key: cord-0489127-pggyh2tm authors: Juneja, Sandeep; Mittal, Daksh title: Modelling the Second Covid-19 Wave in Mumbai date: 2021-05-05 journal: nan DOI: nan sha: 7f3d563cb4067b47688a9693cac5cec64fb2860d doc_id: 489127 cord_uid: pggyh2tm India has been hit by a huge second wave of Covid-19 that started in mid-February 2021. Mumbai was amongst the first cities to see the increase. In this report, we use our agent based simulator to computationally study the second wave in Mumbai. We build upon our earlier analysis, where projections were made from November 2020 onwards. We use our simulator to conduct an extensive scenario analysis - we play out many plausible scenarios through varying economic activity, reinfection levels, population compliance, infectiveness, prevalence and lethality of the possible variant strains, and infection spread via local trains to arrive at those that may better explain the second wave fatality numbers. We observe and highlight that timings of peak and valley of the fatalities in the second wave are robust to many plausible scenarios, suggesting that they are likely to be accurate projections for Mumbai. During the second wave, the observed fatalities were low in February and mid-March and saw a phase change or a steep increase in the growth rate after around late March. We conduct extensive experiments to replicate this observed sharp convexity. This is not an easy phenomena to replicate, and we find that explanations such as increased laxity in the population, increased reinfections, increased intensity of infections in Mumbai transportation, increased lethality in the virus, or a combination amongst them, generally do a poor job of matching this pattern. We find that the most likely explanation is presence of small amount of extremely infective variant on February 1 that grows rapidly thereafter and becomes a dominant strain by Mid-March. From a prescriptive view, this points to an urgent need for extensive and continuous genome sequencing to establish existence and prevalence of different virus strains in Mumbai and in India, as they evolve over time. specific numbers may be far off from the truth. However, the observation that Mumbai had highly infective variants that grew to have substantial presence in March is likely to be true. From a prescriptive view, this points to an urgent need for extensive and continuous genome sequencing to establish existence and prevalence of different virus strains in Mumbai and in India, as they evolve over time. This may help better manage infections that may result from new strains in future. Since, other larger districts of Maharashtra such as Pune, Thane, Nashik and Nagpur are closely synchronised with Mumbai in this second wave in terms of reported cases (see Fig 3) and other regions in the state typically lag Mumbai, this suggests that much of Maharashtra may see a similar trend to Mumbai with a small lag. As is well known, R 0 denotes the expected number of individuals a single randomly selected exposed person infects in a city where everyone else is susceptible. It captures the infectivity of the virus. Later in Section III we also discuss the value of R 0 for the virus strain present last year as per our model, and its value under more infectious variants. Given that early this year, none of the models anticipated the severe second wave that we see in India, the above claims should be accepted with caution. The key caveats that may affect our projection and to watch out for include a large proportion of population infected last year becoming susceptible to severe reinfections. The virus showing significant capacity to escape the vaccine induced immunity. Rise of new strains that may be more infectious and lethal. Of course, the city population needs to continue to guard against laxity in the distancing precautions. Furthermore, while we have conducted extensive time demanding simulations, due to time paucity, we have not been comprehensive. Nor have we been able to implement more sophisticated stochastic optimization techniques. We cannot entirely rule out scenarios where reinfections carefully combined with population laxity may explain most of the observed fatality curve, even though that appears unlikely. While the reported cases never exceeded 1,500 a day up till July 2020 due to low testing, they remained close to or below 2,500 between mid September and October during the minor wave that started in the wake of Ganpati Festival related crowding in late August 2020. In comparison, the reported cases have been over 8,000 for most of the days since April 1, 2021. Thus, the cases are dramatically higher this time around compared to any time in 2020, a fact that agrees with the all India trend. The Covid-19 reported fatalities in Mumbai (Fig 1) further light on the relationship between cases and fatalities in Mumbai. CFR is a popular measure of severity of disease. It equals observed number of fatalities divided by observed number of cases during a specified time window. This can be a poor measure when the cases are rising as fatalities typically lag cases by close to 3 weeks. As per [7] , the average time between a person getting exposed and the person not surviving is around 28 days. Also, a person typically develops symptoms around 5 to 6 days after getting exposed. It is then reasonable to assume that an exposed person reports positive around 8-9 days later. With these numbers in mind, in Fig 4, we also report the Shifted Case Fatality Rate (SCFR) measured over a thirty day window where the cases in the denominator are an average 20 days earlier than the fatalities in the numerator. Thirty days are chosen to be large enough to provide a stable estimator, while small enough to capture time varying trends. For Mumbai we see that while this number was over 3% in September 2020, it is currently around 1%. The figure also illustrates that SCFR may be a better measure of disease severity compared to CFR that behaves poorly due to the lag between cases and fatalities. The fact that SCFR is taking reasonably small values suggests that Mumbai has not seen many avoidable deaths. Apart from increased testing (see Fig 7) , key factor that explains the discrepancy in SCFR in September 2020 and April 2021 may be that the population segment getting exposed now is more likely to get tested compared to some months ago. Other possible reasons include improved medical care, presence of more infectious strain (s) that maybe less deadly, and presence of mild reinfections that get reported as cases but lead to reduced fatalities compared to the first time infections. As an interesting digression, in the appendix (section V), we show SCFR graphs for other districts in Maharashtra, for India, as well as for some of the states with the largest number of people infected in the second wave. The current SCFR of over 6% for Delhi and UP is particularly worrying. In addition to suggesting significant reporting issues, this also suggests presence of a more fatal strain (perhaps the UK Variant), and of many avoidable fatalities. The SCFR for India is around 2.5%, while interestingly, the CFR is less than 1%. Second Wave: An important aim of this report is to develop insights into the second wave in Mumbai through playing out various scenarios and see the ones that better explain the Mumbai fatality data from February to April 2021, with its peculiar late escalation. As in our previous reports, we avoid trying to match observed cases because they are much less predictable -they rely on administrations' testing policy that can vary with time and location. The observed cases also rely heavily on the segment of population getting tested. For instance, a slum dweller, for the same symptoms, is much less likely to get tested compared to someone living in a high rise. Before we conduct scenario analysis, lets understand the observed fatality numbers a little better: Recall that by late January 2021, the reported Covid-19 cases in India, in Maharashtra (see Fig 5) , and in Mumbai (see Fig 1) , were well below their peak from last year. Earlier Serosurveys in Mumbai [8] had shown a high prevalence of the disease especially in the dense slum areas, and it was expected that if the city fully opened by November, then by January it would approach some form of herd-immunity [3] . Serosurveys in other metropolitan cities also showed high prevalence, and even all of India was seen to have a significant proportion of population already exposed to the disease, see [9] and [10] . The general mood was that, given the relatively high prevalence, the young and largely rural population that may have some form of prior immunity, the pandemic was over in India. This reflected in the increased population mobility, increased social gatherings and increased laxness in observing social distancing. The Google Mobility Report [11] in Fig. 6 underscores this increase. Around late January, cases started to increase in Amravati and Nagpur in Maharashtra (see Fig 8) . The city of Mumbai substantially opened up on February 1 [12] , where the restrictions on the people travelling by local trains, the lifeline of the city in normal times, were reduced. Initially, from February 9 onwards, the cases in Mumbai increased at a slow rate (see Fig 1) , however, by early-March there appeared to be a phase change and the rate of growth increased substantially. Similarly, the observed fatalities were low in February and mid-March and saw a steep increase in the growth rate after around late March (see Fig 1) . As we discuss later in Section IV, projections from our October report [3] match the observed fatality graph quite well till December with some underestimation in January and February (with January opening in Figure 11 shifted by a month to roughly depict the Feb. opening). Recall that in Mumbai, restrictions on train travel was substantially relaxed on Feb. 1. The report was written under the assumption that the population compliance to social distancing remains unchanged through early 2021. With the adjustment for population laxity in December and January, the projected fatality data matches the observations reasonably well till early March. However, it completely misses the sharp rise in fatality numbers thereafter. In this report, we conduct an extensive scenario analysis to try and explain the much larger fatalities observed from mid-March to end April. We fix a base case that appears to match the fatality data well, and consider many scenarios that are perturbations to the base case. Our key conclusions -that the Mumbai second wave of fatalities will likely peak around the first week of May, and that the fatalities reduce to Jan. and Feb. levels by June 1 can be seen to roughly hold across all the scenarios that we consider. Mobility: As in the January opening scenario of [3] , we assume that the mobility of the Mumbai population is at 50% of the pre-covid times in the months from September to December, and extend this level to January as well. This increases to 65% on February 1 when the city further opened up [12] and stays at that level till the 'Break the Chain' lockdown on April 15, 2021 (see [13] ). To see that this approximation is in the correct ball park, we observed that the data on average passenger counts travelling daily by buses and local trains between January 13 to January 31, 2021 equals 4.7 million, while this average between February 1 to February 14, 2021 equals 5.8 million (source: MCGM). Given that in normal times travel by train is around 8 million passenger counts daily [14] , and by BEST buses, it is order 2 million [15], are selected mobility percentages appear reasonable. These percentages are also in rough consonance with the Google Mobility Report in Figure 6 . One could fine tune these numbers and also account for slight slowing down in late March and early April in the view of rising cases, but this is unlikely to impact the model fatality profile significantly. Maharashtra government imposed the 'Break the Chain' lockdown restrictions from April 15 to May 1, that were later extended to May 15 (see [16] ). Google Mobility Report suggests that mobility in late April was similar and slightly higher than in June 2020. We impose restrictions in our model similar to those in June 2020, with fraction travelling to work increased from 15% in June 2020 to 20% during the new lockdown. Compliance: In our October report, we had assumed that compliance levels was 60% in non-slums (NS) and 40% (S) in slums throughout the period (except for the festival periods where compliance is reduced to 40% NS and 20% S). The reduction in cases and fatalities likely led to increased laxity in the population in December and January. The increased mobility as seen in Google Mobility Report for retail and recreation, workplaces, transit stations, parks and groceries suggests that population mixing had increased and people likely became more lax in social distancing, wearing masks, etc. in December 2020 and January 2021. We account for this laxity by changing the compliance from 60% NS and 40 % S to to 50% NS and 30% S in December and 40% NS and 20% S in January. In February, till the 18 th, the compliance is assumed to be 20% NS and 10% S but from February 19 onwards, it is increased to 40% NS and 20% S till April 14. This increase in compliance was based on the order from MCGM for compulsory masks [17] . During the lockdown (April 15 -May 15), the compliance is assumed to be 60% NS and 40% S. Reinfection level: While some cases of reinfection have been reported in the literature, the phenomena does appear to be limited so far. Further, if the reinfections are mostly mild, they have little affect on the fatality numbers. We keep the reinfection level to zero in the base case. Other levels do not improve the fit to data. Later, we conduct extensive perturbations of reinfection levels to assess their impact. Variants: We assume that there exists a single variant that accounts for 2.5% of all the infected population on Feb. 1 in our model. These are randomly chosen amongst all the infected on Feb 1. Further we assume that this variant is 2.5 times more infected compared to the original strain. Technically, this means that the transmission rates in our simulation (see [2] ) corresponding to transmission at home, at workplaces, in communities and in local trains is increased by 2.5 when an infected person, infected by the new variant, encounters a susceptible one. This is the transmission rate assigned to all who are thereafter infected by the new variant. Vaccination: The elderly (above 60 years of age) were given the first vaccination dose in Mumbai from March 1 onwards. Vaccines were made available to 45 years and older from April 1. The vaccine starts to provide immunity many weeks later. In the month of March around 5 lac people were vaccinated. Since then the number is closer to 18 lac a month and is expected to become higher provided the supply continues. In our model, we assume that vaccine once administered is immediately effective. Further, to capture vaccine's efficacy of around 75%, we assume that 75% of the people vaccinated have complete protection, while randomly chosen 25% of the vaccinated get no protection. The specific vaccination schedule that we implement in our model is: • 40% (appx. 5 lakh) people above the age of 60 years are vaccinated in month of April (quarter of these numbers are vaccinated during the month at four equally spaced instances). • 15 lakh people above the age of 45 years (this includes 40% above age of 60) are vaccinated in month of May for the first time (again four times each month). • 20 lakh people above the age of 18 years are vaccinated for the first time in each month from May to July. ( four times each month). In our earlier report [18] we propose a methodology to model infection spread in local trains. The train transmission rate (β T ) based on some reasonable approximations corresponding to contact rates is set to 0.4 times β H , where β H denotes the transmission rate at home . We continue with this setting in the base case. In the graphs, the base scenario is compared with the following 4 scenarios: • Scenario without the infectious strain and without the lockdown. This is the contrafactual setting where the mild wave is caused in Mumbai only due to the opening of the economy on February 1, and increased laxity in population before that. • Scenario without vaccination and without lockdown. This helps us estimate the combined benefits of the lockdown and the vaccination effort. • Scenario without lockdown and with vaccination. This helps isolate the benefits of the lockdown. • Scenario with lockdown and without vaccination. This helps isolate the benefits of the vaccination effort. The following configurations match those in [3] and are common to all the figures hereafter unless the perturbation is specified: Workplace opening schedule is 5% attendance from May 18 to May 31st, 2020, 15% attendance in June, 25% in July, 33% in August, 50% from Some observations suggested from the base case: Figure 9 suggests that the lockdown substantially speeds up the return to normalcy. Under the blue curve one sees that daily fatalities return to Jan-Feb levels by June 1 due to the lockdown and vaccinations. The red curve corresponds to no lockdown and vaccinations and in this scenario the normalcy is delayed by a month. Importantly, vaccines make this normalcy lasting. Else, as the green curve shows, the fatality numbers would have started to increase from mid July onwards. Comparing the violet and the blue curve in Figure 10 suggests that the more infectious strain may have cost the city from 1,500 to 2,500 extra fatalities by September 2021. Figure 11 maps the projected daily infections in different scenarios. It suggests that even under the blue curve base case, we may see a few thousand new infections each day in September. Various sero-surveys suggest in Indian metropolitan cities roughly that 15-30 infections result in one reported case, so that we may still see a few hundred cases daily in September. Figure 12 illustrates the speed with which a highly infective variant can come to dominate once the economy starts to open. The scenarios around the base case that we consider include (these are summarized in Table 13 ): 1) Economy opening up: Figure 14 shows the fatality curve when the city opened up to level 75% on February 1 (red curve) instead of the base case (blue curve) where it opened at a level 65%. This leads to an increase in fatality levels from around March 1 onwards and somewhat earlier peak time for the fatalities. This illustrates the broader phenomena that increasing or reducing economic activity would lead to a shift in the fatality curve that does not help align the base curve with the steepness of the observed curve. A variable shift, where we first reduce the economic activity and later increase it may help in a better curve fit, but that does not match with our experience of the city situation. 2) Compliance: In Figure 15 , we show a lower compliance scenario (red curve) and compare it to the base scenario (blue curve). Again, the red curve leads to an increase from March 1 onwards in the fatality curve. Its not clear how to reduce or increase compliance in a realistic manner that would achieve the steepness of the observed fatality data. In Figure 16 we consider a more promising scenario where the compliance is lowered as before, however the variant virus infectiousness is reduced to 2 times from 2.5 times. This matches the data equally well except at the peak. Further lowering compliance levels however may be an unrealistic match to reality. Nonetheless, this example illustrates that infectiousness of the new variant of order 2 is also consistent with the data. Matching any curve too closely leads to over fitting and is not desirable. Further, peak fatality data around mid to end April may be high due to avoidable fatalities as Mumbai medical infrastructure was quite stretched around mid-April, so we do not look for a close match there. 3) Reinfection: In Figure 17 we consider the cases where reinfection is 5% (green curve) on February 1. Technically, this means that we convert randomly chosen 5% of the recovered population on February 1, and treat it thereafter as susceptible. Red curve captures the case where the reinfection is set to 10% on February 1. As Figure 17 shows, increase in reinfection as specified simply leads to higher fatality numbers that more-or-less increase linearly from March 1 with a high slope. In the orange curve we consider a case where reinfections are introduced gradually at 2.5% each month from February 1 to July 1. The curve again increases very steeply. While these are some ad-hoc numbers, its clear that reinfections would need to increase in a very specific In Figure 19 , the red curve corresponds to the case where the new strain is assigned 10% of the infected on Feb 1. The orange curve corresponds to 5% , the green curve corresponds to 1%, and blue curve with 2.5% denotes the base case. These curves suggest that values close to but smaller than 2.5% on Feb. 1 perhaps with slightly higher infectivity may match the observed data a little better than the base case. However, our broad conclusions are unchanged by this. In the base scenario we considered that the new strain was 2.5 times more infectious than the original strain. In Figure 20 , we compare this base scenario with the scenarios where new strain is 1.5 times (green curve), 2 times (orange curve) and 3 times (red curve) more infectious than the original strain. The July. Figure 22 suggests that even a moderately successful vaccination drive will help keep the fatality numbers low in August and September from the 'third wave' that may otherwise result from school opening. Again, as we noted earlier, the decision to open the schools on July 1 or later is best made closer to those times, when one has a better idea of the infections resulting from the opening. Variant virulence: Figure 23 shows the scenario where the new strain virulence or fatality rate is set to 1.3 times that of the original strain (red curve). This appears to better match the observed fatality data compared to the base case (blue curve) around the peak values in late April (when the effect of the new strain is more pronounced as it becomes dominant around late March). Again, since some of those high values may be due to avoidable deaths, it is difficult to claim from the experiments that the new strain maybe more virulent. This is also not suggested by the current Mumbai SCFR graph. Implementation: Technically, due to lack of medical data, the increased virulence is implemented in our model by increasing the probability of an individual transitioning from symptomatic state to hospitalised state, from hospitalised to critical state, from Recall that R 0 denotes the expected number of individuals a single randomly selected exposed person infects in a city where everyone else is susceptible. It captures the infectivity of the virus. All else being equal, a higher R 0 implies a more infective virus. Below, in Figure 26 , we report the R 0 for our model when it is fitted to the fatality data last year, the base case. See [2] for details of how our model was fit to data. We also report the R 0 from a variant that in our model is two times or two and half times more infective in terms of the transmission rates compared to the base case. The R 0 for overall city corresponds to the case where the exposed individual is randomly selected from across the whole city. We also report R 0 when the exposed individual is randomly selected from a non-slum area as well as from a slum area. It is higher in the latter case, because in a more dense setting, an individual is likely to interact with more people and infect more of them. Since the current infections in Mumbai are largely in non-slums, the increase in R 0 from non-slums is a better measure of the impact of more infective variants to the city. Figure 26 suggests roughly that in the non slum areas the R 0 has increased from around 2 to 2.5 for the base case to over 3 under the variants with transmission rates increased by a factor from 2 to 2.5. Relative infectiousness w.r.t. base case R 0 for non slums R 0 for overall city R 0 (inferred) for slums Below we recall our simulation dynamics. These help illustrate the R 0 estimation procedure as well as the methodology to incorporate more infective variants in our simulation model. Recall that our simulation model works as follows (see [2] ): • At a well chosen start time for our simulation, a fixed number of exposed individuals are seeded in the city where every one else is susceptible. • The simulation proceeds iteratively over time, incrementing it by ∆t at each time step. In our simulation ∆t corresponds to 1/4 of a day, or six hours. At each time t, for every susceptible individual n, its infection rate λ n (t) is the sum of infection rates coming from all the infected individuals in his interaction spaces including home (h), workplace (w), school (s), community spaces (c) and transport (T ). There are other categories in our model and they may be similarly handled. Thus, if λ n ,a n (t) denotes the rate at which individual n in the interaction space a ∈ (h, w, s, c, T ) infects individual n at time t (this would be zero if n is not infective or not interacting with n), we have λ n (t) = n ,a∈(h,w,s,c,T ) λ n ,a n (t). • Then at time t + ∆t, each susceptible individual moves to the exposed state with probability 1 − exp (−λ n (t)∆t), independently of all other events. With the remaining probability it continues to be susceptible. • Individuals once exposed, follow a disease progression probabilistic dynamics as specified in [2] . Some exposed become asymptomatic or symptomatic and they may infect others in the coming time periods. Asymptomatics recover after a short duration. Symptomatics may either recover or a small age dependent fraction may be hospitalised. Hospitalised may recover or a small age dependent fraction may become critical. Critical cases may recover or a small age dependent fraction may pass away. • Time is then incremented to t + ∆t and the condition of individuals are updated. The simulation iteratively continues till some large specified terminal time. Above, if a susceptible individual becomes infected at time t, there remains an issue of identifying the individual who infected this individual. In our algorithm, this blame is assigned uniquely to one individual n . And this assignment happens with probability a∈(h,w,s,c,T ) λ n ,a n (t)/λ n (t). This appears to be a fair blame allocation that can be shown to probabilistically asymptotically valid as ∆t → 0. Estimating R 0 : At day zero, a randomly selected individual is marked as exposed to the disease, while all others are marked as susceptible. The selected individual follows the disease progression dynamics. When in an infective state, he may infect others. We count the total number of individuals infected by the selected individual until he is no longer infective. The above specified allocation rule helps in arriving at this number uniquely. This is a random quantity. This experiment is repeated independently many times to arrive at an estimator for Then, this person is assigned a new strain with probability λ new n (t) λ n (t) . and old strain with the remaining probability. Else, the algorithm proceeds as before. IV. OCTOBER REPORT RESULTS Figure 27 reproduces the fatality projections from our October Report [3] where the scenario of the economy opening up on November 1, green curve, and the economy opening up on January 1, orange curve, are considered. Figure 28 shows the same curve with the observed fatality numbers updated till February 1. Since opening up on January 1 does not impact the fatalities till the end of January, the projections are valid till that time although the major opening up in Mumbai happened on February 1. While the projections hold quite well till mid-December, Figure 28 shows that they underestimate thereafter. As the violet curve in Figure 9 shows, this is significantly corrected once we assume increased laxity in population in December and January. In fact, that does a pretty good job of explaining fatalities right up to the first week of March. Of course, as we mentioned earlier, the subsequent increase in cases and fatalities were difficult to explain simply by assuming laxity in the population, and are best explained by assuming highly infective variants. The contact rates used for experiments (see Fig 29) are same as in our earlier report [3] . Class 9 · β school 4.83507 Neighbourhood 9 · β community 0.19368 Close friends 9 · β community 0.19368 Figure 29 : Interaction spaces, subnetworks and contact rates Following are the SCFR graphs for India (Fig 30) , for other districts in Maharashtra ( Fig 31) as well as for some of the states (Fig 32, 33) with the largest number of people infected in the second wave. We refer the reader to our October Report [3] , Page 5 for additional caveats associated with this report. Explained: B.1.617 variant and the Covid-19 surge in India City-scale agent-based simulators for the study of non-pharmaceutical interventions in the context of the covid-19 epidemic Covid-19 epidemic in Mumbai: Projections, full economic opening, and containment zones versus contact tracing and testing: An update Projections for fatalities in second Covid-19 wave for Mumbai Projections for cases in second Covid-19 wave for Mumbai (reported on April 15, 2021) Reopening Schools After COVID-19 Closures; THE LANCET COVID-19 COMMISSION INDIA TASK FORCE Estimates of the severity of coronavirus disease 2019: a model-based analysis Seroprevalence of SARS-CoV-2 in slums and non-slums of Mumbai Delhi's 5th sero survey: Over 56% people have antibodies against Covid-19 21 per cent seroprevalence across India, shows survey, ICMR says prevention key Covid-19 Google Mobility Report Mumbai's lifeline back on track! local trains to resume services w Break The Chain Order BEST's daily ridership climbs to 23 lakh, back to 2018's average Break The Chain Order BMC issues fresh guidelines as cases rise in Mumbai COVID-19 Epidemic Study II: Phased emergence from the lockdown in Mumbai Pulse survey on continuity of essential health services during the COVID-19 pandemic: interim report We thank our colleagues Prahladh Harsha, Ramprasad Saptharishi and Piyush Srivastava for many useful suggestions that helped our analysis. We thank them as well as our IISc collaborators R. Sundaresan, P. Patil, N. Rathod, A. Sarath, S. Sriram, and N. Vaidhiyan for their tireless efforts in developing the IISc-TIFR Simulation model [2] and their key role in our earliers report on Mumbai. We thank IDFC Institute for sponsoring Daksh Mittal's work with the TIFR COVID-19 City-Scale Simulation Team.We thank Mrs. Ashwini Bhide, AMC, MCGM for her insights and for her crucial data inputs. We also thank Shri Saurabh Vijay, Secretary, Higher & Technical Education Department, We acknowledge the support of A.T.E. Chandra Foundation for this research. We further acknowledge the support of the Department of Atomic Energy, Government of India, to TIFR under project no. 12-R&D-TFR-5.01-0500.