturally resistant to COVID-19. Much of the evidence is anecdotal but a few studies have been performed. For example, there was an outbreak of COVID-19 in the south of France during a festival¹⁰ and the measured infection rates amongst adults was 4 x the rate in children. In other studies, the impact of sustained BCG vaccinations¹¹ has shown a reduction of severe infections from COVID-19 by up to 50%. Other studies of Vitamin D sufficiency have shown a correlation with reduced infection rates. We don’t know in L Ward or Mumbai in general what levels of pre-existing resistance may exist but we can run tests with the Localisation model to ask the question: what if 5%, 10%, 15%, 20% or 25% of the population has some form of pre-existing resistance? This effectively excludes that proportion of the people from ever being infected and shifts them to the Resistant status from the outset. Exhibit 9 shows what happens if we implement this variation in the slums:

With varying levels of assumed pre-existing resistance in the slums, little difference in the path of the outbreak or the conclusions is found. In other words, even though it is possible there are a significant number of people who have natural resistance, the level of pre-existing resistance is not a factor for our recommended strategy unless, as we will show later, reinfection risk is significant.

However, the situation in the non-slums is different. Exhibit 10 shows a similar variation of pre-existing resistance with bigger steps of 10%, 20%, 30% 40% and 50%. At 10% pre-existing resistance, the story is much the same if a little less severe. We still have a very significant possibility of a second wave. But at 30% and above the second wave is significantly reduced and is much more like what happened in April and May at 40%. We do not know the level of pre-existing resistance prevalent in the non-slums. However, if there is pre-existing resistance that there will be a material impact on the recommended strategy.

What about reinfection risk?

So far our scenarios have assumed low reinfection risk. However there is evidence of re-infections in Mumbai city. Large scale research has been published by Imperial College in the UK suggesting 26.5% of antibodies are lost in 3 months¹². Loss of antibodies is not the same as loss of resistance because the memory T and B cells can restore antibodies rapidly for many infections. However it is genuinely uncertain how significant the impact will be. If we assume an exponential decay path for loss of antibodies then 65% of the population would have lost resistance over 1 year. Exhibit 8 shows the result of a series of experiments to check if re-infections can occur in large numbers after 90, 180, 270, 360 and 450 days:

If the Imperial College research is used as a guide then our 360 day or 450 day scenarios are the likely path we would expect in the slums: a lull in new infections to a low level over the next few months and then a modest resurgence in February or March of 2021. That would a conservative estimate based on the Imperial evidence unless subsequent research shows resistance loss accelerates faster than an exponential decay. The critical insight from this is it is very important to regularly measure reinfection rates in the slums: this is a powerful lead indicator of which loss of resistance path Mumbai is likely to follow.

We ran the same experiment for the non-slums and a very different picture emerges. Exhibit 9 shows the non-slums are lagging the slums in the timing of infections peak by about 6 months: consequently the next surge of infections would be delayed until September or October of 2021. The model also shows where and when to look for signals that quantify the pace of loss of resistance:

We have shown a range of possible outcomes for Covid 19 in L Ward. And it seems clear to us that the best strategy in L Ward is not homogenous. It must be tailored to the conditions in each part of L Ward. Also there are major uncertainties about the level of pre-existing resistance and reinfection risk. By comparing the model outputs to actual data in both slum and non-slum areas the model can help deliver better estimates of the level of pre-existing resistance and the rate of loss of resistance. Improved estimates can help with better preparation and planning.

This is not just a lesson for L Ward: it applies to every neighbourhood in Mumbai and other similar cities in India and most likely across the world: the strategy in each area must be linked to the living conditions and insights from the cases data and subsequent seroprevalence surveys.

Summary and Conclusions:

There are two important themes we believe are important to take away from our work Localising L Ward:

1. We have revealed distinctive strategic insights and identified specific actions that could help better manage the pandemic for locations with very different living conditions
2. In a short time we have learned an entirely new approach linking rigorous Data Analytics and Modelling to Policy Recommendations.

What are the Strategic insights and implied Actions?

Data analysis allows the authorities to predict negative consequences of subject-matter under study and take requisite measures to prevent them by formulating due policies. The localised model is beneficial to calculate the early arrivals of infections in any region it is brought into use and acquaints us to recognise the vital role of seroprevalence survey in reflecting actual infection trajectory. One big value of setting up the model for a location is to help you ‘see’ invisible stuff and give you a better chance to develop a robust strategy.

Considering such results, concentrated local policies can be tailored to avert mis happenings. Following are suggestions for the authorities in L Ward in Mumbai:

1. Developing distinct strategies for slums and non-slums as they have different characteristics;
2. Promote synergy between the elected representatives of the urban local bodies and ward level medical officials;
3. Conduct health protocol awareness drives with help of community leaders;
4. Incentivise adherence to hygiene measures by positive nudge or penalties;
5. Enhance mechanism for door-to-door surveillance, contact tracing and isolation
6. Rigorously record reinfections in the slums vs non slums
7. Use the model to make better estimates of pre-existing resistance and loss of resistance

Further, the insights from L Ward, Mumbai can be applied to the metropolitan cities of Kolkata and Delhi due to their close resemblance in terms of demographic composition, standard of living, structure of inhabitation and pressure on existing resources. By examining the model, we understand that the prospects of a second way are limited in the slums of Mumbai and are relatively higher in the non-slum areas. Significant slum population can result in an increase of the infections reported due to inadequate infrastructure and poor adherence to the health protocols as evident in L Ward Mumbai. Similar has been noticed in Delhi and Kolkata. The population of Mumbai is about 2 crores similar to that of Kolkata (1.4 crores) and Delhi (2.9 crores). The co-existence of the slum and non-slum population is the common characteristic amongst these cities thus making the results of the model applicable to them. It is pivotal for the administration to undertake seroprevalence surveys for slums and non-slums areas so as to strategise area specific policies to combat the virus. In the slum area higher infections were reported in the first wave of the virus thus chances of a second wave are averted due to presence of antibodies in a significant count of population. The non-slum areas are at risk of witnessing a second wave firstly, because of low infection rate in the first wave, reopening of the economy and relatively lower count of antibodies.

The second wave of Covid 19 virus can be prevented in the non-slum areas of densely populated cosmopolitans in India by strict adherence to health protocols. The authorities should intensify testing and tracing to curtail the spread of the virus. Simultaneously, there is an urgent need to build health infrastructure and ensure smooth accessibility to medical facilities.

What have we learned?

As a group we started this journey knowing only what we could read or see on the News about the pandemic and we all had different views about what was really happening and what that meant we should do. Now we have experienced a process using the Covid 19 Localisation model and the course that shows how to use it and we have each learned something different. When asked what are biggest insights for each of us, here are the observations we made:

Harshita Magroria: “Throughout the process of building this model we learnt analysis, fact checking and examining situations from every perspective. These learnings are comprehensive in nature and can be applied to tackle diverse social issues in India.”

Sukhada Gole: “The entire process of modelling was enriching where I learnt to observe things and understood the process of data collection. Now, I more of believe in facts rather than abruptly made assumptions.”

Sonam Maske: “We were involved in positive discussion with our TC-19LMG coach, which indeed was a learning experience. I initially thought language would be a barrier although crossing such boundaries I learnt to build the model. I see myself different from the crowd because I understand the COVID-19 virus a little better.”

Viraj Shende: “The COVID-19 modelling helped me understand with the progress in technology it becomes necessary to adapt, learn and implement such modelling methods to deal with other issues such as the climate change.”

Yashashree Rane: “With the localised modelling I understood the way to reach at the roots of various problems and come forward with realistic solutions.”

This experience has transformed our general understanding of how to start with data, check its validity and accuracy and use the data to get insights from a sophisticated model of Covid 19 in any location. In the process we have become more confident advocates of actions we believe will make a positive difference and these actions are distinct for each area. Our learning will, we believe, not only help us make a difference where we live here in Mumbai; it will also help us understand and analyse other situations we may encounter when we move on from University and become active in the workforce.

Postscript

“We all have agreed to voluntarily assist you (TC-19LMG) further in the process of building the localised COVID-19 models. It shall be a one of a kind learning experience for us.”

Thanks to our Directors at Praja Foundation and Maurice Glucksman, Andre Nemec and Dr Kim Warren for their guidance and contributions to this article

Notes and References

1. Praja Foundation Elected Representatives Fellowship 2019–2020
2. Guidance. Praja Foundation: Priyanka Sharma, Pooja Verma, Milind Mhaske, and Yogesh Mishra; The COVID-19 Localisation Modelling Group: Maurice Glucksman, Andre Nemec, Dr Kim Warren
3. “Mumbai Covid 19 Localisation Update: Emigration has reduced new cases and prompted celebrations — It may be too soon to celebrate” Medium article
4. The Covid 19 Localisation Modelling Group. https://covid-19-localisation-modelling.thinkific.com/
5. Includes: https://stopcoronavirus.mcgm.gov.in/key-updates-trends; Hospital, L Ward interviews
6. “Technical details: SARS-CoV2 Serological Survey in Mumbai by NITI-BMC-TIFR”, https://doi.org/10.1101/2020.08.27.20182741
7. “The infection fatality rate of COVID-19 inferred from seroprevalence data”, https://doi.org/10.1101/2020.05.13.20101253
8. We uprated seroprevalence data by 10% (from 57% to 63% in the slums and from 16% to 18% in the non-slums) as an approximation to account for people who develop resistance after an infection but do not have measurable Covid-19 anti-bodies
9. When formulating strategy we re-scale the results to the levels expected in L Ward
10. “Clusters of COVID-19 associated with Purim celebration in the Jewish community in Marseille, France, March 2020”, https://doi.org/10.1016/j.ijid.2020.08.049
11. “Impact of Routine Infant BCG Vaccination on COVID-19”, https://doi.org/10.1016/j.jinf.2020.08.013
12. “Declining prevalence of antibody positivity to SARS-CoV-2: a community study of 365,000 adults”, Imperial College Oct 2020. https://doi.org/10.1101/2020.10.26.20219725