key: cord-0926433-ktj5oeyu
authors: Kumar, C. K.; Balasubramanian, R.; Ongarello, S.; Carmona, S.; Laxminarayan, R.
title: SARS-CoV-2 Testing Strategies for Outbreak Mitigation in Vaccinated Populations
date: 2022-02-06
journal: nan
DOI: 10.1101/2022.02.04.22270483
sha: 5ed3a88f9c70d2bc9b52e57acbc997d2e520068d
doc_id: 926433
cord_uid: ktj5oeyu

Although COVID-19 vaccines are globally available, waning immunity and emerging vaccine-evasive variants of concern have hindered the international response as COVID-19 cases continue to rise. Mitigating COVID-19 requires testing to identify and isolate infectious individuals. We developed a stochastic compartmentalized model to simulate SARS-CoV-2 spread in the United States and India using Reverse Transcriptase Polymerase Chain Reaction (RT-PCR) assays, rapid antigen tests, and vaccinations. We detail the optimal testing frequency and coverage in the US and India to mitigate an emerging outbreak even in a vaccinated population: overall, maximizing frequency is more important, but high coverage remains necessary when there is sustained transmission. We show that a resource-limited vaccination strategy still requires high-frequency testing and is 16.50% more effective in India than the United States. Tailoring testing strategies to transmission settings can help effectively reduce cases more than if a uniform approach is employed without regard to differences in location.

Since the emergence of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) in 28

Wuhan, China in late 2019, the coronavirus disease (COVID-19) pandemic has resulted in more 29 than 280 million reported cases and 5.4 million reported deaths worldwide as of January 1, 30 2022 1 . Despite efforts to curb the spread of SARS-CoV-2 through restrictions on travel 2 , 31 business openings 3 , and personal measures 4 -including mask wearing and social distancing-32 cases have continued to rise in many countries 5 . Although vaccines against COVID-19 are 33 available, the emergence of variants of concern 6 that are only partially neutralized by existing 34 antibodies or prior vaccination 7 and are more contagious along with waning immunity 8 has 35 resulted in widespread COVID-19 outbreaks even in highly vaccinated populations 9 . Likewise, 36 many populations, particularly in low-and middle-income countries, still lack widespread access 37 to vaccines and continue to experience significant excess mortality 10 . All these factors must be 38 simultaneously considered when developing mitigation strategies for emerging outbreaks. 39

Consequently, it appears likely that SARS-CoV-2 will continue to pose a threat to public health 40 for many years even if vaccines are distributed widely because of the rapid evolution of variants 41 of concern. Thus, testing and containment will continue to be critical to COVID-19 response and 42 mitigation. 43

Because of the high transmission rate of SARS-CoV-2 and prevalence of asymptomatic 44 carriers 11 , accurate, efficient, and pervasive testing methods are needed to track and contain 45 disease spread. Currently, two main diagnostic methods are widely used 12 . Reverse Transcriptase 46

Polymerase Chain Reaction (RT-PCR), which is considered to be the gold standard, detects the 47 presence of viral RNA in respiratory samples 13 . Although highly sensitive, test results typically 48 require two to three days 14 , during which an infected individual may continue transmitting the 49 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. While we observe the same general trends in both settings, frequency is more important in 131 India than in the United States: Figs. 1 and 2 show a strong gradient of increasing cases as testing 132 frequency is decreased, most noticeably for antigen testing and especially for India. However, we 133 observe that at low coverage, the effect of frequency is much less in India. High-frequency, low-134 coverage testing can still be useful in the United States, but we did not observe the same pattern 135 in India, where coverage must be relatively high for effective mitigation. Although increasing 136 coverage when it was low had little benefit, increasing coverage from half of the population 137 surveilled to the whole population surveilled had a larger effect in India. Ultimately, though 138 frequency may still have dominated overall, increasing coverage was also critical in certain 139 testing scenarios in India. Additionally, we conducted sensitivity analyses by running canonical 140 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint 8 strategies with R eff from 2.5 (used in all main figures and analyses; Fig. S6, Fig. S7 ) and 141 decreasing R eff to R eff = 2 (Fig. S8 ) and 1.5 (Fig. S9 ). Our main findings and trends still held 142 across all values of R eff , though there are fewer cases with a lower R eff . 143

Emerging variants of concern are not only vaccine-evasive but often more transmissive. To 145 identify and isolate transmissive individuals as quickly as possible, we propose a mixed strategy 146 that utilizes the complementary nature of antigen tests and RT-PCR assays (i.e., antigen tests 147 excel when used frequently whereas RT-PCR assays are more effective when used widely). All 148 individuals were tested weekly using antigen tests and all negative results were followed up 149 immediately by an RT-PCR assay (Fig. 3) . Since RT-PCR assays have higher sensitivity, their 150 use as a follow-up should allow for the detection of individuals with a viral load too low to be 151 detected by an antigen test. Nevertheless, this was resource intensive: it potentially required 152 more than just one test per individual. 153 Such a mixed strategy was effective in both settings-resulting in minimal hospitalizations 154 and deaths-but more effective in reducing cases India than in the United States. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint 9 factors are not considered in our simulations as they are highly variable and differ between 164 transmission setting). Nevertheless, we show that effective surveillance is critical especially in 165 the early stages of transmission to mitigate disease burden from contagious variants of concern. 166 Likewise, the rapid emergence of variants of concern has placed an emphasis on widespread 167 vaccination campaigns and now even booster shots for some populations. Nevertheless, key 168 questions remain about optimal testing strategies and epidemic trajectories for populations as 169 vaccines are administered or additional immunity is conferred through booster shots. Thus, we 170 determined how proactive testing and simple vaccination strategies can be used together to 171 promote SARS-CoV-2 reaching the endemic phase. Disease surveillance, continued testing, and 172 vigilance remain critical even as more individuals become vaccinated. Note that while there are 173 studies detailing optimal vaccine allocation strategies 27 , our goal is not optimize the distribution 174 of vaccines but rather determine the impact of vaccines in a resource-limited distribution scheme 175 and how vaccines must be coupled with testing to mitigate disease spread. We couple daily 176 vaccinations with antigen testing 100% of the population weekly and antigen testing 33.3% of 177 the population weekly (i.e., see Fig 

Our results provide insight into constructing testing strategies with maximum effectiveness. 207

First, antigen tests are more effective than RT-PCR tests across both transmission settings 208 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint because they enable faster action to reduce transmission; our results agree with real-world 209 evidence that antigen tests have been used successfully in nation-wide testing campaigns 28,29 . We 210 observe that RT-PCR assays are more comparable to antigen tests as their turnaround time is 211 decreased 30 (Fig. S2) . Nevertheless, the increased mitigation of antigen tests compared with RT-212 PCR assays with standard turnaround times is most pronounced when 100% of the population is 213 tested weekly. Since antigen tests have a quicker turnaround time, infected individuals are more 214 likely to self-isolate faster. Our simulations show that use of antigen tests results in a lower peak 215 of daily cases compared with RT-PCR assays. Because disease spread is greatest in the early 216 stages, when most of the population is still susceptible, early isolation of infected individuals is 217 critical to mitigating disease spread 31 , especially critical when considering highly contagious and 218 vaccine-evasive variants of concern. Additionally, SARS-CoV-2 transmission to secondary 219 individuals is significant even immediately after initial infection 32 (Fig. S9) , further underscoring 220 the need for isolating infectious individuals quickly. 221

We show that that high-frequency testing must be prioritized when fighting an emerging 222 outbreak driven by a contagious variant of concern, though the relative importance of frequency 223 versus coverage differs by setting and the type of test used. Maximizing frequency has the 224 greatest importance for antigen testing. This is likely driven by its lower sensitivity but quicker 225 turnaround: since antigen tests are unable to detect infected individuals with low viral loads, they 226 must be used frequently to identify when individuals become infectious past detectable levels 227 and force them to isolation. On the other hand, RT-PCR assays can still be effective when used 228 widely because they can identify infectious individuals with low viral loads. Moreover, frequent 229 use of antigen tests is not substantially better than even more extensive disease mitigation 230 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint strategies, such as coupling antigen and RT-PCR tests, and is still effective in scenarios where 231 there may even be unmitigated disease spreaders. 232

Although we observe mostly similar trends in mitigation strategies between the two 234 countries, some differences are important for tailoring mitigation solutions. While our simulations overall indicate that high-frequency testing must be an urgent 246 priority, we also find that the importance of frequency and coverage differs by transmission 247 setting. Whereas increasing frequency is overall more important in India, increasing coverage 248 beyond half of the population surveilled at each testing occurrence is critical for markedly 249 improved mitigation and for the benefits of frequency to be most noticeable. Notably, at lower 250 coverages, increasing frequency is more beneficial than increasing coverage. Consequently, we 251 suggest that with limited resources, frequency should be prioritized unless coverage can be 252 increased beyond half of the population surveilled; likewise, at those high coverages, the 253 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint 13 importance of frequency is most evident. Ultimately, the need for widespread and frequent 254 antigen testing is urgent in both countries, but the trade-off between frequency and coverage 255 should be tailored to community needs. 256

Moreover, we find antigen testing is not only more effective but also substantially cheaper 257 than use of RT-PCR assays. Our simulations show that given a constant budget constraint, 258 antigen testing can be done more frequently or at wider coverage and result in fewer cases than 259 use of RT-PCR assays. Nevertheless, we also observe that the same testing scenario may have 260 different costs in the United States versus India. In our simulation, we assume that all individuals 261 who have not been infected must be tested. Since in the United States the peak in cases occurs 262 earlier, more individuals are infected in the early stages and thus a typical individual is removed 263 from the testing pool faster than in India. Although the cost of testing thus should be lower in the 264 United States and we do observe this in many of our simulations, in certain scenarios (e.g., where 265 antigen tests are used at high frequency and coverage; Fig. 2 ), the cost is less in India because the 266 testing strategy is less effective. Since in India more individuals are infected and do not need to 267 be tested, the cost of the strategy falls. However, given the differential nature of disease spread, 268 testing frequency and coverage can change as the epidemic progresses, which may also change 269 the cost (not considered in our simulations). Finally, we do not consider the cost of hospital beds 270 or self-isolation, which likely differ heavily between settings. Additionally, our analysis does not 271 explicitly consider contact tracing or self-isolation of individuals who experience symptoms, so 272 our results more directly indicate the impact of proactive testing and immediate quarantining. 273

Finally, we show that vaccines and testing can be combined to create mitigation strategies 274 that mitigate the duration of sustained transmission and can usher in an endemic phase earlier. 275

Even a resource-limited vaccine allocation strategy of simply distributing vaccines randomly to 276 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint susceptible individuals in addition to testing some of the population weekly is effective in 277 minimizing cases and ending sustained transmission earlier in both transmission settings (Fig. 4) . 278

However, we show that vaccinations have different impacts in each transmission setting. In 279 particular, the vaccination strategy utilized is more effective in reducing cases in India than the 280 US. This is likely due to the sustained nature of SARS-CoV-2 transmission observed naturally in 281 India (Fig. S1) ; consequently, the impact of continued vaccinations is greater in India as 282 supposed to the US where infections peak much earlier. Thus, our findings show that vaccines 283 are critical to minimizing the chance of future waves of COVID-19 cases especially as much of 284 the world's population still remains susceptible to SARS-CoV-2. Nevertheless, we note that 285 widespread testing is still critical, especially in the early phases of vaccine distribution when 286 vaccines are limited. Moreover, testing will likely continue to be critical as further variants of 287 concern that may be vaccine-resilient or even vaccine-resistant continue to emerge and must be 288 monitored to ensure that resurgences of SARS-CoV-2 infections do not occur 37 . 289

Generally, the most effective of the testing strategies discussed in this paper are frequently 290 not the most expensive but rather those that are most closely tailored to the dynamics of the 291 setting. Therefore, identifying transmission dynamics across a wide range of settings and 292 applying specialized testing scenarios to specific environments are critical to effective 293 mitigation. Our study suggests that contact matrices specific to the setting must be used as 294 opposed to generic contact matrices 38 commonly used in modeling studies. Our simulation shows 295 that social mixing patterns affect the efficacy of mitigation strategies. However, we acknowledge 296 that in developing tailored scenarios, considering social factors is also crucial, since health 297 behaviors have been shown to be related to social clustering 39-41 . Nevertheless, we believe the 298 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint trade-offs presented in this paper present a useful set of heuristics that can inform testing 299 strategies and health policy across a wide range of settings. 300

Accurately simulating SARS-CoV-2 dynamics requires two types of data: details of disease 302 outcomes (i.e., data regarding COVID-19's effects in humans), and attributes of virus spread 303 (i.e., properties of SARS-CoV-2 transmission). Details of disease outcomes include such effects 304 as whether an infection will result in hospitalization. Attributes of virus spread refer to the 305 disease's epidemiology, such as the Reff and viral shedding by day. We gathered both kinds of 306 data through freely and publicly available sources for both transmission settings. 307

We obtained data on COVID-19 cases and death counts from the beginning of data 309 acquisition in the United States and India: US data were obtained from the National Center for 310

Health Statistics (of the US Centers for Disease Control and Prevention, CDC) 26 , and data from 311 India were obtained from previous studies 18,42 . We assign each individual an age, gender, and 312 comorbidity based on census data and comorbidity prevalence 43 . We used these data to estimate 313 COVID-19 probability of death given age, gender, and comorbidities. Additionally, the CDC 314 keeps data on the probability of hospitalization and summary statistics (i.e., the 25 th , 50 th , and 315 75 th percentile) for time spent in the hospital due to severe COVID-19 infection 44 . Although 316 having the distribution of hospital stay duration would be ideal, based on empirical evidence that 317 such data follow a negative binomial distribution 45 , we construct negative binomial distributions 318 with the same summary statistics to generate an estimated probability for various lengths of 319 hospital stays. We had different hospitalization rates in the United States and India, but we 320 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint assume that the pattern of hospital stay length is maintained between the two countries, since 321 there is a shortage of data on COVID-19 hospitalizations in India. Ultimately, drawing from 322 empirical evidence of COVID-19 effects, we project the expected outcome of individuals' 323 infection given their age, gender, and comorbidities, including whether they are hospitalized, 324 time spent in hospital, and whether they die. 325

We considered two types of tests in our simulations: RT-PCR assays and antigen tests. The 327 sensitivity and specificity as a function of viral load of each test are given in Table S3 . From 328 current data, we used a turnaround time of three days for RT-PCR assays 14 . We assumed antigen 329 test results come back quickly enough that an infectious individual will not further spread the 330 virus while waiting for results. Finally, drawing from current estimates of costs for these tests, 331

we assumed a cost of each test (Table S3) increase the chance that a vaccinated but infected individual will not transmit to their contacts. 337

Although the above data determine the effects of COVID-19 for an individual, they do not 339 detail how SARS-CoV-2 spreads in a population. Thus, to simulate SARS-CoV-2 transmission, 340 we gathered data on the incubation period 46 , transmission probability since infection 32 , and 341 inferred viral load after symptoms 47 . Together, these variables detailed the necessary information 342 for SARS-CoV-2 transmission from an infected individual to secondary contacts by day and 343 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint whether the individual will be flagged as infectious by a test. Despite extensive data along with 344 quantifiable uncertainty for the incubation period, transmission probability since infection, and 345 viral load after symptom onset of SARS-CoV-2 (see Table S3 for the parameters used in our 346 analysis), quantifying viral load prior to symptom onset is more difficult and there is sometimes 347 contradictory evidence on the peak viral load 48,49 . Thus, we drew from previous viral kinetics 348 models to infer the viral load prior to symptom onset; note that our results are likely robust to 349 changes in viral load distribution as there is variability in our estimated viral loads by individual 350 (Fig. S9) . Specifically, we say that viral load peaks anywhere from day 0 to day 4 after symptom 351 onset 50 ; the peak is anywhere from 5 to 11 log10 virions per mL 47 , that log10 viral load increases 352 linearly from negative infinity on the day of infection to the aforementioned peak, and that log10 353 viral load decays from peak to the end of the individual's infection linearly with a slope drawn 354 from meta-analyses 47 . See Fig. S9 for our inferred viral load distributions. 355

Additionally, we gathered data on contact matrices and the distribution for the number of 356 secondary cases arising from an infected individual for each transmission setting 18,51 . These 357 variables were used for determining how many infections may arise from a single infected 358 individual and the likely age of the consequently infected individuals. Table S3 shows a 359 complete list of parameters compiled to simulate SARS-CoV-2 transmission in our model. From 360 these parameters, we simulated realistic disease spread in a population across settings. 361

We developed a stochastic, compartmentalized, empirically driven agent-based model 363 (ABM) to project COVID-19 cases, hospitalizations, and deaths given a variety of testing 364 strategies. We adopted the following structure in our model: individuals in the population start as 365 "susceptible" or "recovered" if they have previously been infected before entry into our 366 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted February 6, 2022. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint

An interactive web-based dashboard to track COVID-19 in 398 real time

Impact of international travel and border control measures on the global 400 spread of the novel 2019 coronavirus outbreak

Inferring the effectiveness of government interventions against COVID-403 19

Causal impact of masks, policies, behavior 405 on early Covid-19 pandemic in the US

Southeast Asia 407 is an emerging hotspot for COVID-19

SARS-CoV-2 variants of concern in the United 409

States-Challenges and opportunities

SARS-CoV-2 Omicron has extensive but incomplete escape of Pfizer 411

BNT162b2 elicited neutralization and requires ACE2 for infection

Waning immunity after the BNT162b2 vaccine in Israel

CoV-2 spread

Types of Assays 422 for SARS-CoV-2 Testing: A Review

Saliva as a non-424 invasive specimen for detection of SARS-CoV-2

Quest Diagnostics Media Statement about COVID-19 Testing

Clinical evaluation of three 427 sample-to-answer platforms for detection of SARS-CoV-2

Comparison of automated SARS-CoV-2 antigen test for COVID-19

infection with quantitative RT-PCR using 313 nasopharyngeal swabs, including from seven 430 serially followed patients

Dynamic interventions to control COVID-19 pandemic: a multivariate 432 prediction modelling study comparing 16 worldwide countries

Epidemiology and transmission dynamics of COVID-19 in two 435 Indian states

Poverty and access to health care in developing countries

Simulating preventative testing of 439

SARS-CoV-2 in schools: policy implications

Modeling between-population variation in COVID-19 dynamics in Hubei

The authors are grateful for the computational resources managed and supported by Princeton 524

simulation. Susceptible individuals can become "infected and not expressing symptoms" after a 367 positive transmission event with another infected individual. Individuals can either stay as 368 "infected and not expressing symptoms" (i.e., "asymptomatic") for the duration of their infection 369 or move to "infected and expressing symptoms". Infected and symptomatic individuals may 370 either recover or become hospitalized. Finally, hospitalized individuals may either recover or die 371 (see Table S3 for the probability and duration of each event). Throughout each phase, the 372 individual's probability of transmitting the virus changes (peaking near symptom onset), as does 373 the viral load (peaking shortly after symptom onset). We inferred the viral load before symptom 374 onset based on previous studies 30 and drew the viral load after symptom onset from meta-375 analyses 47 . Nevertheless, not all individuals will transmit the virus: in accordance with 376 "superspreading" 52 , we drew the number of positive contacts for infected individuals from a 377 negative binomial distribution, and whom they are likely to infect, from contact matrices. 378Individuals interacted homogeneously with each other in Brownian fashion in an open space 379 with dimensions tuned to ensure the R eff is 2.5 without any mitigation. ABMs present two 380 benefits over traditional deterministic compartmentalized models: (i) implementing individual 381 specificity is easier, and (ii) they are inherently stochastic and thus can provide credible ranges of 382 the epidemic trajectory given initial conditions. Each model was run in the following way: there 383 are 5,000 individuals with age and genders drawn from US and Indian census data, and 384 comorbidities drawn from recorded prevalences given age and gender in 2017. Note that these 385 parameters can be easily changed so that policymakers can determine which mitigation and 386 testing strategies are most effective for specific communities. We ran each model for 200 days 387 (until a steady state is reached), 200 times (i.e., independent replications), and present the 50th, 388 2.5th, 25th, 75th, and 97.5th percentiles (i.e., the "credible intervals") as summary statistics in 389 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)The copyright holder for this preprint this version posted February 6, 2022. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint 20 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 6, 2022. 

The authors declare no potential competing interests. 532

All data used in this paper is freely and publicly accessible through the US Center for Disease 534 Control or peer-reviewed studies. The provided supplement details data sources. 535

The code used in this paper is available upon request and is on GitHub. 537 538 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint 27 FIGURES 539 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)The copyright holder for this preprint this version posted February 6, 2022. ; https://doi.org/10.1101/2022.02.04.22270483 doi: medRxiv preprint