Introduction
Health systems have a finite capacity to provide care. During the COVID-19 pandemic, this capacity was stretched by demand in most countries worldwide, and in many, was exceeded.1 Some health systems responded to unprecedented capacity challenges by instituting formal resource allocation policiesReference Tyrrell, Mytton and Gentry2; other health systems responded by instituting efforts to increase capacityReference Powell and Chuang3; some health systems allocated resources according to ability to pay or other sociodemographic characteristics outside of the pandemic settingReference Berger and Miller4; finally, some health systems did not enact formal guidance on resource allocation.Reference Spector-Bagdady, Laventhal and Applewhite5 In the presence of on-the-ground capacity limitations, many clinicians were forced to enact informal, or implicit, rationing.Reference Powell and Chuang3, Reference Spector-Bagdady, Laventhal and Applewhite5
Crisis standards of care exist when a health system lacks capacity to provide usual care to patients.Reference Hanfling and Altevogt6 An established body of literature outlines methods for developing rationing protocols during crisis standards of care,Reference Powell and Chuang3, Reference Maves, Downar and Dichter7–Reference McGuire, Aulisio and Davis10 and guidelines for health system resource reallocation during the COVID-19 pandemic have been published.Reference Tyrrell, Mytton and Gentry2, Reference White and Lo11–Reference Riccioni, Bertolini and Giannini13 However, to date, no empiric evidence exists on the effects of implementation or failure to implement these guidelines.Reference Bollyky, Hulland and Barber14 Furthermore, available literature centers around a few highly resourced health systems in the United States and Europe.Reference Hick, Hanfling and Wynia15–Reference Emanuel and Persad18
This study directly examines policymaking within health systems, in contrast to the majority of studies to date, which have focused on non-pharmaceutical interventions outside of the health system (e.g., mask mandates, school and business closures, and travel restrictions).Reference Mendez-Brito, El Bcheraoui and Pozo-Martin19 Resource allocation policies within a health system are defined as “government policies which affect the material (e.g., medical equipment, number of hospitals for public health) or human (e.g., doctors, nurses) health resources of a country,”Reference Cheng, Barceló and Hartnett20 and could include, for example, redeployment of staff to other roles, procurement of medical supplies, or rescheduling of elective procedures and appointments. They can be instituted during, and perhaps in response to, surges in health system demand, which we consider a simultaneous policy. Alternatively, policymakers may choose to preemptively reallocate resources in an attempt to avert mortality. Effects of preemptive resource reallocation, e.g., in the prior month or week, can be associated with outcomes by examining temporal lags in statistical panel models.
Most available literature which has assessed policies has assessed policy interventions’ impact on COVID-19 cases, spread (e.g., R0), or case-fatality ratio.Reference Mendez-Brito, El Bcheraoui and Pozo-Martin21, Reference Yeoh, Chong and Chiew22 These measures can be subject to bias from variable access to testing across time and countries, and from variable coding of cause of death in vital statistics systems, a particularly important issue in cross-national comparisons.Reference Lipsitch, Donnelly and Fraser23–Reference Weinberger, Chen and Cohen25 This study uses an outcome measure which is more robust to international and temporal comparisons: excess mortality.
Therefore, this study assessed the effects of health system resource reallocation policies worldwide. The study’s hypothesis is that the scope, scale, and timing of resource reallocation within health systems were associated with all-cause mortality during the COVID-19 pandemic. Given the wealth of data and experience generated in responding to the COVID-19 pandemic, and the potential consequences of future similar crises, it is imperative to assess the impact of policies enacted across health systems worldwide.
Methods
Objectives and Design
This is a multinational, country-level longitudinal evaluation of the association between the scope and scale (summarized as intensity) of resource reallocation within health systems and excess mortality during the pre-vaccination COVID-19 pandemic. The study’s objective is to model timing and intensity of health system resource reallocation policies to predict excess mortality. As detailed below, all countries for which data were available were included (see Figure 1 or Online Data Supplement A1), comprising 85 countries from all continents except Antarctica. Data covered the period January 1, 2020-January 31, 2021. Although the study models the effects of policies rather than synthesizes health estimates, it includes multiple data inputs, so for transparency, it adheres to the Guidelines for Accurate and Transparent Health Estimates Reporting.Reference Stevens, Alkema and Black26
Figure 1. Timing of health system resource reallocation policymaking intensity, in black, and excess all-cause mortality, in gray, during the early COVID-19 pandemic in 81 countries worldwide (Figure 1A), and in the United States of America (Figure 1B). Excess mortality units are p-score, calculated as percentage of deaths greater than historical average. Note: Bosnia and Herzegovina abbreviated Bosnia & Herz.
Data Inputs
The predictor variable, intensity of health system resource reallocation, is derived from publicly available work by the CoronaNet project,Reference Cheng, Barceló and Hartnett20 which captures intensity of policymaking about health system resources. Data for CoronaNet were collected and cleaned by a worldwide network of research assistants during 2020-2021. Statistically validated indices were developed using Bayesian ideal point modelling to discover each country’s changing position on a scale of policymaking activity for each day of the pandemic.Reference Kubinec27–Reference Kubinec, Barceló and Goldszmidt29 For example, a national health system which redeploys all physicians toward the COVID-19 response has a high score on health resource policymaking intensity, whereas policies directed at subnational regions or which only redeploy nonessential staff are lower-intensity.
The study’s outcome variable, excess mortality, is a measure of how many more deaths (from any cause) occurred compared to an expected number of deaths for a given location and time period. This study uses a historical comparison to calculate expected deaths for each country during each week (or month) of the pandemic, compared to the mean deaths that week (or month) over a 5-year reference period preceding the pandemic:
$$ {\displaystyle \begin{array}{l}Excess\ mortality\hskip0.5em p\ score\ \\ {}=\frac{Death{s}_{Period\hskip0.5em 2020-21}-{Average\ deaths}_{Period\hskip0.5em 2015-19}}{{Average\ deaths}_{Period\hskip0.5em 2015-19}}\times 100\end{array}} $$
The outcome of excess mortality captures deaths from causes other than COVID-19, such as non-communicable diseases or maternal mortality, which may increase in the setting of health system crises.Reference Palmer, Monaco and Kivipelto30, Reference Kluge, Wickramasinghe and Rippin31 Considering all causes of mortality is critical to improving overall health system resource allocation and readiness.
This study uses excess mortality data from Our World in Data,Reference Ritchie, Mathieu and Rodes-Guirao32 which aggregates data from the Human Mortality DatabaseReference Wilmoth, Andreev and Jdanov33 and the World Mortality Dataset,Reference Karlinsky and Kobak34 provided by national statistical agencies and Eurostat. Depending on the vital statistics system of each country, excess mortality is calculated on a weekly or monthly basis. Excess mortality p-score can be negative (if deaths were fewer than expected) or positive.
Covariates include health system capacity parameters, including hospital beds per 10,000 population; domestic government and private expenditure on health per capita ($100 USD); completeness of cause-of-death reporting in national vital statistics systems; and whether medical devices are procured at the national level.35 Two additional composite scores were included, based on country self-reports to the World Health Organization (WHO) used to monitor compliance with the International Health Regulations (IHR): a score for health system crisis preparedness (Preparedness Score) and a score for legislative efficacy of IHR implementation (Legislation Score).
Finally, weekly or monthly COVID-19 incidence using case count and population data from Johns Hopkins University Center for Systems Science and EngineeringReference Dong, Du and Gardner36 were calculated per 1,000 population. These data may be biased by testing availability, particularly in the early pandemic or in lower-resourced settings, so sensitivity analyses were conducted prior to including them as covariates. Detailed descriptions of all data inputs are in Online Data Supplement A2.
Data Analysis
Distinct groups of countries report mortality weekly and monthly, so countries reporting mortality weekly (N = 42) and monthly (N = 43) were first separately analyzed. Finally, a combined analysis was conducted for all 85 countries.
The predictor variable, health resource allocation policymaking index, is normally distributed and was summarized with weekly or monthly means, as appropriate for each country’s mortality data. Descriptive analysis of all variables by country and region was conducted, identifying outliers through visual inspection of scatter plots and verifying data. Peak p-scores for Ecuador, in April 2020, were approximately double the next highest country’s peak p-score; this has been documented.Reference Cevallos-Valdiviezo, Vergara-Montesdeoca and Zambrano-Zambrano37 Sensitivity analyses with and without Ecuador produced similar models; Ecuador was included in the final analysis.
Next, random-intercept linear regression by maximum likelihood using the mixed command for panel data in Stata was performed with first-order autoregressive residuals. Country-specific error was assumed to vary and was thus estimated using random effects.
Autocorrelation was a priori assumed to be present in the policy predictor variable because it is estimated as t = t − 1 + Gaussian-distributed noise σi. Serial correlation of the residuals was examined with the Wooldridge-Drukker test for both monthly and weekly data, and the Q(p) test for monthly data, and was found to be significant at the first order.Reference Wursten38 Testing for autocorrelation at higher orders was nonsignificant using Inoue and Solon’s test; however, this test is limited with many time observations relative to N,Reference Wursten38, Reference Inoue and Solon39 so models with autoregressive residuals were iteratively examined until the Bayesian Information Criterion (BIC) was minimized. Adding higher-order autoregressive residuals beyond second-order did not improve model fit. Quadratic and cubic terms, which allow estimation of effect reversal, were added, and model fit was assessed. The direction of the associations was tested using Granger’s test, which was significant in the modeled direction of the association, resource allocation ➔ excess mortality, and was nonsignificant in the reverse direction, providing reassurance that the data support the modeled direction of this association.Reference Dumitrescu and Hurlin40 Lags of the predictor variable (health system resource allocation index) help capture delayed effects of policymaking activity on mortality or effects related to time delays in reporting of mortality during surge periods.Reference Angelopoulos, Pathak and Varma41 Lags were iteratively added and subtracted until BIC was minimized. Separately, vector autoregression models were estimated to determine optimal lag length for each country, which for most countries was 4 months based on BIC, consistent with the above analysis. The most conservative lag length (4 months) was thus chosen, although additional lags between 2 and 3 months did not improve model fit and were dropped, leaving only the first and fourth lags of resource allocation index.
Health system covariates were added to the best-performing lagged models, and fit was examined. The addition of all health system covariates improved model fit. For 4 locations (Hong Kong, Macao, Taiwan, and Kosovo), covariate data are not available from international agencies. Due to their political status, these data may not be missing at random, so these locations were not imputed and were excluded from the analysis (final N = 81). Unadjusted analyses were similar with or without these locations. Two covariates, cause of death completeness in national vital statistics systems, and procurement of medical devices at the national level, were not available for several of the 81 countries (see Online Data Supplement A2) and were neither statistically significant nor improved fit in the model including the 74 countries for which data were available, so were excluded from the final model to maximize geographic coverage.
Finally, to examine the entire dataset together, for weekly-mortality countries, variables were summarized as monthly means (incidence was recalculated as monthly), and the analysis was repeated with the entire dataset of 81 countries for time period t = 1 month. Results are reported below. In this combined dataset, several planned additional analyses were conducted. WHO region was investigated as an effect modifier of the relationship between the resource allocation index and excess mortality, to account for regional differences in health system policymaking and assess the representativeness of study results. Month and year were included as covariates to explore the contribution of potentially decreasing mortality as the pandemic evolved (due to changing supply chain constraints, prevention, and treatment modalities); these were nonsignificant and did not improve model fit. Finally, an age-stratified analysis was undertaken for the 33 countries where age-stratified mortality data were available, since resource constraints and rationing may explicitly or implicitly affect age groups differently.Reference Comas-Herrera and Fernández42, Reference Aron and Muellbauer43 Two sensitivity analyses were conducted to assess for the possibility of nonlinearity and collinearity in the measured associations and lags. First, the simultaneous predictor (lag 0) was removed from the model, and then quadratic and, separately, cubic, transformations of the predictor were added. Second, distributed lag linear and nonlinear models were constructed using the dlnm packageReference Gasparrini, Armstrong and Kenward44 in R 4.4 using natural cubic splines with 2-4 degrees of freedom.
This research did not involve human participants and, as such, was exempt from review per the regulations of the New York University Grossman School of Medicine Institutional Review Board. All analyses were conducted using Stata 13.1,45 except as mentioned above.
Results
Countries across the world had heterogeneous epidemic curves and different approaches to health system resource reallocation, as shown in Figure 1A. Some countries’ resource reallocation preceded increases in mortality, while in others (as in Figure 1B), the early peaks of both were simultaneous. The included covariates also had substantial variability, as shown in Table 1.
Table 1. Descriptive characteristics of country-months and countries for health system resource reallocation policy, excess mortality, and health system covariates
Notes: See Online Data Supplement A2 for further details on data sources.
Abbreviation: SD, Standard deviation.
Across these heterogeneous epidemic curves in the sample of 81 countries and 752 country-months of data, there was a consistent and strong association between health system resource reallocation policies and excess mortality. Specifically, as demonstrated in Table 2, health system resource reallocation intensity was positively associated with increased excess mortality when both occurred simultaneously (b = 0.80, 95% confidence interval, CI 0.42–1.18). However, the previous month’s (preemptive) resource reallocation (the first lag of the predictor variable with monthly data) was associated with a decrease in excess mortality (b = −0.58, 95%CI −0.93 to −0.23). Adding a 4-month lag, in addition to this prior month lag, improved model fit but was not statistically significant. The scale of these coefficients is substantial: for example, a 1-point increase in resource reallocation (which is on a scale of 0–100), such as occurred between April and May 2020 in the United States (demonstrated graphically in Figure 1B), was associated with an 80% increase in mortality. For April 2020, this 1-point increase translates to an additional 16,944 deaths in the United States. Reallocating resources a month earlier, conversely, was associated with 12,496 fewer deaths in April 2020 in the United States. Across all study countries taken together, each unit increase in resource reallocation occurring in February 2020 corresponded with a decrease of approximately 42,010 deaths in March 2020.
Table 2. Full and age-stratified models for the association between health system resource reallocation policy index and all-cause excess mortality, adjusted for COVID-19 incidence, baseline health system capacity, and preparedness, demonstrating lower mortality with proactive resource reallocation and increased mortality with reactive resource reallocation
Notes. Abbreviations as follows: COVID-19 incid., COVID-19 Incidence per 1000 population; IHR Legislation, International Health Regulations Legislation Score; IHR Preparedness, International Health Regulations Preparedness Score; Private spending, Private health expenditure per capita, $100 USD; Gov’t spending, Government health expenditure per capita, $100 USD; Hospital beds, Hospital beds per 10,000 population. Main model N = 752 country-months, 81 countries; Age-stratified models N=330 country-months, 33 countries. See Online Data Supplement A2 for further details on data sources and Online Data Supplement A3 for full model specification and fit statistics.
*P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.
Health system covariates were associated with excess mortality in expected directions (Table 2). COVID-19 incidence was strongly associated with excess mortality (b = 2.52, 95%CI 2.11–2.94). An IHR Preparedness Score 1% higher predicted a lower excess mortality with b = −0.25 (95%CI −0.43 to −0.064). Government health expenditure per capita ($100 USD) was associated with reduced mortality: for every $100 USD spent, a decrease in p-score of −0.36 (95%CI −0.64 to −0.073) was observed. Finally, baseline health system capacity was protective, with one additional hospital bed per 10,000 people predicting lower mortality, b = −0.19 (95%CI −0.33 to −0.057). IHR Legislation Score and private health expenditure per capita were not associated with excess mortality, but improved model fit and were retained in the final model.
A sensitivity analysis of countries which reported mortality weekly produced similar results to the overall model for the main predictors, lag structure, and covariates. The sensitivity analysis on the subgroup of countries reporting mortality monthly was nonsignificant for all variables except COVID-19 incidence. The overall model was similar with or without an indicator variable for mortality reporting (weekly vs. monthly).
In the smaller group of countries reporting mortality data by age group (N = 33; 330 country-months of observation), separate models stratifying excess mortality by age group found consistent results, with monotonically increasing coefficients at older ages for the impact of resource reallocation policy. As demonstrated in Table 2, in the youngest age group (15–64 years), the coefficient for simultaneous resource reallocation was b = 0.30 (95%CI 0.067–0.52), whereas preemptive resource reallocation (1 month prior) was again protective with b = −0.29 (95%CI −0.48 to −0.10). The impact increased until, in the age group over 85 years, simultaneous resource reallocation had a coefficient b = 0.92 (95%CI 0.61–1.23), while preemptive resource reallocation had a mirroring protective effect with b = −0.94 (95%CI −1.23 to −0.65).
In stratified analyses, WHO Region was an effect modifier of the relationship, with results similar to those reported above for the Americas and Europe regions, nonsignificant relationships in the Africa, Southeast Asia, and Western Pacific regions, and a trend toward effect reversal in the Eastern Mediterranean region, which included 5 countries (see Online Data Supplement A4).
Sensitivity analyses for collinearity of the lags and nonlinearity redemonstrated the effects observed in the main analyses and are presented in the Supplementary material A5.
Limitations
This study has several limitations. Other work has focused more specifically on time delays in mortality reporting.Reference Wang, Paulson and Pease46 This study did not evaluate for this possibility directly; however, the consistency of the empirically determined lag structure in the models, including in vector autoregression models for each country separately, argues against this as an explanation for the findings, since reporting delays are highly variable between countries.Reference Wang, Paulson and Pease46 It is possible that the lack of association when countries reporting mortality monthly were analyzed may be related to heterogeneous data quality in this subgroup of countries: in a post-hoc Student’s t-test, monthly-reporting countries had lower completeness of vital statistics cause-of-death information. In a similar vein, this multinational study included data from many disparate countries; however, country-specific intercept and error were allowed to vary in the aggregated models using random effects, which is a standard approach in multinational analyses.Reference Laird and Ware47–Reference Basagaña, Pedersen and Barrera-Gómez49
Debate exists surrounding mortality ascertainment. To avoid compounding uncertainty by modelling with modeled input data, this study uses a historical comparison to real mortality data for each country. Other methods require more stringent assumptions to model excess mortality, which produce roughly similar global estimates but vary widely on an individual country level, even among highly resourced health systems.Reference Wang, Paulson and Pease46 However, using actual mortality data necessitates the exclusion of many countries from the analysis, notably, lower-resourced health systems in Africa. Region modifies the effect of health system resource allocation on excess mortality, with the association appearing to be driven by Europe and the Americas in the stratified analysis. This may be related to data quality and availability (with 56% of countries from Europe and 23% from the Americas in this study’s sample); the conclusions should be further validated outside of these regions. However, resource allocation is arguably of even greater importance in lower-resourced settings50; ongoing mixed-methods work conducted in partnership with clinicians and researchers in lower-resourced settings addresses this. These issues provide a reminder that strengthening vital statistics systems is critical to global health security.
Finally, this study examines excess mortality in the acute phase of a pandemic. Health system resource reallocation may include interventions which, for example, delay diagnosis or care for chronic conditions, the effects of which may still not be fully realized. Untangling these effects statistically on a population level poses technical challenges given the myriad of competing risks present. Later peaks in excess mortality occurred in some countries, such as the United States, despite high overall levels of health system resource reallocation (e.g., Figure 1B), which could relate to subnational-level changes in resource reallocation and could be further investigated if data are available.
Discussion
Health system crisis readiness is universally acknowledged as important in academic literature51 but is variably implemented. Despite extensive analysis of COVID-19 pandemic responses, this study is the first to empirically examine the timing and intensity of health system policy change in anticipation and response to an international health security threat. These results support the importance of readiness and suggest there may be a “critical window” to intervene in health systems early in the course of a crisis. If that window is missed, whether due to legislative, capacity, or readiness gaps, these results suggest that simultaneous policies are associated with increased mortality. Overall, preemptive health system resource reallocation, approximately 1 month early, was strongly associated with protection against excess mortality in this sample of countries over the course of the early pandemic. These results also suggest an explanation for why certain countries fared worse than expectedReference Nuzzo, Bell and Cameron52: a lack of preemptive resource reallocation may have set countries up to implement simultaneous policies, reacting to peaks in mortality. These associations persist after controlling for COVID-19 incidence, baseline health system capacity, and preparedness covariates, which are also each statistically significant predictors of mortality. The increasing strength of these associations in older individuals across study countries underscores the importance of health system policies in either protecting or endangering vulnerable groups. While global mortality from COVID-19 was highest in older demographics,Reference Naghavi, Ong and Aali53 these results suggest that policy effects were also magnified in these particularly susceptible individuals.
This study may provide support to policymakers in deciding whether, and when, to institute resource reallocation policies to protect health and economic security. Policy is made in real-time with incomplete information; however, political pressures can also impair policymakers’ ability to act using available evidence or guidelines, even when they do exist.Reference Nuzzo, Bell and Cameron52, Reference Gutmann Koch and Han54 “The month before the surge”—when this study suggests action should be taken to prepare health systems—is only identifiable in retrospect in an individual location; however, epidemiologic evidence and data sharing can provide predictions about case surges. However, low trust in government has been identified by others as an important risk factor for infection fatality rate and may prevent policymakers from implementing lifesaving policy, even when epidemiologic evidence and resources are available.Reference Bollyky, Hulland and Barber14, 51, Reference Nuzzo, Bell and Cameron52
This study has several strengths in comparison to prior literature. Many studies attempting to estimate effects of non-pharmaceutical interventions on COVID-19 outcomes, particularly R0 and other case-related parameters, struggle to suggest causality due to the concern that more stringent measures were actually undertaken in response to increases in COVID-19 burden, so must introduce a time delay between exposure and COVID-19 outcome (ranging from several days to 6 weeks in the literature) to avoid reverse causality.Reference Besançon, Meyerowitz-Katz and Zanetti Chini55 By focusing specifically on the outcome of excess mortality (which does not in itself place a burden on the health system, unlike incidence or case severity), this study eliminated this bias. Note that, unlike the protective effect of preemptive resource reallocation, for which the clear counterfactual is inaction (no change in resource reallocation index), the simultaneous resource reallocation occurs at the same time as increased mortality, which can pose challenges in establishing the direction of effect. However, this study statistically confirmed Granger non-causality and empirically examined the lag structure of the effects to account for time delays in the modeled policy ➔ mortality effect. Sensitivity analyses also redemonstrated the importance of the prior month’s preemptive resource reallocation without the inclusion of the simultaneous resource reallocation predictor. The finding that policymaking in the fourth and first months prior to the index month were important determinants of mortality is logical when considering, for example, that COVID-19 surged in March-April 2020 in many countries after beginning in Wuhan, China, 4 months prior.
This study’s outcome measure, as well as its longitudinal panel methodology and its focus on health system policies, may account for differences in findings between this study and other work which has investigated the impact of health system capacity parameters and preparedness on COVID-19 outcomes.Reference Bollyky, Hulland and Barber14 Because of a lack of association between externally evaluated preparedness metrics such as Joint External Evaluations or Global Health Security Index scores and mortality in high-quality grey literature,51, 56 this study instead analyzed each country’s self-evaluation of preparedness, and found it consistently protective against excess mortality.
This study includes the time period covering mainly the year 2020 for 2 reasons: first, vaccinations became available beginning in early 2021; because this study did not aim to explore vaccination coverage and effects, this avoided bias related to differential vaccine coverage and rollout across countries. Second, this study aimed to understand the impacts of health system resource reallocation to provide lessons for future pandemics or disasters, and thus hypothesized that the early pandemic, before supply chains and system operations could be adjusted semi-permanently to adapt to a pandemic “new normal,” provides the best evidence to answer this question.
This study’s intent was not to discover which countries had better outcomes over the course of the pandemic. Rather, by incorporating a diverse sample of countries and analyzing their policymaking activity longitudinally, its aim was to understand which approaches to the timing of resource reallocation were protective or detrimental. A single country may have had multiple different approaches; countries had policy responses that occurred simultaneously with mortality at the beginning of the pandemic due to a variety of factors.51
Conclusions
In a longitudinal sample of 85 countries containing 2.4 billion people, preemptive health system resource reallocation was meaningfully protective against excess mortality during the early COVID-19 pandemic, while simultaneous health system resource reallocation was significantly associated with increased excess mortality. These effects increased markedly in older populations. Health system capacity, government expenditure on health, and country self-reports of preparedness were protective against excess mortality. This is the first empiric study of its kind to explore health system policymaking during the COVID-19 pandemic and confirms expert opinion51 about the key roles of readiness and early response in mitigating the effects of a pandemic. Future work is urgently needed to qualify key actions for health system crisis readiness and expand the scope of readiness to strengthen global health security.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/dmp.2025.10297.
Acknowledgments
This work was funded by a Global Health Pilot Grant from the Institute for Excellence in Health Equity of the New York University Grossman School of Medicine. During manuscript preparation, SM was supported by the National Institute for Environmental Health Sciences, K12 ES033594.
Author contribution
SM conceived and designed the study, obtained the data, conducted the analyses, and prepared and edited the manuscript. SW and CD advised on the analytic approach and results and edited the manuscript. LG co-conceived and co-designed the study and edited the manuscript.
Competing interests
SM has received funding for academic work from the National Disaster Medical System. The remainder of the authors declare no competing interests.
Open access
