Increased Intensity Of PCR Testing Reduced COVID-19 Transmission Within Countries During The First Pandemic Wave

No medication has proven effective in slowing transmission of the novel coronavirus (SARS-CoV-2), which causes coronavirus disease 2019 (COVID-19). Lacking a vaccine, governments have relied on non-pharmaceutical interventions (NPIs), including physical distancing, travel restrictions, and hygiene measures. After China sequenced and shared the viral genome, detection of active infection using reverse transcription–polymerase chain reaction (PCR) testing has been part of the response, extending beyond clinical care and epidemiological tracking. Country experience and simulations show that testing, tracing and isolation can reduce transmission,^1,2 since substantial asymptomatic transmission occurs in COVID-19. The World Health Organization (WHO) has urged countries to “test, test and test,”³ and has suggested a rate of 10 negative to 1 positive test result as an indicator of adequacy.⁴

Despite this, decision-makers disagree on what constitutes adequate testing.^5,6 The legacy of pandemic influenza planning, which focused on reducing morbidity and mortality and never envisaged testing for controlling spread, may contribute to this disagreement, and most research focuses on other NPIs.^7–10 In a PubMed search, we found only 30 quantitative analyses of the impact of testing: almost all involved modeling and simulation, and none quantified real-world impacts.¹¹ Research problems include: difficulties of isolating impacts when multiple NPIs are simultaneously implemented; increases in per capita testing rates with cases which makes them a poor indicator of testing strategy; lack of a global testing database; failure to control for confounding factors; and unrepresentative geographical samples.¹²

Our study addressed these research problems by compiling data on numerous factors to quantify the association of PCR testing with COVID-19 spread during the initial pandemic wave—when some countries, such as China and New Zealand, achieved near-elimination—using a study design that robustly managed data gaps to maximize sample size and covariates.

Study Data And Methods

To model impacts on COVID-19 transmission, we adapted methods from previous epidemiological studies that assessed impact of interventions, such as school closures, on COVID-19 and other respiratory viruses.^13,14 Specifically, we estimated a cross-sectional, linear regression model of transmission intensity against the average intensities of interventions and other factors.

The online eAppendices provide a full description and justification of all our data sources and methods.¹⁵

Data

Our unit of observation was all territories with daily cases reported in the COVID-19 Data Repository published by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University (JHU).¹⁶ We included subnational jurisdictions with substantial autonomy over health policies and borders, such as Macau, and territories, such as the US state of Hawaii, which are not part of the contiguous United States. We defined the observation period for each to start from the date of peak incidence in March, typically March 28–31, and ending June 15, 2020.

We obtained data on daily PCR test numbers from multiple online sources, including the Our World in Data (OWID) data repository,¹⁷ other online collections, official communications and news reports. We enumerated tests done instead of persons tested, since few countries report persons tested. We interpolated data gaps and corrected inconsistencies, giving preference to official sources and OWID, and we excluded data, such as from Venezuela, which substantially included antibody tests, as these do little to stop transmission.

We obtained daily data on eleven categorical COVID-19 policy indicators from the Oxford COVID-19 Government Response Tracker (OxCGRT).¹⁸ These tracked containment and closure interventions, and public information, testing, and contact tracing policies.

To track changes in individual movement, which may be voluntary or in response to government NPIs, such as lockdowns, we obtained mobility metrics from Google and Facebook,^19,20 derived from locational data sent by mobile devices. As these were highly collinear, we combined them into two composite measures that proxy population level mobility: (i) increases in time spent at home; and (ii) reductions in time spent in non-residential locations (eAppendix 1).¹⁵

We found and collated data on other relevant factors, giving preferences to datasets with wide coverage, from UN or other official agencies, and from research groups whose sources were well documented. We filled many remaining gaps by search of online sources, including government sites and news media. eAppendix 1 describes all variables and sources.¹⁵

For some interventions and factors, we obtained daily data. These included national school closures, the percentage of people wearing face masks, face mask mandates, temperature, specific humidity and relative humidity. For other factors, we obtained estimates for the most recent available year. These included population exposure to air pollution by particulate matter (PM_2.5—particulate matter with diameter less than 2.5 micrometers), physical inactivity, geographical latitude, distances to the major pandemic epicenters during March (Wuhan, Korea, Italy, and Iran), having a policy for universal Bacillus Calmette-Guérin (BCG) vaccination for tuberculosis (TB), TB incidence, the number of SARS cases to proxy SARS experience, an index of democratic development, and two measures of health systems capacity to manage infectious disease threats that have been developed by defense analysts: the Infectious Disease Vulnerability (IDVI)²¹ and the Global Health Security (GHSI)²² indices.

As general controls, we sourced a range of socioeconomic and health systems measures from the World Bank’s World Development Indicators (WDI)²³ and other sources, including Gross Domestic Product (GDP) per capita, life expectancy, hospital beds and poverty rates. As initial high rates might constrain countries from increasing testing faster than incidence, we added as controls the peak March incidence rate and the date on which the fiftieth COVID-19 case was reported.

Outcome Of Interest

The time varying effective reproduction number ($R_{eff}$) quantifies the transmissibility of a virus at any given time. It represents the average number of secondary infections generated by one infected person. In the absence of interventions or behavioral changes, the SARS-CoV-2 virus is highly infectious with a reproduction number of 2–4, indicating that each case will typically infect that many other people.^1,2 Interventions that slow transmission reduce the reproduction number: if it falls below 1.0, incidence declines, and eventually the virus will disappear. For our analysis, we consider the impact of interventions on the average level of the reproduction number during the study period. This is the average transmissibility of the virus, which is independent of the average number of cases.

Focusing on average transmissibility has two advantages. First, alternatives used in other studies, such as cumulative cases or incidence rates, are compatible with different averages of transmissibility, since they also depend on how transmissibility varies over time. For example, a country that employs an intervention for one month that reduces transmissibility below one and then abandons it in the next month will experience a U-shaped outbreak with fewer cumulative cases and a lower average incidence rate than another country that starts with the same initial incidence rate but employs the same intervention only during the second month, leading to an inverted-U shaped outbreak. From our perspective, the intervention had equal effectiveness in both, but this will only be evident from the average reproduction number, which will be identical, and not from the raw case or incidence numbers. Second, within a linear regression model when the reproduction number is logged, the covariate coefficients can be directly interpreted as their percentage effect on the reproduction number. This is more meaningful from the policy perspective as the key epidemiological goal in controlling an epidemic is to find a mix of interventions that reduce the reproduction number below one to achieve control.

We estimated the average reproduction number in each territory by adapting an approach¹⁴ that approximates it by assuming that spread is exponential and that negligible numbers of people have immunity²⁴ (eAppendix 2).¹⁵ Specifically, we compute the constant reproduction number that would have been required to change the incidence rate at the start in March to that on June 15. By assuming that the increase in incidence is exponential, we can obtain the average daily percentage increase in incidence. From this, we derive the average transmissibility or reproduction number by making an assumption about the average number of days it takes one person to infect the next (the generation interval). Incidence rates were derived from the center-weighted, 7-day moving average of new cases, and we followed the EpiForecast group’s assumption of a generation interval of 3.6 days.²⁵ External validation of our method confirmed no bias and close consistency with EpiForecast estimates.

Interventions And Other Covariates

We quantified PCR testing intensity using the test-to-case ratio (TCR), defined as the ratio of tests to new cases reported. This controls for increases in test numbers with cases and increases in detected cases with testing, but it is also inflated by multiple tests on the same person. It aligns with the WHO benchmark⁴ and is comparable to others using test positivity rates, such as those adopted by Trump administration.²⁶ The TCR proxies overall intensity, which will depend on policies for testing international arrivals, contact tracing and testing of contacts, repeat testing in detection and clinical care, and symptomatic thresholds for community testing, but it does not reflect inefficiencies, such as reporting delays, which reduce impact on transmission. In the absence of global data on these details, the TCR was the most comparable metric.

We quantified exposure variables, which include interventions, ecological factors and other controls, by their mean daily value during an exposure period that lagged the observation period by seven days. This was based on estimates of an incubation period of 2–12 days²⁷ and a delay between symptom onset and case reporting of 2–7 days.²⁸

Analysis And Models

Many variables had missing values: the frequency ranging from zero for most general indicators and 5% for TCR to 34% for Google mobility variables. Half the territories lacked data for at least one variable. As values were not missing completely at random (MCAR), analyzing only territories with complete data would result in a small sample, biased results, and reduced precision. To overcome this, we used multiple imputation to impute missing values, estimating all models with 300 multiply imputed data sets. We also log transformed TCR, per capita GDP, the peak March incidence rate and other variables to ensure normality as desirable in multiple imputation. For full details refer to eAppendix 3.¹⁵

Out of 221 territories in our dataset, we included in analysis only 173 that reported more than 100 COVID-19 cases (for listing, refer to eAppendix 1),¹⁵ as small case numbers make incidence estimates unreliable. These accounted for 98% of the world’s population and 99% of COVID-19 cases.

We pre-defined two lists of covariates: (A) exposures to be retained on the basis of theory and prior evidence; and (B) exposures without strong evidence. List A comprised the TCR, mobility changes, school closures and face-mask usage. List B consisted of BCG policy, TB incidence, latitude, temperature, specific humidity, the IDVI and GHSI indices, and per capita GDP and life expectancy as general controls. Other covariates formed a third category C. We included the OxCGRT policy indicators in C, as they were highly correlated with the mobility measures, which reflect actual behavior, and added little explanatory power.

We evaluated covariates for model inclusion using stepwise backward selection. In Model 1, we retained lists A and B, and eliminated others based on model fit (adjusted $R^2$), and coefficient meaningfulness and significance at univariate $p$ value <0.20. In Model 2, we prioritized parsimony to avoid overfitting and forced inclusion of the A list only.

The institutional review board of the Institute for Health Policy assessed the study as not requiring full ethics review, as all data were anonymized, aggregate and publicly available. Analyses were performed in Stata 14.2,²⁹ with some data processing done using R.³⁰ The statistical significance level was set at 5% and statistical tests were 2-tailed.

Limitations

Our analysis has several limitations. First, as an observational study, it cannot support causal inferences and relationships remain associations. Second, reported cases understate actual incidence, and testing increases detection. Although using ratios of incidence rates removes country differences, it cannot eliminate temporal changes, but since countries with higher TCRs increased these levels faster (see eAppendix 5),¹⁵ any bias is likely downwards. Additionally, our data cannot differentiate imported from local cases, and our estimates of transmission will be biased upwards when imported cases predominate. Third, our analysis cannot account for heterogeneity within large countries, such as temperature, which will reduce precision. Fourth, we cannot account for factors that our data do not capture, such as physical distancing or differences in isolation strategies, although our extensive covariates mitigate this. Fifth, the exposure lag may not adequately control for endogeneity arising from countries intensifying interventions in response to increased cases. Finally, time series analysis would be better, but was not possible owing to data limitations.

Study Results

Descriptive Information

During March to June, COVID-19 transmissibility fell globally with the median reproduction number declining from over 2.5 in early March to fluctuate above 1.0 during April–June (supplementary exhibit S1).¹⁵ The falls in transmissibility varied by region. The World Bank’s East Asia and Pacific (EAS) region, which groups East Asian and Pacific Island countries, Australia and New Zealand, was most successful in controlling COVID-19 spread, with the average reproduction number falling to 0.9, and several territories coming close to eliminating the virus. In contrast, in several regions and countries, such as South Asia and Brazil, the average reproduction number remained above 1.1 with no slowdown in spread.

Amongst interventions to control spread, testing intensity showed the largest variation, with the TCR varying from 1 to over 1,400 across countries (online exhibit 1).¹⁵ Greater testing intensity was also associated with lower incidence, cumulative cases and deaths (supplementary exhibits S5–6).¹⁵ Amongst other interventions (supplementary exhibits S2–8),¹⁵ there was considerable uniformity in the imposition of school closures, and little variation by income level in mobility reductions and lockdown measures. However, mask usage varied considerably across countries, the country average being greatest in Latin America, and lockdown measures and mobility reductions were least in East Asia and Pacific.

Impact Of Testing, Interventions And Other Factors

Our models regress COVID-19 transmissibility (represented by the natural logarithm of the average reproduction number) against the average intensity of the included interventions and factors. The estimated coefficients convey the size of impact of each factor, and by exponentiating these we obtain their percentage effect on the reproduction number. The models also provide us with 95% confidence intervals (CI) for the effect sizes.

Our final model (Model 2) fitted the data well, explaining 81% (adjusted $R^2=0.81$) of the variation in average COVID-19 transmissibility across countries (full estimates detailed in supplementary exhibit S10).¹⁵

Of all intervention measures, testing intensity was the most influential and was highly significant ($p<{10}^{–16}$). Its effect is logarithmic, so a ten-fold increase in the TCR would reduce the average reproduction number by 8.6% (95% CI 6.8–10.3), and a hundred-fold increase by 16.4% (95% CI 13.1–19.6) (computed from the Model 2 estimates reported in supplementary exhibit S10).¹⁵ Since TCR levels varied so much between countries, this translates into the largest relative impact of all intervention measures. This is shown in online exhibit 2¹⁵ which illustrates the relative impacts of key interventions and factors. In contrast to the effect estimates discussed below and given in supplementary exhibit S10,¹⁵ this chart uses the corresponding standardized coefficients, which allow direct comparison of their relative effects.

None of the other intervention measures were statistically significant ($p>0.05$), although school closures and face mask use were associated with reductions in transmissibility. Increased time spent at home was associated with increased transmissibility, although not statistically significant ($p=0.15$). This was expected, as epidemiologists assume this leads to more transmission within households.³¹ However, there was no reduction in transmissibility associated with reduced time spent in non-residential locations, implying that globally the mobility changes usually associated with lockdowns increased overall transmission, although none of these effects were statistically significant.

Several of the ecological factors were associated with substantial and statistically significant protective effects. A 1”C increase in temperature and a 1% increase in the share of the population that was elderly (65+ years) reduced transmissibility by 0.4% (95% CI 0.2–0.5; $p<{10}^{–4}$) and 0.7% (95% CI 0.3–1.0; $p<{10}^{–4}$) respectively. Increases in the share of the population that was aged 5–14 years were also protective, but this was not statistically significant ($p=0.15$). In contrast, air pollution, urbanization and poverty were associated with statistically significant increases in transmission. A 1 μg/m³ increase in fine particulate air pollution (PM_2.5), a 1% increase in the urban share of the population and a 1% increase in the share of the population (%) living below the five dollar international poverty line were associated with 0.1% (95% CI 0.0–0.2; $p=0.002$), 0.1% (95% CI 0.0–0.2; $p=0.003$) and 0.1% (95% CI 0.0–0.2; $p=0.02$) increases in transmissibility respectively.

In contrast, physical inactivity, specific and relative humidity, latitude, and the various measures of health systems capacity, such as health spending, hospital beds and the IDVI and GHSI indices were not associated with any statistically significant effects ($p>0.20$) during model building (results not shown). Our general controls of life expectancy and per capita GDP also exhibited no relationship with transmission (joint F-test: $p=0.70$). All were dropped from the final model.

Two results were unexpected. The March peak incidence rate was highly influential with a ten-fold increase associated with a 7.4% (95% CI 5.6–9.1; $p<{10}^{–6}$) reduction in transmissibility during the subsequent three months, and universal BCG vaccination was associated with a 4.2% increase (95% CI 1.0–7.7; $p=0.01$).

Additionally, increases in the days that public gatherings of more than 10 people were banned ($p=0.015$) and increased distance from the Wuhan epicenter ($p=0.02$) were significantly associated with increases in transmissibility. Being a US territory was also significantly associated with a 12.7% (95% CI 6.5–19.2; $p<{10}^{–4}$) increase in transmissibility compared to the rest of the world.

Simulations

We explored our results using counter-factual simulations (full details, eAppendix 5).¹⁵ In a hypothetical country with median levels of other interventions and characteristics, increasing testing intensity to TCR levels of 60 and above would have reduced the average reproduction number significantly below one (online exhibit 3).¹⁵ Simulations also showed that increased testing might have reduced the reproduction number close to or below one in many territories, including Peru, Chile and Indonesia, where lockdowns failed to achieve this (online exhibit 4).¹⁵ Other simulations indicated that better performance in East Asia and Pacific was driven primarily by greater testing, and that other regions might have done much better with a similar mix of testing, masks and mobility restrictions (supplementary exhibit S9).¹⁵

Robustness Checks

We undertook a range of robustness checks to assess possible limitations (eAppendix 4).¹⁵ These evaluated changes in parameter assumptions and samples, including the observation period, the exposure lag, and the case threshold for including territories. Throughout, our results remained robust, meaningful and statistically significant. We evaluated our choice of multiple imputation by re-estimating models using only complete cases, which indicated that a complete case analysis would have produced biased over-estimates of testing’s impact. As linear regression assumes linear relationships, we also re-estimated our results by specifying the logged TCR with a restricted cubic spline to allow it to have a non-linear relationship, which showed that our estimated effect holds across TCR levels of 1–1,000 (supplementary exhibit S7).¹⁵

We also investigated our failure to detect net benefits from mobility reductions associated with lockdown measures, but we were only able to detect net benefits, although not statistically significant, in the World Bank’s Europe and Central Asia (ECS) region. We also confirmed a beneficial impact in the group of eleven European countries where Flaxman and colleagues⁸ previously reported that lockdown reduced transmission, indicating that our study is picking up real geographical differences in lockdown impacts (eAppendix 4).¹⁵

Discussion

Unlike previous studies, our analysis explicitly quantifies testing for COVID-19. Its strengths are its comprehensive global sample, its accounting for both interventions and ecological factors, and extensive controls that minimize selection and omitted variable bias. These may explain why results differ from previous studies.

Although our finding of a strong association between testing intensity and transmissibility cannot prove causality, the robustness of the relationship across countries and TCR levels is consistent with a known mechanism. Around half or more of COVID-19 transmission is caused by people who are asymptomatic or who have only minor symptoms, so only increases in PCR testing make it possible to increase detection and isolation of infectious cases, and then to increase the numbers of their potentially infectious contacts who are isolated. This remains the only known approach that blocks person-to-person transmission sufficiently to stop the epidemic.¹

We lack systematic data on who actually is tested in different countries and what happens after testing, so we did not adjust for this. However, as online exhibit 5 illustrates,¹⁵ differences in testing rates between countries are associated with significant differences in how testing is targeted and acted on, and its likely impact on transmission. Where testing intensity was low, as in the USA (TCR=11) and UK (TCR=15), testing mostly diagnosed and isolated the most symptomatic cases, which has limited impact on transmission. Only with greater testing can countries screen wider circles of asymptomatic case contacts and lower the symptomatic threshold for testing individuals without obvious exposure. At the most intensive testing levels (TCR=100–1,500), countries were either actively encouraging anyone with respiratory symptoms or fever to get tested or routinely testing such patients, in addition to testing international arrivals and isolating and testing all case contacts. Such countries tolerated very low positivity rates in order to increase detection of cases in the community. To the extent that such differences explain differences in testing rates, they provide the link between higher testing rates and more effective control of COVID-19 transmission.

Our findings on the effects of temperature, air pollution (PM_2.5) and age structure confirm previous studies (see eAppendix 1).¹⁵ The role of air pollution may warrant more attention as it may explain higher transmission levels in countries such as India and Nepal, although we note that we use mostly 2018 pollution estimates that over-estimate 2020 levels in most countries. However, our inability to detect additional effects of humidity and latitude suggests that these do not have substantial effects independently of temperature.

Our results strongly reject previous findings that BCG vaccination is protective. We speculate that earlier studies failed to adequately control for confounding factors. Our finding that economic and health capacities have no relationship with transmission suggests that national resources are not usually a constraint, and instead it is the strategies that countries choose that matter. Surprisingly, the GHSI index, which claims to assess country capability to prevent and mitigate epidemics and pandemics, and the IDVI index, which assesses country vulnerability to transnational infectious disease outbreaks, exhibited no relationship with COVID-19 transmissibility, indicating that they are poor measures of country capacity and vulnerability and that other country characteristics need to be looked at.

The findings that locations closer to Wuhan and with worse COVID-19 outbreaks in March did significantly better in subsequent control of transmission are intriguing. We speculate that in places that were closer to or confronted the initial epidemic earlier, this engendered more fear and forceful reactions by governments and societies, which our data do not capture, than in places that had time to habituate to the threat.

The estimated effects of our two mobility measures are consistent with previous research in that they confirm an increase in transmission associated with time spent at home, but their overall effect, which increases transmission, except in Europe, was not. Although neither effect was statistically significant, this may explain an anomaly in current knowledge. Studies have found that lockdowns and mobility reductions slowed COVID-19 transmission in Europe and North America,^8–10 but no empirical analysis with adequate controls has demonstrated net benefits at the global level. Problematically, many countries elsewhere, such as India, Indonesia, Peru and Chile, failed to slow the epidemic with stringent lockdowns. Additionally, a recent analysis of OxCGRT data that evaluated the global impact of NPIs found that the evidence on impact of several lockdown-related NPIs was inconsistent and inconclusive.¹² We offer three linked explanations, noting that household transmission accounts for a substantial part of transmission in most countries. First, in developing regions, where personal living space is less, there may be larger increases in transmission at home during lockdowns. Second, outside developed regions, the necessities of subsistence may make it more difficult for people to remain in their homes. Third, home confinement was only effective in Wuhan in achieving epidemic control when residents were tested and quarantined if positive to prevent them infecting other household members.³² In many countries with ineffective lockdowns, testing, quarantine and household support mechanisms may be inadequate to obtain substantial benefits. Given the economic costs, better understanding of the performance of lockdowns in developing countries should be a global research priority.

Policy Implications

Our findings indicate that there is no single optimal level of testing. At any level, increases in testing further reduce transmission (online exhibit 3).¹⁵ When incidence is high and uncontrolled, all measures, including testing, might need to be intensified to achieve control and to make widespread testing and tracing feasible. When the virus is close to elimination or the reproduction number is substantially below one, increases in testing could be traded for relaxing other interventions, such as school and work closures, face masks and social distancing, as aptly demonstrated by the sustained return to normalcy in countries with intensive testing, such as China, New Zealand and Vietnam.

At the same time, almost all countries which reduced transmission to levels compatible with elimination were testing at TCR levels of 100 and above (online exhibit 1).¹⁵ This implies that most benchmarks suggested by the WHO, the US government and other agencies are inadequate. Given the effectiveness of whatever other interventions they were doing, most countries would likely have needed a TCR of at least 100 to achieve epidemic control. Supplementary exhibit S6,¹⁵ which shows that the bulk of all deaths occurred in countries that tested at the WHO benchmark level or lower, underlines this.

Our results imply that in “flattening the curve” strategies, which originate in CDC’s pandemic influenza planning,⁷ critical care capacity is the wrong threshold to target for COVID-19. At high incidence rates, even the wealthiest nations, such as the USA, UK, and Qatar, cannot expand testing and tracing fast enough (or they give up altogether) to achieve epidemic control. Early and continuous aggressive testing to keep incidence within capacity to test, trace and isolate may be the best implementation of flattening the curve.

Conclusion

We provide empirical evidence that testing intensity was the common factor explaining the success of places that achieved near-elimination, such as China, Cambodia and New Zealand, and the most important predictor of performance elsewhere. Given the costs and uncertainties associated with other NPIs, a strategy that relies much more on increased testing and isolation deserves serious consideration and resource allocation outside East Asia and the Pacific. It is likely to be less costly in terms of money, economic growth and human life.

ACKNOWLEDGMENTS

The authors thank Saroj Jayasinghe, Rob Condon, Owen O’Donnell, Delan Devakumar, David Scheerer and two anonymous reviewers for their comments and helpful observations, and Anabella Pinton for valuable research assistance. This study was not supported by any external funder, and the Institute for Health Policy would like to express its appreciation to members of the research team who accepted reductions in compensation during their work on the study at a time of financial stringency for the Institute.

NOTES

• 1 Hellewell J, Abbott S, Gimma A, Bosse NI, Jarvis CI, Russell TWet al. Feasibility of controlling COVID-19 outbreaks by isolation of cases and contacts. Lancet Glob Health. 2020;8(4):e488–e96. Crossref, Medline, Google Scholar
• 2 Cowling BJ, Ali ST, Ng TW, Tsang TK, Li JC, Fong MWet al. Impact assessment of non-pharmaceutical interventions against coronavirus disease 2019 and influenza in Hong Kong: An observational study. Lancet Public Health. 2020;5:e279–88. Crossref, Medline, Google Scholar
• 3 World Health Organization. WHO Director-General’s opening remarks at the media briefing on COVID-19—16 March 2020 [Internet]. Geneva: WHO; 2020 Mar 16 [cited 2020 May 28]. Available from: https://www.who.int/dg/speeches/detail/who-director-general-s-opening-remarks-at-the-media-briefing-on-covid-19—16-march-2020 Google Scholar
• 4 World Health Organization. COVID-19—virtual press conference—30 March 2020 [Internet]. Geneva: WHO; 2020 Mar 30 [cited 2020 May 28]. Available from: https://www.who.int/docs/default-source/coronaviruse/transcripts/who-audio-emergencies-coronavirus-press-conference-full-30mar2020.pdf?sfvrsn=6b68bc4a_2 Google Scholar
• 5 Graham D, Oré D. Mexico’s coronavirus death surge puts testing regime under the microscope. Reuters [serial on the Internet]. 2020 Jun 17 [cited 2020 Jun 20]. Available from: https://www.reuters.com/article/us-health-coronavirus-mexico-tests/mexicos-coronavirus-death-surge-puts-testing-regime-under-the-microscope-idUSKBN23N392 Google Scholar
• 6 Hogan L. Fighting alone. Washington Post [serial on the Internet]. 2020 Jul 16 [cited 2020 Jul 19]. Available from: https://www.washingtonpost.com/outlook/2020/07/16/larry-hogan-trump-coronavirus/ Google Scholar
• 7 Centers for Disease Control and Prevention. Interim Pre-pandemic Planning Guidance: Community Strategy for Pandemic Influenza Mitigation in the United States: early, targeted, layered use of nonpharmaceutical interventions [Internet]. Atlanta (GA): CDC; 2007 [cited 2020 Jun 10]. Available from: https://stacks.cdc.gov/view/cdc/11425 Google Scholar
• 8 Flaxman S, Mishra S, Gandy A, Unwin HJT, Mellan TA, Coupland Het al. Estimating the effects of non-pharmaceutical interventions on COVID-19 in Europe. Nature. 2020;584:257–61. Crossref, Medline, Google Scholar
• 9 Nouvellet P, Bhatia S, Cori A, Ainslie KEC, Baguelin M, Bhatt Set al. Reduction in mobility and COVID-19 transmission [Internet]. London: Imperial College London; 2020 [cited 2020 Jun 10]. (Report 26). Available from: https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/2020-06-08-COVID19-Report-26.pdf Google Scholar
• 10 Courtemanche C, Garuccio J, Le A, Pinkston J, Yelowitz A. Strong Social Distancing Measures In The United States Reduced The COVID-19 Growth Rate. Health Aff (Millwood). 2020;39(7):1237–46. Go to the article, Google Scholar
• 11 We searched PubMed on 29 August 2020 for studies evaluating the impact of testing on COVID-19 spread, using keywords including “COVID-19”, “2019-nCoV”, “SARS-CoV-2”, “transmission”, “intervention”, and “testing”, and identified other studies not reported in PubMed. This yielded 3,179 published studies and preprints, of which 30 reported quantitative analyses of the impact of testing (with or without tracing and isolation).
• 12 Liu Y, Morgenstern C, Kelly J, Lowe R, Jit M. The impact of non-pharmaceutical interventions on SARS-CoV-2 transmission across 130 countries and territories. medRxiv [serial on the Internet]. 2020 Aug [cited 2020 Nov 20]. Available from: https://www.medrxiv.org/content/10.1101/2020.08.11.20172643v1 Google Scholar
• 13 Cowling BJ, Lau EHY, Lam CLH, Cheng CKY, Kovar J, Chan KHet al. Effects of school closures, 2008 winter influenza season, Hong Kong. Emerging Infect Dis. 2008;14(10):1660–2. Crossref, Medline, Google Scholar
• 14 Huang KE, Lipsitch M, Shaman J, Goldstein E. The US 2009 A(H1N1) influenza epidemic: quantifying the impact of school openings on the reproductive number. Epidemiology. 2014;25(2):203–6. Crossref, Medline, Google Scholar
• 15 To access the appendix, click on the Details tab of the article online.
• 16 Johns Hopkins University. Covid-19 Data Repository by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University [Internet]. Baltimore (MD): Johns Hopkins University; 2020 [cited 2020 Sep 9]. Available from: https://github.com/CSSEGISandData/COVID-19 Google Scholar
• 17 Hasell J, Ortiz-Ospina E, Mathieu E, Ritchie H, Beltekian D, Macdonald Bet al. Coronavirus (COVID-19) Testing [Internet]. Cambridge (MA): Our World in Data; 2020 [cited 2020 Sep 09]. Available from: https://ourworldindata.org/coronavirus-testing Google Scholar
• 18 Hale T, Angrist N, Kira B, Petherick A, Phillips T, Webster S. Variation in government responses to COVID-19 [Internet]. Oxford: University of Oxford, Blavatnik School of Government; 2020 [cited 2020 Jun 10]. (BSG Working Paper Series). Available from: https://www.bsg.ox.ac.uk/sites/default/files/2020-05/BSG-WP-2020-032-v6.0.pdf Google Scholar
• 19 Google COVID-19 Community Mobility Reports [database on the Internet]. 2020 [cited 2020 Sep 9]. Available from: https://www.google.com/covid19/mobility/ Google Scholar
• 20 Facebook Data for Good. Movement Range Maps [serial on the Internet]. 2020; [cited 2020 Nov 20]. Available from: https://dataforgood.fb.com/tools/movement-range-maps/ Google Scholar
• 21 Moore M, Gelfeld B, Okunogbe A, Paul C. Identifying future disease hot spots: Infectious Disease Vulnerability Index [Internet]. Santa Monica (CA): Rand Corporation; 2016 [cited 2020 Jun 10]. Available from: https://www.rand.org/pubs/research_reports/RR1605.html Google Scholar
• 22 Nuclear Threat Initiative, Johns Hopkins Bloomberg School of Public Health. Global Health Security Index: building collective action and accountability [Internet]. Baltimore (MD): Johns Hopkins University; 2019 [cited 2020 Jun 10]. Available from: https://www.ghsindex.org/wp-content/uploads/2020/04/2019-Global-Health-Security-Index.pdf Google Scholar
• 23 World Bank. World Development Indicators [Internet]. Washington (DC): World Bank Group; 2020 [cited 2020 Sep 9]. Available from: https://data.worldbank.org/indicator/ Google Scholar
• 24 Various serological surveys indicate that even in most affected places, with the exception of New York, less than 5% of the population would have been infected prior to June 2020.
• 25 Abbott S, Hellewell J, Thompson RN, Sherratt K, Gibbs HP, Bosse NIet al. Estimating the time-varying reproduction number of SARS-CoV-2 using national and subnational case counts [version 1; peer review: awaiting peer review]. Wellcome Open Research [serial on the Internet]. 2020,5:112 [cited 2020 Nov 20]. Available from: https://wellcomeopenresearch.org/articles/5-112 Google Scholar
• 26 White House. Opening Up America Again [Internet]. Washington (DC): White House; 2020 [cited 2020 Jul 19]. Available from: https://www.whitehouse.gov/wp-content/uploads/2020/04/Guidelines-for-Opening-Up-America-Again.pdf Google Scholar
• 27 Lauer SA, Grantz KH, Bi Q, Jones FK, Zheng Q, Meredith HRet al. The Incubation Period of Coronavirus Disease 2019 (COVID-19) From Publicly Reported Confirmed Cases: Estimation and Application. Ann Intern Med. 2020;172(9):577–82. Crossref, Medline, Google Scholar
• 28 Gloeckner S, Krause G, Hoehle M. Now-casting the COVID-19 epidemic: The use case of Japan, March 2020. medRxiv [serial on the Internet]. 2020 Mar [cited 2020 Nov 20]. Available from: https://www.medrxiv.org/content/10.1101/2020.03.18.20037473v1 Google Scholar
• 29 StataCorp. Stata Statistical Software: Release 14. College Station (TX): StataCorp LP; 2015. Google Scholar
• 30 R Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing; 2020. Google Scholar
• 31 Ferguson N, Laydon D, Nedjati-Gilani G, Imai N, Ainslie K, Baguelin Met al. Impact of non-pharmaceutical interventions (NPIs) to reduce COVID-19 mortality and healthcare demand [Internet]. London: Imperial College London; 2020 [cited 2020 Jun 10]. (Report 9). Available from: https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/2020-03-16-COVID19-Report-9.pdf Google Scholar
• 32 Hao X, Cheng S, Wu D, Wu T, Lin X, Wang C. Reconstruction of the full transmission dynamics of COVID-19 in Wuhan. Nature. 2020;584(7821):420–4. Crossref, Medline, Google Scholar