Importation models for travel-related SARS-CoV-2 cases reported in Newfoundland and Labrador during the COVID-19 pandemic

1 Introduction

On January 31, 2020, due to an outbreak of SARS-CoV-2 in Wuhan, China, the World Health Organization (WHO) advised other countries to expect SARS-CoV-2 cases and be prepared for outbreak containment, but measures that would restrict travel and trade were explicitly not recommended (Grépin et al, 2021; World Health Organization, 2020c). State Parties were to notify WHO within 48 hours of the public health rationale and justification of measures that would significantly interfere with international traffic (World Health Organization, 2020c). Despite broad consensus prior to the pandemic that during a public health emergency travel measures significantly impacting travel and trade should not be implemented, during the initial phase of the COVID-19 pandemic most countries implemented such travel measures (Grépin et al, 2021; Piccoli et al, 2022; Shoichet, 2020). The next two years saw substantial variation in the implementation and strictness of travel measures between (Piccoli et al, 2022) and within (Reddy et al, 2021; Studdert et al, 2020) countries. In July 2021, WHO provided updated recommendations stating that international travel-related measures should be ‘proportionate to the public health risk’ and adapted to a country’s ‘specific epidemiological, health system and socioeconomic context’, and recommended a risk-based approach (World Health Organization, 2021).

The inconsistent implementation of travel measures during the first two years of the pandemic may have been due to the low quality of evidence to support policy. Systematic reviews (Burns et al, 2021; Grépin et al, 2021) report that travel measures may have had a positive impact on infectious disease outcomes, and reduced and delayed imported SARS-CoV-2 cases from Wuhan, but that overall the quality of evidence was low. Most evidence was due to modelling studies with a lack of ‘real world’ data (Burns et al, 2021), with inconsistent parameter estimates and assumptions, and that overlooked the impact of undetected cases outside of China (Grépin et al, 2021).

To better support modelling studies with ‘real world data’, and to consider the specific context in which travel measures are applied (World Health Organization, 2021), our analysis focuses on Newfoundland and Labrador (NL), a Canadian province that implemented travel measures during the COVID-19 pandemic. Our analyses contribute data, parameter estimates, and models to support decisions and inform best approaches to surveillance. In particular, we note there are few data and guidelines to inform travel measure implementation within countries during a pandemic. We develop models that consider data describing travel volume arriving in NL, infection prevalence at a traveler’s origin, and reported travel-related cases in NL. Three studies have previously considered all of these data types (Arnold et al., 2024; McCrone et al, 2022; Yang et al, 2021), but many importation modelling studies have had to proceed without all of these data sources available, for example, Godin et al (2021) and Steyn et al (2021) do not consider travel volumes. We develop different statistical models, and a mechanistic model to determine the impact of data gaps on the reliability of importation models.

Our analysis considers multiple sources of travel volume data, as many travel volume data sources have exclusions. For example, the OpenSky database and the Official Aviation Guide are the travel volume data sources used in Russell et al (2021), and these sources report only travelers that arrive by air, and not by land or sea. Exclusion of travelers arriving by some travel modes is just one type of exclusion that occurs in data sources, and we describe in detail the types of exclusions that occurred in the NL data in the Supplementary Information of this manuscript. We corrected the travel volume data for exclusions, and estimate how travel volumes to NL changed during the pandemic, and given the enacted travel measures.

During the COVID-19 pandemic, the government of NL reported travel-related cases, and the reliable public reporting of these data aided in the development of importation models. Travel-related case data was less reliably reported in the other Canadian provinces and the territories (see Hurford et al 2023). While post-arrival testing of travelers occurred in Canada, data describing infections detected in international travelers arriving in Canada was not rapidly reported or publicly available during the pandemic. Several studies have estimated the epidemiological risk due to imported infections without any data reporting travel-related cases (Hincapie et al, 2022; Linka et al, 2020; Menkir et al, 2021; Milwid et al., 2024; Russell et al, 2021), and other studies have used data that does not distinguish between travel-related and community cases (Chinazzi et al, 2020; Costantino et al, 2020; Hossain et al, 2020; Wells et al, 2020). To replicate the approach to importation modelling when travel-related case data is unavailable (i.e., Milwid et al. 2024), we formulate a mechanistic model that considers the processes that give rise to travel-related cases, but is not fit to the travel-related case data. This mechanistic modelling approach estimates general parameters from published studies, as an approach to overcoming the limitations arising due to data unavailability. The predictions of the mechanistic model are validated with the travel-related case data which are completely independent of the mechanistic model’s parameterization.

2 Methods

2.3 Modelling overview

2.3.1 Statistical models

We formulated nine statistical models with a linear model structure. The statistical models are generalized linear models with a Poisson error distribution, a logarithmic link, and where the response variable is either the daily travel-related cases reported in NL from Canada (Table 4, Models 1-6) or from international origins (Table 4, Models 1-3). The models have explanatory variables that are daily total travel volume from an origin (section 2.2.1), daily prevalence proportion at origin (section 2.2.2), or the product of daily prevalence proportion and travel volume for a given origin. For models to predict daily travel-related cases of Canadian origin, we considered stratification of all model variables for each Canadian province and the territories, and aggregation of variables for all of Canada (Table 4, Canada: Models 4-6). We fixed the model intercept as zero because if travel volume or prevalence proportion is zero, then no travel-related cases should be predicted.

View this table:

• View inline
• View popup
• Download powerpoint

Table 3. Variables and parameters that are estimated from the published literature or assumed for the mechanistic model

View this table:

• View inline
• View popup
• Download powerpoint

Table 4. Model selection for predicted daily travel-related cases reported in NL from Canadian and international origins.

Models where variables are stratified for Canadian provinces and the territories are denoted with ‘Provinces’ in parenthesis. Models were variables are aggregated for the Canadian provinces and the territories are denoted with ‘Canada’ in parenthesis.

2.3.2 Mechanistic model

The mechanistic epidemiological model predicts the number of travel-related cases of Canadian and international origin that would be reported in NL from September 2020 to June 2021 given travel volumes, prevalence proportion at a travelers’ origin, days since exposure for infected travelers, and testing policies in NL (see Table 3 for descriptions of variables and parameters). Unlike the statistical models which are fit to the data for travel-related cases, the mechanistic model is not fit, it is developed based on the processes typically represented in mechanistic importation models, with the number of travel-related cases predicted based on the testing of travelers that occurred in NL during the pandemic (see Figure 1).

• Download figure
• Open in new tab

Fig. 1.

An overview of the mechanistic model. The number of inbound infected travelers arriving on day t is the product of v_i(t): the number of travelers arriving from i on day t, and f_i(t): the prevalence proportion at i on day t summed over all travel origins, i. The days since exposure for infected travelers that arrive in NL depends on: a) positive pre-departure test results when such tests are required, b) the probability that symptom onset occurs before departure, and c) if the traveler was infected on the flight. Given a) and b) the distribution of days since exposure in arriving travelers are shown. Lastly, to be reported as a travel-related case in NL, the traveler must test positive on post-arrival test administered in NL. See Section 2.3.2 for complete details.

For travelers arriving in NL, Polymerase Chain Reaction (PCR) tests were mandatory for anyone symptomatic, for travelers that were passengers on flights for which exposure notifications were issued, and for rotational workers upon their return to NL (see Tables D6 and D5 in the Supplementary Information).

For travelers from Canadian origins, the predicted number of travel-related cases reported in NL at t is, where the sum i is across all Canadian provinces (except NL) and the territories, and where: is the number of regular travelers that are predicted to test positive and are tested due an exposure notification; is the number of regular travelers that are predicted to test positive and requested a test because they developed symptoms post-arrival; and is the number of rotational workers that are predicted to test positive on at least one mandatory post-arrival test.

For travelers from international origins, the predicted number of reported travel-related cases in NL is, which is similarly defined as equation 1 except that international travelers are not stratified by origin and rotational workers are not considered because, by definition, rotational workers do not return from international origins. The percentage of reported cases of each type (regular travelers and rotational workers of Canadian origin, and international travelers) predicted by the mechanistic model is shown in the Supplementary Information in Figure ??4)fig: (Table 4)e accuracy of the predictions of the mechanistic model are assessed by calculating the likelihood of R_CA(t) given the reported number of travel-related cases of Canadian origin, and the likelihood of R_INT(t) given the reported number of travel-related cases of international origin, and assuming a Poisson error distribution.

The mechanistic model requires information on days since exposure for infected travelers as this affects the PCR test sensitivity, and the timing of symptoms post-arrival, which affects when test results are reported. We assume that travelers that test positive due to an exposure notification were infected on their flight. For other travelers, we assume that the exposure time is uniformly distributed between 0 and 13 days prior to departure. This range was selected because any traveler infected more than 13 days prior to departure is unlikely to be infectious after arrival (Ferretti et al., 2020). We assume that a fraction of infected travelers originating from Canada and internationally, and all travelers that test positive on pre-departure tests, do not travel, which impacts the distribution of days since exposure for infected travelers, given that they have arrived in NL.

Regular travelers: symptomatic after arrival

Regular travelers (i.e., travelers that are not rotational workers) that departed from i, may have been reported as a travel-related case in NL on day t, if they developed symptoms post-arrival and requested a PCR test. The predicted number of such travel-related cases is, where is the number of infected regular travelers arriving from i, with an infection of age, a,where ā + t_r + t_rep is the number of days after the travellers arrival when the travel-related case would be reported, ρ is the probability that infected travelers have symptomatic infections, σ is the probability that travelers with symptoms request a test, and γ(a) is the probability of a positive test result for an infection that is a days since exposure. The mean time to develop symptoms after a traveler’s arrival is ā days (see section C.1 in the Supplementary information for the derivation). The time between requesting a PCR test and the test being performed is t_r days, and the time between the test and reporting of the results is t_rep days.

Regular travelers: exposure notification

Regular travelers that departed from i, may have been reported as a travelrelated case in NL on day t, if they were asked to complete PCR testing due to an exposure notification on their arriving flight. For the arrival dates corresponding to exposure notifications (see Table D6 in the Supplementary Information), the predicted number of such travel-related cases is, where is the number of regular travelers arriving from i, the number of days after the travelers’ arrival when the travel-related case would be reported is t_e + t_r + t_rep, the number of days after arrival when the exposure notification is issued is t_e, and is the probability of being infected on the flight is ϕ.

Equation (4) assumes that travelers who were exposed on flights were not infected prior to departure. While this many not always be the case, with few exposure notifications it is unlikely that our equations will result in substantial double counting of infected individuals because we counted these individuals both as infected during a flight, and infected pre-departure.

Rotational workers: mandatory testing

From the early stages of the pandemic, in NL there were specific postarrival testing measures that applied to rotational workers (see Supplementary Information, Table D5). Rotational workers were required to complete up to 3 post-arrival PCR tests. We define the number of rotational workers that test positive on their first post-arrival test occurring t₁ days after arrival from origin i, as, where is the number of infected rotational workers arriving from i, with an infection of age, a, where t₁ + t_rep is the number of days after the travelers’ arrival when the travel-related case would be reported.

For positive results on the second and third post-arrival tests, occurring t₂ and t₃ days after arrival, but not any post-arrival tests prior (i.e., for the third post-arrival test, but not testing positive on either the first or second post-arrival test) is, The predicted number of travel-related cases reported due to positive test results from rotational workers is calculated by summing the number of positive results for first, second, and third post-arrival tests (where applicable) that would be reported on day t, We do not consider tests for rotational workers based on developing symptoms when at least one post-arrival test was mandatory for rotational workers.

2.3.3 Statistical models representing data gaps

The first hypothetical data gap that we consider is ‘what if only one source of travel volume data were available’, rather than the three travel volume data sources that we combined to estimate total travel volume. A second hypothetical data gap that we consider is unavailability of data to estimate under-reporting coefficients. We develop additional statistical models to determine the accuracy of travel-related case predictions when just one source of travel volume data is considered, and when prevalence proportion is calculated without any corrections for under-reporting.

2.3.4 Model selection and validation

We used corrected Akaike Information Criteria (AICc) to determine which of the statistical models best predicted travel-related cases reported in NL without over-fitting. Model fit was assessed by calculating the likelihood ratio test statistic (calculated as −2 times the difference in the log likelihoods for the model and the null model), and assuming a Poisson distribution of error for each observation recorded at t. We validated the mechanistic model by calculating the likelihood of the model’s predictions, given the travel-related case data, and assuming a Poisson distribution for error for each observation recorded at t.

3 Results

We found that the total travel volume arriving in NL declined by 82% during the pandemic while travel measures were enacted (September 2020 to May 2021) compared to the same period a year prior (Fig. 2A; red line). Relative to the same 9 months one year prior to the pandemic, the average percentage of travelers arriving in NL increased from British Columbia (2.5 to 5.5%), Alberta (7 to 17%), the Canadian territories (0.3 to 3.3%), and international origins not the United States (8 to 15%; Fig. 3). The average percentage of travelers arriving in NL during the pandemic decreased from Ontario (27 to 21%), Québec (14 to 4%), and the United States (14 to 8%; Fig. 3).

• Download figure
• Open in new tab

Fig. 2.

A) The total travel volume arriving in Newfoundland and Labrador before and during the pandemic (red line, equation A1 in the Supplementary Information), and travel volumes reported by individual data sources: International Air Transport Authority (IATA, green line); Travel Declaration Forms (TDF; purple line); and Frontier Counts (FC; blue line). The numbered vertical lines correspond to travel measures implemented in NL (see Table 2). B) Total travel volume arriving from Canadian and international origins. C) and D) The average number of travelers arriving daily before and during the pandemic and their region of origin, where these averages are taken over 9 months.

• Download figure
• Open in new tab

Fig. 3.

Effect of the pandemic on the percentage of travelers arriving in NL from different origins. A) The percentage of travel volume each month from each of the 9 non-NL Canadian provinces, the Canadian territories (TR is all three territories combined), the United States, and non-United States international origins. B) The average percentage of arrivals from particular origins for 9 months during the pandemic (September 2019 to March 2020, and April to May 2019), and C) this same average for the same 9 months during the pandemic (September 2020 to May 2021).

Travel-related cases arriving in NL from Canada were best predicted by a statistical model with explanatory variables that were daily travel volume multiplied by daily prevalence proportion where both variables are stratified for each of the provinces of Canada of origin and the Canadian territories (Figure 4A), and this model was 532 times more likely than the null model (Table 4). The mechanistic model was validated with the travel-related cases of Canadian origins (Figure 4B). The higher AICc score for the mechanistic model is because the mechanistic model is not fit to the travel-related case data and has a lower log likelihood.

• Download figure
• Open in new tab

Fig. 4.

Reported travel-related cases as predicted by the best model (model 1 in Table 4) and the mechanistic model (lines). The shaded region is a 95% prediction interval corresponding to Poisson-distributed uncertainty in the predictions with no uncertainty in the parameter estimates. Data are shown as black dots.

Travel-related cases arriving in NL from international origins were best predicted by a statistical model with an explanatory variable that describes travel volume (Figurec 4C), and this model was 653 times more likely than the null model (Table 4). The mechanistic model was validated with the travel-related cases of international origins

(Figure 4D), and had the second lowest AICc value (Table 4). The fitted coefficients for the best models (both Canada and international) are reported in Section ?? of the Supplementary Information.

Only the TDF travel volume data source reports daily travel volume, so to determine the effect of using just one travel volume data source it was necessary to fit to the travel-related case data aggregated by month (9 observations). Results are reported in Table F7 of the Supplementary Information, however, generally, the lack of daily travel volume estimates (only one per month) reduced our ability to formulate and test the models. We found that correcting the prevalence proportion for under-reporting had no effect on model predictions (Table F8 in the Supplementary Information). This result applied even to models that were stratified for travelers originating from each non-NL Canadian province and the territories. Correcting the prevalence proportion by considering underreporting in each non-NL Canadian province and the territories affected the estimated model coefficients, but not the model fit.

4 Discussion

In July 2021, the World Health Organization (WHO) updated guidelines for international travel (World Health Organization, 2021) to advise that local epidemiology, and public health and health system performance and capacity should be considered to determine if travel measures are appropriate during a pandemic. Analysis of regional data and regionally-specific models are needed because decisions of whether travel measures are appropriate depend on regional characteristics.

We estimated that travel measures (described in section A.2 of the Supplementary Information), reduced the total travel volume arriving in NL by 82% (September 2020 - May 2021) compared to the same 9 months prior to the pandemic (Figure 2A, red line). This finding is similar to an estimated 79% reduction in non-resident visitor volume to NL for 2020 compared to 2019 (Government of Newfoundland and Labrador, 2020). In other countries, air travel is estimated to have declined, on average, 63% for May 2020 (during the pandemic) as compared to May 2019 (before the pandemic; Russell et al 2021). The more substantial reduction in travel to NL may have been due to stricter travel measures for entry to NL than for other countries and regions during the study period. Most notably, non-NL residents entering NL from other regions of Canada were required to complete Travel Declaration Forms (TDFs) and have a self-isolation plan to submit to a government representative at entry. These results help quantify the impact of travel measures implemented during a pandemic and modelling that predicts clinical cases depend on assumptions describing how travel measures affect travel volume (Hurford et al, 2021).

We found that during the pandemic there was a 10% increase in arrivals from Alberta, and a 10% decrease in arrivals from Québec (Fig. 3) to NL. Evidence for this result can be seen in Table A3 of the Supplementary Information, where the percentages before the pandemic (January 2019 - March 2020) are calculated from the International Air Transit Authority (IATA) data, and the percentages during the pandemic (September 2020 - May 2021) are calculated from the TDF data. This shift in the percentage of travelers arriving from different provinces may be explained by the high percentage of NL’s rotational workers that work in Alberta (57%) as compared to Québec (2%; Table A3 in the Supplementary Information), while the travel volume of regular travelers from Québec before the pandemic is relatively high (11-26%, third highest behind Ontario and Nova Scotia, Table A3 in the Supplementary Information). Rotational workers likely continued working during the pandemic, while regular travelers may have delayed or canceled trips. Importation models used during public health emergencies often consider travel volume data. An implication of our findings is that the travel volume data used for these calculations needs to be reported daily, otherwise the type of importation modeling that can be done is very limited (i.e., cannot be stratified by province of origin), and that travel volume data collected before the pandemic, may not accurately represent the proportion of travelers arriving from different travel origins. It is necessary to consider travel volume and prevalence proportion estimates that are stratified by Canadian province of origin, as when these data are not considered the AICc score is reduced by 191.6 (Table 4, Model 1 vs. Model 4).

Our results provide insights into whether mechanistic importation models that are formulated independently of travel-related case data (i.e., Milwid et al. 2024 and Hincapie et al 2022) can be relied upon. The mechanistic modelling approach is to mathematically describe the process that give rise to imported cases, and to estimate parameters from the published literature. The predictions of mechanistic models (Hincapie et al, 2022; Linka et al, 2020; Nakamura and Managi, 2020; Russell et al, 2021) have not been tested or validated, and so the accuracy of this modelling approach is not known. Although our mechanistic model was not fit to the travel-related case data, the predicted number of travel-related cases was still reasonably accurate (Figure 4A and C), although overall models that are fit to travel-related case data perform better (Table 4). One reason that the mechanistic model preformed this well is because the mechanistic model’s predictions are sensitive to the number of travel-related cases that are reported due to testing of passengers on flights that were subject to exposure notifications. We assumed that the probability of infection during a flight is ϕ = 0.01, but there is no data to inform this estimate (Table 3).

There are number of reasons why the predictions of mechanistic importation models may not hold and should be validated. Many studies consider only air travel volumes (Chinazzi et al, 2020; Hincapie et al, 2022; Nakamura and Managi, 2020; Russell et al, 2021; Wells et al, 2020), although travelers are likely to arrive via other travel modes and this proportion may change seasonally. No one travel volume data source reports the epidemiologically-relevant travel volume due to their scope and exemptions. Travel documents completed at international borders omit travelers that originate from within the country, and not all travel between countries requires such documents to be completed (i.e., travel between countries in the European Union). Provinces may conduct surveys to understand tourist preferences, but such surveys (i.e., Government of Newfoundland and Labrador 2020-2021) usually focus only on non-resident travel, while returning residents and workers are a potential source of imported infection. We combined four data sources (see section A in the Supplementary Information) to overcome limitations of each travel volume data source and to estimate total travel volume; however, this was very time-consuming, and to model importations urgently during a public health emergency, it is necessary that daily total travel volume data is available in real time.

There are some differences between the predictions of the best statistical models and the reported travel-related cases (Figure 4). Arnold et al. (2024) performs statistical modelling to describe the number of cases reported in international travelers isolating in government-managed quarantine facilities in New Zealand during the COVID-19 pandemic. In Arnold et al. (2024), countries of origin that have low numbers of arrivals, low numbers of cases, or both, are referred to as ‘low information countries’. In our study, the differences between the statistical model’s predictions and the reported travel-related cases in NL are likely due to low number of cases and low travel volumes, particularly for travel-related cases of international origin. Nonetheless, importation modelling, and data of the type analyzed in this study are rare in the published literature, and the model fits shown in Figure 4 demonstrate that the model’s predictions have good correspondence with reported travel-related cases.

The statistical modelling approach predicts travel-related cases more accurately than the mechanistic modelling approach (Table F7), but it can be difficult to give a meaningful epidemiological interpretation to the fitted coefficients when the statistical modelling approach is used. In our study, the best model to predict travel-related cases of Canadian origin is a generalized linear model that considers the product of travel volume and prevalence proportion in each non-NL Canadian province and the territories. Given the fitted coefficients, travel volume and prevalence proportion in Québec contributes negatively to travel-related cases reported in NL, while travel volume and prevalence proportion in Saskatchewan and Manitoba contribute positively (Supplementary Information, Section E). For an epidemiological perspective, it is puzzling as to what the negative coefficient estimated for Québec could mean, yet, for what the statistical model lacks in meaningful coefficient interpretation, it makes up for with more accurate predictions of travel-related cases. We could constrain the fitted statistical model coefficients to be positive, but this would be one step towards building a mechanistic model, and rather than build statistical models with some mechanistic components, we built a fully mechanistic model representing the importation pathways described in Figure 1, and with the parameters estimated to correspond with their published epidemiological values (Table 3).

Regarding forecasting of imported infections arriving in a jurisdiction during a pandemic, the appropriate modelling approach depends on data availability. If travel-related case data reported in the destination jurisdiction are available, then statistical modelling is a good option. Surprisingly, our study found that when infection prevalence is an explanatory variable, if it is not possible to correct prevalence for under-reporting, i.e. unavailability of seroprevalence data, this may not matter (Supplementary Information, Table F8). If travel-related case data in the destination jurisdiction are not available, then mechanistic modelling is the only possible approach. However, even when travel-related case data are available, it is good practice to develop both a mechanistic model and statistical models, to contribute to our knowledge of how to best model importations, to better understand the processes that predict imported infections, and so we can avail of the strengths of each approach.

Testing of travelers at points of entry is a core surveillance method to support early warning for change in epidemiological patterns (World Health Organization, 2022). Early in the COVID-19 pandemic, WHO situation reports described transmission classifications as imported cases only, sporadic cases, clusters of cases, local transmission, and community transmission where this information was self-reported by the State Parties (World Health Organization, 2020a,b). We do not consider it necessary for all regions to report the number of cases with a travel history (see also Martignoni et al 2024), but reporting of data describing the number of travel-related cases in a jurisdiction that is not experiencing widespread community transmission is necessary to support the development, and refinement, of importation models.

During the COVID-19 pandemic many countries implemented some type of travel measures. The unprecedented implementation of travel measures during the initial phase of the COVID-19 pandemic, may have been because COVID-19 was more challenging to detect and contain than H1N1 influenza A (Anderson et al, 2020; Grépin et al, 2021). Guidelines that recommended against the implementation of travel measures may have overlooked the value of coordinated implementation of travel measures and community non-pharmaceutical interventions as occurs for an elimination strategy (Grépin et al, 2023; Martignoni et al, 2024). Generally, data, models, and analysis are needed to inform decisions to implemented travel measures in a region, and our results aim to support decision-making and inform future approaches to model development.

Declarations