https://orcid.org/0000-0002-7987-2199
Figures
Abstract
During the outbreak of the COVID-19 pandemic, Non-Pharmaceutical and Pharmaceutical treatments were alternative strategies for governments to intervene. Though many of these intervention methods proved to be effective to stop the spread of COVID-19, i.e., lockdown and curfew, they also posed risk to the economy; in such a scenario, an analysis on how to strike a balance becomes urgent. Our research leverages the mobility big data from the University of Maryland COVID-19 Impact Analysis Platform and employs the Generalized Additive Model (GAM), to understand how the social demographic variables, NPTs (Non-Pharmaceutical Treatments) and PTs (Pharmaceutical Treatments) affect the New Death Rate (NDR) at county-level. We also portray the mutual and interactive effects of NPTs and PTs on NDR. Our results show that there exists a specific usage rate of PTs where its marginal effect starts to suppress the NDR growth, and this specific rate can be reduced through implementing the NPTs.
Citation: Luo W, Guo W, Hu S, Yang M, Hu X, Xiong C (2021) Flatten the curve: Empirical evidence on how non-pharmaceutical interventions substituted pharmaceutical treatments during COVID-19 pandemic. PLoS ONE 16(10): e0258379. https://doi.org/10.1371/journal.pone.0258379
Editor: Yajie Zou, Tongji University, CHINA
Received: July 5, 2021; Accepted: September 26, 2021; Published: October 11, 2021
Copyright: © 2021 Luo et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The codes and dataset are publicly available on Figshare. The link is https://doi.org/10.6084/m9.figshare.16641358.v1.
Funding: The authors received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
COVID-19 has brought unprecedented global problems. As of June 10, 2020, the rapidly spreading COVID-19 virus has infected 7,492,360 people and claimed 422,150 lives across the world. The first case of COVID-19 in the US was confirmed in Washington State on January 21, 2020. On March 13, 2020, only two days after the World Health Organization (WHO) announced the COVID-19 as a world-wide pandemic, the U.S. government proclaimed a national state of emergency concerning the COVID-19 outbreak [1]. By mid-April 2020, stay-at-home orders were issued across all but 8 states in mid-west region of the US. This indicates that at least 316 million people were being urged to stay home, reduce unnecessary contact, and keep social distance [2]. However, the restrictions did not last long. On April 16, 2020, the guidelines of reopening the nation were released by the White House. In consonance with that, most of the lockdown states began to reopen in some manner. As of May 1, 2020, 18 states had lifted their stay-at-home orders or partially selected some regions or businesses to reopen [3]. Within only 3 months and 6 days, the number of confirmed cases exceeded 1 million on April 27, 2020 [4].
Until the wide availability of a vaccine, traditional interventions such as social distancing, hand sanitizing, wearing masks and utilizing ventilators remain the primary mechanisms to slow the spread of COVID-19. The governments of different countries have promulgated various countermeasures which can be roughly classified into Non-Pharmaceutical Treatments (NPTs) and Pharmaceutical Treatments (PTs). NPTs are actions, apart from getting vaccinated and taking medicine, that people and communities can take to help slow the spread of illnesses like pandemic influenza [5]. NPTs have been proven to be considerable in delaying and containing the spread of the virus [6–12]. Many state governments issued stay-at-home orders, shut down businesses, and limited gatherings to restrict human mobilities. Several research reveal a positive relationship between human mobility and COVID-19 infections [13, 14]. Also, some researchers mentioned that if people actively cooperate and comply with adaptations such as hand washing, masking, and social distancing, the infectivity and impact of the virus can be alleviated [7, 15]. However, moderate interventions such as school closures, self-isolation of symptomatic individuals, or shielding of older people would probably not be sufficient to control the epidemic and to avoid far exceeding available ICU capacity, even using these measures in combination [16]. Travel restriction policy is particularly useful in the early stage of an outbreak when it is confined to a certain area acting as a major source, but it may be less effective once the outbreak is more widespread at a later stage [17]. Geographic Information Systems (GIS) and big data technologies have played an important role in many aspects to fight against the virus, e.g., spatial tracking, prediction of regional transmission, spatial segmentation of the epidemic risk and prevention level and so on [18]. Mobility patterns are found to be strongly correlated with decreased COVID-19 case growth rates for the most affected counties in the USA [19].
From the perspective of PTs, recommendations on managing the Intensive Care Unit (ICU) as well as the infrastructure, supplies, and staff have been proposed to help ICU practitioners, hospital administrators, governments, and policy makers prepare for the upcoming challenge [20]. The World Health Organization (WHO) developed a suite of complementary surge calculators to help governments, partners, and other stakeholders to estimate potential requirements for essential supplies to respond to the current COVID-19 pandemic [21].
Due to the limited time of the pandemic and the availability of data resources and analytical algorithms, several major knowledge gaps exist and are worthy of attention. First, during a second spike in outbreaks without wide availability of vaccines, governments and people become more concerned with finding the balance point for the economy against the pandemic. However, the vast majority of related research only focuses on non-pharmaceutical intervention methods and ignores the pharmaceutical methods, and there is an even greater lack of discussions on how to make a trade-off on both methods. Second, many previous research oversimplified their models; some critical covariates such as socio-demographics, points of interests, and so on are largely ignored [22].
The main contributions of this research can be summarized as follows. First, this research leverages the aggregated data from large amounts of location-based service data which is more informative and accurate than assumptions on disease transmission rate or reproduction rate [23]. Second, most of the previous researchers solely focused on NPTs or PTs, and no one has studied the substitution effect of NPT and PT. This research is also among the first to quantify the relationship between NPTs and PTs, and to visualize their substitutional impact. This is innovative and timely research that will provide practical information to the medical system. It also provides a good prediction on New Death Rate, New Case Rate, and New Mild Case Rate. Third, this study contributes to broader general interests. Utilizing the daily updated web portal [24] on the pre-trained models mentioned above, we can deliver days- or weeks-ahead predictions to the public.
2. Data source
The data source of our study is a comprehensive national human mobility dataset from the University of Maryland COVID-19 Impact Analysis Platform [24, 25]. This platform incorporated over 150 million anonymous individuals monthly active mobile devices to develop the dataset of person movements for the U.S. These data are collected from individual devices including iOS and Android OS. The risk of re-identification is reduced by applying privacy-preserving techniques, such as aggregating to census block group level or county level. More technical details can be found in this paper [13] and the project report [26]. We utilize the data from February 22, 2020 to May 27, 2020 and conduct all the regressions at county level. The NPT and PT metrics are shown in Table 1. We also employ social demographic metrics as covariates in our models to rule out some static impacts. In addition, we also consider the factors referring to time-varying, weekend dummy, governors’ approval rates among states, random effects of different states and federal emergency status in our model.
4. Models
4.1. GAM and data preprocessing
This section provides a detailed description of the GAM (Generalized Additive Model) we employed to examine the NPTs v.s. PTs on controlling New Death Rate, New Case Rate and New Mild Case Rate. GAM was originally developed by Hastie & Tibshirani [27] to blend properties of generalized linear models with additive models. Its linear response variables blend in unknown smooth functions of additional independent variables. The inference of these non-linear smooth functions is the highlight of the GAM methods and often draws more research interests. GAM is more flexible on the assumptions for each independent variable. The functions of independent variables can be either a specified parametric form (such as a polynomial), a non-parametric form, or a semi-parametric form, simply as smooth functions.
As previously mentioned, the data starts from Feb. 22, 2020 when there were no cases in most counties. We could simply “one-size-fits-all” the data at a specific date for all counties; however, due to the variation of the first confirmed case among all counties, doing so may exclude significant information from several counties. In order to preserve as much and as precise information as possible and to differentiate the periods before and after the pandemic, we choose the first case occurrence time point of each county as the cutting point and set to zero all independent variables previous to the first case in the panel data.
Typical linear models fail to capture non-linear effects. To address this model limitation, GAM is adopted to incorporate highly complex non-linear relationships with its flexibility on smooth function forms. Moreover, linear predictors can be mixed in with non-linear predictors to enhance the interpretability of the model. In terms of the GAM structure involved in this discussion, the functional form is written as below:
In order to smooth out the in-week pattern and some short-term fluctuations, dependent variables and some predictors are calculated by a 7-day moving average. Yti represents New Death Rate, New Case Rate, and New Mild Case Rate on day t of county i. g(.) represents the link function between independent variables and the dependent variable; Several research assumed that the virus reproduction numbers are gamma distributed in transmission modeling [28–30], here we also assume the dependent variable following a Gamma distribution applying the log link function on the left-hand side of the formula. Xtim is the mth fixed effect linear covariate on day t of county i with a coefficient βm in a set of covariates with total number M. Xtin is the nth non-linear covariate on day t of county i in a set of covariates with total number N. fn represents a smooth function for Xtin. fpq(Xtip, Xtiq) denotes the smooth interaction of Xtip and Xtiq with a function form fpq. The models use a thin plate regression spline basis for each smooth function [31]. P denotes the set of variables on which we want to observe the effects of interaction.
We utilize variance inflation factor (VIF) to check the multicollinearity for the linear parts of the model. A threshold of 10.0 for each variable is introduced to filter out the highly multicollinear ones.
4.2 Summary of variables
A summary of all variables involved in the models is shown in Table 2. Each variable has 301536 observations.
4.3. Model selection
Our aim is to fit a statistical model for each county and the whole nation to understand how the social demographic variables, NPTs, and PTs affect the New Death Rate. Using this model, we predict the future trend of New Death Rate and inspect the relationships between NPTs and PTs. The panel data we leverage is from all counties in the US from Feb. 22, 2020 to May 27, 2020. For a robustness check, another two models fitting New Case Rate and New Mild Case Rate are used. Unequal lag periods are applied to the three models due to different lag time from symptom onset to case confirmation and death. Referring to the previous research [19], an optimal lag of days of 11 is calculated by optimizing the correlation between the pre-defined “mobility ratio” and “COVID-19 growth rate ratio”, therefore we chose 11 days for the New Case Rate and New Mild Case Rate models. Referring to the previous research [32], the median time delay is 13 days from illness onset to death. The number “13 days” can also be supported by another previous research [20] which said, “the median time from symptom onset to severe hypoxaemia and ICU admission is approximately 7–12 days”. Therefore, we used 11+13 = 24 days as the lag number for the New Death Rate model. The results of the three GAMs are shown in Table 3. The table is divided into two parts: upper half for coefficients corresponding to the linear fixed effects, and the lower half for non-linear smooth terms and interaction terms.
In the lower half, the smooth terms with parameter “bs = ‘re’” are set to produce a random coefficient for each level of the factor. The abbreviation “e.d.f” is short for “effective degrees of freedom”, which reflects the non-linearity of each smooth term. A larger “e.d.f” represents more wiggliness of the smooth term. The adjusted R-sq. is calculated by the formula as follows:
The percentage of deviance explained [27] in this paper is calculated as below:
- Null deviance: θ0 refers to the null model (i.e., intercept-only model).
- Residual deviance: θ0 refers to the trained model.
- y represents the outcome.
represents the estimate of the model.
and
are the parameters of the fitted saturated and proposed models, respectively. A saturated model has as many parameters as it has training points, that is, p = n.
- p(y|θ) is the likelihood of data given the model.
We’ve also applied the concurvity function in the “mgcv” package [33] on the three models to measure the concurvity among all the smooth terms in the three models. Generally, a value over 1.0 indicates a term may be a smooth curve of another. In the three GAMs, the concurvity of each combination of two variables among the non-linear parts are less than 1.0, indicating that the concurvity of the formulation is at the acceptable level.
5. Results
5.1. Inferential analysis
The “R-sq.” and “deviance explained” of the New Death Rate model are higher than the New Case Rate and New Mild Case Rate models, reflecting a better explanation of variances and a better improvement from null model to the fitted model for New Death Rate than for New Case Rate from the predictors. Intuitively, the type I and II statistical error of New Case Rate is harder to be reduced than New Death Rate due to the possibility of misdiagnoses, which may result in bias and pollute the data. As for linear effects in the three models, independent variables such as “Population Density” and “Hospital Bed” show the same positive direction of effects on dependent variables in all three models at a significant level. This indicates a higher risk for people in an area of higher population density and more points of interest. The race factors (“African percentage” and “Hispanic percentage”) are fitted as positive impacts on New Death Rate at a significant level. The age factor (“age60”) shows significantly positive impacts in the New Death Rate model and insignificant in New Case Rate and New Mild Case Rate model. This indicates that older people are more vulnerable than younger people. With regard to the NPTs (“Cross County trip” and “Cross State trip”), “Cross State trip” significantly influences the dependent variables positively in the three models. However, “Cross County trip” shows relatively weaker linear impacts compared with “Cross State trip” in all three models. As the virus spreads quickly and widely, we continue to observe a positive correlation between PTs (“ICU utilization” and “Daily Test”) and dependent variables. This indicates that the pandemic continues to exert great pressure on our medical system throughout the time interval among all counties in our research.
With respect to the non-linear effects in these three models, the estimated degrees of freedom of the significant smooth terms are largely greater than 1.0, indicating a strong non-linearity for the relationship between the smooth terms and dependent variables. The “Weekday” smooth term is insignificant in all three models, indicating there are no weekly patterns of the dependent variables. The P-values of the smooth terms are smaller than 0.001 indicating their statistical significance. Fig 1A–1C show the interactive effect of NPT and PT on the dependent variable in both 3-D and contour plotting. Of particular note is that the New Death Rate decreases when the “Cross State trip” drops. Additionally, where the “Cross State trip” equals 0, the value of effect peaks when “ICU utilization” is approximately 30 and then decreases along the axis. In other words, the marginal effect of “ICU utilization” on “New Death Rate” sharply drops at value 30 where “Cross State trip” equals 0. At the value of “Cross State trip” under 150, the effect remains steady at a lower value. However, as the value of “Cross State trip” becomes larger, a higher utilization rate of ICU is needed to flatten the growth curve of the effect on New Death Rate. For instance, with the value of “Cross State trip” exceeding 150, we may observe a drastic rising effect on “New Death Rate”. What stands out in Fig 1A is that the derivative of the effect on New Death Rate reaches 0 at a lower value along the “ICU utilization” axis given a lower “Cross State trip” (the cross-section curve becomes flat) comparing with that given a higher “Cross State trip” value. Given a specific “Cross State trip” value, the point at where the effect of “ICU utilization” reaching its peak and then reducing can be regarded as a “status shift point” of our medical system. This peak point gradually disappears along with the increasing of “Cross State trip” value. Therefore, the observation above supports that the “status shift point” is expected to be at an earlier stage along the “ICU utilization” axis when the value of “Cross State trip” decreases. This supports with intuitive thinking that when more people travel in a pandemic, the greater pressure will be imposed on our medical system, and hence make it more challenging for our medical system to subdue the virus. The curved surfaces in Fig 1A–1C indicate quantifiably substitutional impacts of NPT to PT. For example, suppose the “ICU utilization” equals 120. Then, as long as the “Cross State trip” reduces from 225 to 100, the effect on Ln(New Death Rate+1) after 24 days decreases approximately from 2.75 to 1.0 (decrease from 14.64 to 1.72 by converting the effect on Ln(New Death Rate+1) to New Death Rate). Similar findings also appear in the New Case Rate model and the New Mild Case Rate model in Fig 1B and 1C. Same as the New Death Rate model, the more people travel, the “status shift point” is expected to be at a later stage or even disappear on “ICU utilization” axis.
(a) Interaction smooth terms of New Death Rate model (Gamma). (b) Interaction smooth terms of New Case Rate model (Gamma). (c) Interaction smooth terms of New Mild Case Rate model (Gamma).
5.2 Robustness check
To do further robustness checks, we fit 3 new models as comparison by applying an identity link function on dependent variables and assuming them to have a Gaussian distribution. The fitting results are listed in Table 4. The findings in the new fitted models are similar to those reported in section 5.1. It shows the robustness on all explanatory variables. Table 5 makes a brief comparison for all 6 models. The adjusted R-sq., deviance explained, AIC and BIC of Gamma models are better than the Gaussian models, showing that the Gamma models fit better than Gaussian models. It indicates that the assumption of Gamma distribution is closer to reality than Gaussian distribution. Comparing the goodness-of-fit among different dependent variables in Gamma models, we observe a better fit for the New Death Rate model than for the New (Mild) Case Rate model. One convincing explanation is that the daily published death number is more reliable than the case number, as case numbers are more likely to be impacted by diverse diagnostic modes, misdiagnosis, restricted test capacity or even political reasons. In contrast, when people die from illness, the related identification documents registered in government agencies have to be modified, which makes the death number more convincing and more difficult to manipulate.
In order to check the robustness of the predictive ability, we also applied the 6 different models to make predictions starting from May 28 to June 10. The training and prediction process is looped weekly. The models are trained by all daily data before a specific day to predict the dependent variable on the next week. The results are shown in Fig 2. The black curves represent the New Death Rate, New Case Rate, and New Mild Case Rate in the subplots, respectively. The green curves stand for models with Gamma distribution assumption. The red curves represent models with Gaussian distribution assumption. The “ICU utilization” is plotted in blue histogram in the background. “Cross State trip” is barplotted in orange as background as well. It is observed that all models show good fittings and predictions.
6. Conclusion and discussion
The federal and state governments have enacted various complex combinations of responses to COVID-19. The policies affect various people’s activities and result in different patterns of movement and behavioral change. Many state governments have issued mandatory orders impacting economic behaviors, and they have also increased their production and supplies of masks, sanitizers, ventilators and other medical materials. This landscape makes balancing the effect of NPTs and PTs a non-trivial task. This research is based on location-based service data; we captured people’s real-time behavior instead of making assumptions of the efficacy of NPT. This means our mobility metrics were driven by the actual number of people who moved across state-bounds or county-bounds.
This research builds models for the whole nation from a macro point of view, treating various states as random effects instead of building a unique model for each state or county. Our results show a strong and statistically significant correlation between New Death Rate, New Case Rate, New Mild Case Rate, and the treatments. We portrayed the variation and captured the “status shift point” of PT at various levels of NPT. Crucially, the quantified interactive and substitutional impact among NPT and PT should serve to support more accurate policy making for state governments to find a better trade-off at an early stage in a pandemic. If cross-state trips are reduced, this would potentially lessen not only new deaths and cases but also new deaths and cases per ICU unit; far less efforts of PTs would be needed to stop the spread of virus; the medical system would operate more smoothly hence the unit efficiency of the medical system would increase.
The limitations of this research are summarized below. Though factors regarding spatial density (“PD”: Population density, “Hot”: Point of Interest) are captured as significant factors, they are both static variables. A combination with travel behaviors and spatial encounters should be involved in a next-step study which reflects dynamic effects for factors referring to spatial density. For further studies aiming at microscopic analysis, it might be possible to utilize zip-code level data or more fine-grained trajectory data to analyze individual encounters and incorporate additional details. Despite this research is mainly focusing on finding the substitution effect, one next step on making better predictions is also significant. Several previous research suggest methods to overcome the model uncertainty and make better predictions. For instance, we might explore applying different types of models and aggregate them by Bayesian Model Averaging methodology [34, 35].
This research mainly discussed a one-way relationship between human mobility and the spread of the COVID-19 virus. However, the opposite direction or the bi-directional relationship also exist between the two variables. Previous research explored the opposite direction on these two variables [11]. For the next step, the authors also plan to apply SEM (Structural Equation Modeling) methodology [36, 37] to explore further on the bi-directional relationship. In addition, people’s risk perception over time might significantly affect their behaviors and affect the spread of the virus. In this case, we plan to utilize social media data such as Facebook, Twitter, Instagram and so on as input to our model to capture people’s changing in the risk perception. Several studies have utilized social media data to extract people’s perception on COVID-19 related incidents like travelling [38], lockdown [39] and risk mitigation strategies [40].
As local governments will have to lift part of the mandatory orders as the time goes by, more refined analysis on balancing the NPTs and PTs are urgently needed. Our findings highlight a difficult point of decision-making in a pandemic, and also provide a macro view model to support decision making. We hope our study may motivate both individuals and governors to make more optimized decisions to help slow down the spread of pandemic.
Acknowledgments
We would like to thank and acknowledge Amazon Web Service and its Senior Solutions Architect, Jianjun Xu, for providing cloud computing and technical support.
References
- 1.
The White House. (n.d.). (2020). Proclamation on Declaring a National Emergency Concerning the Novel Coronavirus Disease (COVID-19) Outbreak | https://www.whitehouse.gov/presidential-actions/proclamation-declaring-national-emergency-concerning-novel-coronavirus-disease-covid-19-outbreak/
- 2.
The New York Times. (n.d.). (2020a). See How All 50 States Are Reopening (and Closing Again) - https://www.nytimes.com/interactive/2020/us/states-reopen-map-coronavirus.html
- 3.
The New York Times. (n.d.). (2020b). See Which States and Cities Have Told Residents to Stay at Home - https://www.nytimes.com/interactive/2020/us/coronavirus-stay-at-home-order.html
- 4.
1Point3Acres. (n.d.). COVID-19/Coronavirus Real Time Updates With Credible Sources in US and Canada | https://coronavirus.1point3acres.com/en
- 5.
Nonpharmaceutical Interventions (NPIs) | CDC. (n.d.). Retrieved May 13, 2021, from https://www.cdc.gov/nonpharmaceutical-interventions/index.html
- 6. Chinazzi M., Davis J. T., Ajelli M., Gioannini C., Litvinova M., Merler S., et al. (2020). The effect of travel restrictions on the spread of the 2019 novel coronavirus (COVID-19) outbreak. Science, 368(6489), 395–400. pmid:32144116
- 7. Courtemanche C., Garuccio J., Le A., Pinkston J., & Yelowitz A. (2020). Strong social distancing measures in the united states reduced the covid-19 growth rate. Health Affairs, 39(7), 1237–1246. pmid:32407171
- 8. Cowling B. J., Ali S. T., Ng T. W. Y., Tsang T. K., Li J. C. M., et al. (2020). Impact assessment of non-pharmaceutical interventions against coronavirus disease 2019 and influenza in Hong Kong: an observational study. The Lancet Public Health, 5(5), e279–e288. pmid:32311320
- 9. Lee M., Zhao J., Sun Q., Pan Y., Zhou W., Xiong C., et al. (2020). Human mobility trends during the early stage of the COVID-19 pandemic in the United States. PLoS ONE, 15(11 November), 1–15. pmid:33166301
- 10. White E. R., & Hébert-Dufresne L. (2020). State-level variation of initial COVID-19 dynamics in the United States. PLoS ONE, 15(10 OCTOBER), 1–13. pmid:33048967
- 11. Xiong C., Hu S., Yang M., Younes H., Luo W., Ghader S., et al. (2020a). Mobile device location data reveal human mobility response to state-level stay-at-home orders during the COVID-19 pandemic in the USA. Journal of the Royal Society Interface, 17(173), 20200344. pmid:33323055
- 12. Xiao Y., Yang M., Zhu Z., Yang H., Zhang L., & Ghader S. (2021). Modeling indoor-level non-pharmaceutical interventions during the COVID-19 pandemic: a pedestrian dynamics-based microscopic simulation approach. Transport Policy, 109, 12–23. pmid:34025048
- 13. Hu S., Xiong C., Yang M., Younes H., Luo W., & Zhang L. (2021). A big-data driven approach to analyzing and modeling human mobility trend under non-pharmaceutical interventions during COVID-19 pandemic. Transportation Research Part C: Emerging Technologies, 124(June 2020), 102955. pmid:33456212
- 14. Xiong C., Hu S., Yang M., Luo W., & Zhang L. (2020b). Mobile device data reveal the dynamics in a positive relationship between human mobility and COVID-19 infections. Proceedings of the National Academy of Sciences of the United States of America, 117(44), 27087–27089. pmid:33060300
- 15. Park S. W., Sun K., Viboud C., Grenfell B. T., & Dushoff J. (2020). Potential role of social distancing in mitigating spread of coronavirus disease, South Korea. Emerging Infectious Diseases, 26(11), 1–21. https://doi.org/10.3201/EID2611.201099
- 16. Davies N. G., Kucharski A. J., Eggo R. M., Gimma A., Edmunds W. J., Jombart T., et al. (2020). Effects of non-pharmaceutical interventions on COVID-19 cases, deaths, and demand for hospital services in the UK: a modelling study. The Lancet Public Health, 11–15. https://doi.org/10.1016/s2468-2667(20)30133-x
- 17. Kraemer M. U. G., Yang C. H., Gutierrez B., Wu C. H., Klein B., Pigott D. M., et al. (2020). The effect of human mobility and control measures on the COVID-19 epidemic in China. Science, 368(6490), 493–497. pmid:32213647
- 18. Zhou C., Su F., Pei T., Zhang A., Du Y., Luo B., et al. (2020). COVID-19: Challenges to GIS with Big Data. Geography and Sustainability, 1(1), 77–87. https://doi.org/10.1016/j.geosus.2020.03.005
- 19. Badr H. S., Du H., Marshall M., Dong E., Squire M. M., & Gardner L. M. (2020). Association between mobility patterns and COVID-19 transmission in the USA: a mathematical modelling study. The Lancet Infectious Diseases, 0(0), 1–8. pmid:31876483
- 20. Phua J., Weng L., Ling L., Egi M., Lim C. M., Divatia J. V., et al. (2020). Intensive care management of coronavirus disease 2019 (COVID-19): challenges and recommendations. The Lancet Respiratory Medicine, 8(5), 506–517. pmid:32272080
- 21.
WHO. (2019). Essential resource planning. (n.d.). https://www.who.int/emergencies/diseases/novel-coronavirus-2019/technical-guidance/covid-19-critical-items
- 22. Glaeser E. L., Gorback C. S., & Redding S. J. (2020). How much does COVID-19 increase with mobility? Evidence from NEW YORK and four other U.S. cities. NBER Working Paper Series, 53(9), 1689–1699. pmid:33106711
- 23. Ray D, Salvatore M, Bhattacharyya R, Wang L, Du J, Mohammed S, et al. (2020). Predictions, Role of Interventions and Effects of a Historic National Lockdown in India’s Response to the the COVID-19 Pandemic: Data Science Call to Arms. Harvard Data Science Review, 176(3), 139–148. pmid:32607504
- 24.
Maryland Transportation Institute. (2020). University of Maryland COVID-19 Impact Analysis Platform. University of Maryland, College Park, USA. https://data.covid.umd.edu
- 25. Zhang L., Ghader S., Pack M. L., Xiong C., Darzi A., Yang M., et al. (2020a). an Interactive Covid-19 Mobility Impact and Social Distancing Analysis Platform. Medrxiv, 1–14. https://doi.org/10.1101/2020.04.29.20085472
- 26. Zhang L., Ghader S., Darzi A., Pan Y., Yang M., Sun Q., et al. (2020b). Data Analytics and Modeling Methods for Tracking and Predicting Origin-Destination Travel Trends Based on Mobile Device Data. Federal Highway Administration Exploratory Advanced Research Program.
- 27.
Hastie T. J., & Tibshirani R. J. (1990). Generalized Additive Models, Monographs on Statistics and Applied Probability 43. In Chapman & Hall. https://doi.org/10.1016/j.csda.2010.05.004
- 28. Lloyd-Smith J. O., Schreiber S. J., Kopp P. E., & Getz W. M. (2005). Superspreading and the effect of individual variation on disease emergence. Nature, 438(7066), 355–359. pmid:16292310
- 29. Pasetto D., Lemaitre J. C., Bertuzzo E., Gatto M., & Rinaldo A. (2021). Range of reproduction number estimates for COVID-19 spread. Biochemical and Biophysical Research Communications, 538, 253–258. pmid:33342517
- 30. Zhao H., Merchant N. N., McNulty A., Radcliff T. A., Cote M. J., Fischer R. S. B., et al. (2021). COVID-19: Short term prediction model using daily incidence data. PLoS ONE, 16(4 April), 1–14. pmid:33852642
- 31. Wood S. N. (2003). Thin Plate Regression Splines. Journal of the Royal Statistical Society. Series B (Statistical Methodology), 65(1), 95–114.
- 32. Wilson N., Kvalsvig A., Barnard L. T., & Baker M. G. (2020). Case-fatality risk estimates for COVID-19 calculated by using a lag time for fatality. Emerging Infectious Diseases, 26(6), 1339–1341. pmid:32168463
- 33.
Wood, S. N. (2017). Generalized additive models: An introduction with R, second edition. In Generalized Additive Models: An Introduction with R, Second Edition. https://doi.org/10.1201/9781315370279
- 34. Erdoğdu H., Erdem N., & Nacar F. (2021). Housing appraisal under model uncertainty: Bayesian model averaging method. 1(1), 26–34.
- 35. Zou Y., Lin B., Yang X., Wu L., Muneeb Abid M., & Tang J. (2021). Application of the Bayesian Model Averaging in Analyzing Freeway Traffic Incident Clearance Time for Emergency Management. Journal of Advanced Transportation, 2021. https://doi.org/10.1155/2021/6671983
- 36. Mueller R. O., & Hancock G. R. (2011). Best Practices in Structural Equation Modeling. Best Practices in Quantitative Methods, 488–508. https://doi.org/10.4135/9781412995627.D38
- 37. Zong F., Yu P., Tang J., & Sun X. (2019). Understanding parking decisions with structural equation modeling. Physica A: Statistical Mechanics and Its Applications, 523, 408–417.
- 38. Yu M., Li Z., Yu Z., He J., & Zhou J. (2020). Current Issues in Tourism Communication related health crisis on social media: a case of COVID-19 outbreak. https://doi.org/10.1080/13683500.2020.1752632
- 39. Li L., Ma Z., Lee H., & Lee S. (2021). Can social media data be used to evaluate the risk of human interactions during the COVID-19 pandemic? International Journal of Disaster Risk Reduction, 56(February), 102142. pmid:33643835
- 40. Malecki K. M. C., Keating J. A., & Safdar N. (2021). Crisis Communication and Public Perception of COVID-19 Risk in the Era of Social Media. Clinical Infectious Diseases, 72(4), 697–702. pmid:32544242