2.1. Countries Targeted in the Analysis, Disease Introduction Data, and Viral Genetic Data

According to the Global Initiative on Sharing Avian Influenza Data (GISAID, https://www.gisaid.org/), the global network of expertise on animal influenzas (OFFLU, https://www.offlu.org/) of the World Organization for Animal Health (WOAH), and the Food and Agriculture Organization (FAO) of the United Nations (UN) and the literature, 31 countries were affected by clade 2.3.2.1c (H5N1 subtype) between the third quarter of 2009 and second quarter of 2017. The viruses circulated mainly in western Africa, Asia (south-eastern, eastern, southern and western), and eastern Europe. According to the same sources, 55 countries were affected by clade 2.3.4.4b (subtypes H5N1, H5N2, H5N5, H5N6, and H5N8) between the first quarter of 2013 and the first quarter of 2018. The viruses circulated mainly in Europe (southern, western, and northern Europe [47–57]).

Outbreak data (grouped by events) submitted by national authorities to WOAH did not have clade information, in the timeframe considered for this analysis. Dates and locations of (i) the events reported to WOAH and (ii) the positive samples described in the sources mentioned above were compared. Based on that, we selected the outbreaks compatible with the clades studied in this analysis. When several clades were cocirculating, the same outbreaks were considered compatible with each of the cocirculating clades. According to data submitted by national authorities to WOAH, these countries/territories reported, for the period of analysis, about 2300 outbreaks compatible with clade 2.3.2.1c and about 6500 outbreaks compatible with clade 2.3.4.4b (Figure 1). Based on this data, for each clade and each subtype in each country, the first reporting of HPAI was considered as the first introduction. Also, based on HPAI seasonality, reporting after at least 1 year of absence was considered as re-introduction. The list of introductions/reintroductions defined according to these criteria was used as the binary-dependent variable in the GLMM described below. All the countries affected by at least one of the two clades were included in the GLMM analysis.

For phylogeographic inference (discrete trait analysis [DTA]—see details below) coupled with a GLM, the viral genetic data of the HA segment were retrieved from the epi-based dataset used by Fusaro et al. [49]. Starting from a collection of HA genes for which the number of sequences by geographic area under study was generally proportional to the number of reported outbreaks in that geographic area, Fusaro et al. [49] applied three different subsampling procedures to the sequence data in order to mitigate sampling bias. As it was found that the overall migration pattern is consistent across the downsampling strategies, we opt to use the selection of representative sequences by different epidemiological characteristics that have the most balanced distribution of samples among locations and hosts. For further details on the data set compilation strategy, we refer the reader to the Supporting Information provided by Fusaro et al. [49]. It corresponds to 23 countries/territories for clade 2.3.2.1c and 28 countries for clade 2.3.4.4b (Figure 1).

2.2. Poultry Trade Flows

An exercise was first organized with selected international experts to achieve consensus on poultry and poultry products likely to spread HPAI through international trade, and they recommended to consider the following categories: (1) more than 1-day-old chickens, turkeys, quails, guinea fowls, pheasants, ducks, geese, ostriches, emus (highest likelihood); and (2) chickens, turkeys, quails, guinea fowls, pheasants, ducks, geese, ostriches, emus of 1-day-old or after hatching (lower likelihood). United Nations (UN) Comtrade [58] centralizes data on international trade, provided by exporting and importing countries and territories. Data were extracted for the period from 2009 to 2018, for the countries and territories in Africa, Asia, and Europe, and for Canada with a final update as of April 10, 2020. For the same trade exchange, when information on price and quantity was provided by both exporting and importing countries, the data provided by the importing country were considered more accurate [59]. Quarterly data for chickens, ducks, geese, guinea fowls, and turkeys were available for the period between 2010 and 2018. In the article, based on the categories available in Comtrade, these animals were stratified into the following subcategories: “chicken lighter than 185 g”, “poultry other than chicken lighter than 185 g”, “chicken of 185 g or more”, and “poultry other than chicken of 185 g or more”. Quarterly data for hatching eggs were only available for the period between 2012 and 2018 and were grouped under the names “chicken hatching eggs” and “hatching eggs of poultry other than chicken”. Annual data for live poultry were available for the entire study period.

Estimation of missing quarterly quantity data for the period from 2009 to 2018 was performed as follows: 1) when only the monetary value was available, the estimate of missing quantity data was performed using the methods described in Awada et al. [60]. (2) Moreover, when only a yearly value was available, the estimate of missing quarterly data was performed based on the quarterly average of quantity percentage traded between the same pair of countries, for the same poultry subcategory, during other years. In case the pair of countries had not traded the same poultry subcategory in other years, the yearly quantity was equally distributed among all quarters. (3) Finally, to estimate missing quarterly data for hatching eggs from 2009 to 2011 (no data available in Comtrade), the average of quantities traded between countries during other quarters of the analysis (between 2012 and 2018) was computed, and the best fitting regression of the values through time was exponential (R² = 0.80). Therefore, the exponential function was applied for quantities traded between each pair of countries during other quarters. Such corrections were used in previous publications [60] and are consistent with methods described by the United Nations Conference on Trade and Development [61], as well as with the method later developed in the context of the UN Comtrade upgrade plan in 2019 [62]. In the models described below, trade quantities were centered and scaled. The regression residuals were normally distributed after this transformation.

The transformation of quantitative variables into centered and scaled variables enables the effects of different predictors to be compared, independently of the quantitative units of the different variables.

2.3. Migratory Bird Flows

First, the main migratory wild bird species potentially involved in HPAI international spread were identified. As mentioned above, national authorities provide WOAH, through its World Animal Health Information System (WAHIS), information on diseases including the species affected in disease outbreaks [3]. The list of wild bird species affected by HPAI was extracted from WAHIS. Out of these species, only the ones that were (i) categorized as migratory, (ii) whose range overlaps the main study area (Africa, Asia, or Europe) and (iii) reported as affected by HPAI in at least two countries, were considered as potentially involved in HPAI international spread. For 35 species, all these criteria were fulfilled. The list was completed by four additional high-risk waterfowl species listed in a review paper [63]. The final list of the 39 wild bird species potentially involved in HPAI international spread is presented in Tables S1 and S2. Second, breeding and nonbreeding sites of the 39 wild bird species listed above in the geographical scope of interest were extracted from the International Union for Conservation of Nature (IUCN) website [64], and each country of interest was tagged for having or not having breeding or nonbreeding sites. Finally, for each species, a table with two columns was constructed. All possible pairs of country breeding site/country with nonbreeding were included in the table. The temporality of the flows was estimated based on the corresponding migratory periods in both directions (based on a literature review and expert advice; Table S1). To estimate the size of migratory bird populations in flows, this dataset was finally compiled into a matrix, weighted with the sum of the average centered and scaled global population estimates for species migrating during a given quarter, between 2009 and 2018 (details are available in Table S2). The migration periods and global population estimates during a given semester were considered stable across the years, not taking into consideration variations due to changing environmental and climatic conditions.

2.4. Preventive and Surveillance Measures Implemented by Veterinary Services Over Time and Resources Allocated to Veterinary Services

Surveillance and preventive measures data were obtained from WOAH. These data are submitted to WAHIS by the national authorities of 183 Member Countries that have the legal obligation to report information concerning high-impact animal diseases listed by WOAH, and more than 20 additional countries and territories that provide information on a voluntary basis [3]. This annual list of diseases contains about 120 diseases, of which HPAI [65]. Six-monthly data were extracted for the period from 2009 to 2018, for the 67 countries and territories targeted by the analysis, with a final update as of April 10, 2020. Measures implemented during a given semester were considered valid for both quarters within this time range. These data were for (a) active surveillance in poultry for HPAI (including targeted surveillance and screening, as defined by WOAH); (b) passive surveillance in poultry for HPAI (including disease official notification, general surveillance, and monitoring, as defined by WOAH); (c) any surveillance in wild birds for HPAI (including all the measures listed above); (d) precautions at borders against HPAI, and (e) preventive vaccination against HPAI (considering that, based on previous studies, vaccinated poultry can also be a source of spread, as vaccination can in some cases mask clinical signs without preventing infection [19, 66, 67]). Based on the WOAH definition, precautions at borders include “measures applied at airports, ports, railway stations or road checkpoints open to international movement of animals, animal products and other related commodities, where import inspections are performed to prevent introduction of the disease, infection or infestation into a country/territory or zone” [3]. The variables used for these predictors were binary (“1” when implemented and “0” when not implemented), in each country and for a given quarter. Ten percent of the data were missing in WOAH information systems and were estimated, based on information provided for the previous years and subsequent years, with the hypothesis that countries and territories maintain some consistency in the surveillance and control policies implemented. Additionally, gross domestic product (GDP) per capita was used as a proxy for the resources allocated to veterinary services. Data were extracted from the World Bank [68]. In the models described below, GDP per capita was centered and scaled. The regression residuals were normally distributed after this transformation.

2.5. GLMM for Disease Introduction and Reintroduction

For each of the two clades considered in each country, the list of introductions/reintroductions defined in Section 2.1 was used as the binary dependent variable. The detailed description of the GLMM for disease introduction and reintroduction is presented in Section S1.

The full model was tested for each clade, then variables were removed progressively, comparing the values of the Akaike information criterion (AIC), preferring the model with the lowest AIC value. The final model was selected for each clade according to the AIC criterion. After this step, collinearity between statistically significant predictors was tested with variable inflation factors (VIFs), and no collinearity issue was detected. A GLMM approach was selected because our data consisted of repeated measures of imports from countries over time. R version 3.5.1. and package “glmmTMB” were used for the GLMM, with i (country exposed) as a random variable and “binomial” distribution family [69, 70]. The dependent variable, predictor variables, and units used in the model are summarized in Figure S1a.

2.6. Phylogeographic Inference Coupled With GLM With Time-Dependent Coefficients for Viral Spread

We performed a Bayesian phylogeographic inference using the discrete diffusion model [71], which is also referred to as a DTA. For each clade, a set of 1000 evenly spaced post-burn-in phylogenies were selected from the epianalyses reported by Fusaro et al. [49]. These served as empirical tree distributions [32] for reconstructing the history of spread in discrete space under a time variable GLM [31] in BEAST v.1.10 [72] with BEAGLE v.3 [73] for improved computational speed. The GLM model was coupled to a model averaging procedure (Bayesian stochastic search variable selection, BSSVS) that provides a test to identify significant predictors. To ensure that the Bayes factor (BF) support estimates are not driven by the relative abundances of samples by location, the BF support was estimated with the default (BF_default, [32]) as well as the adjusted (BF_adjusted, [74]) setup. Maximum clade credibility (MCC) summary trees were obtained with TreeAnnotator 1.10 distributed with BEAST v1.10 and visualized with FigTree [75]. Convergence and mixing properties were inspected using Tracer v.1.7 [76]. The detailed description of the GLM coupled with phylogeographic inference with time-dependent coefficients for viral spread is presented in Section S2.

The periods covered by the predictors are 2009–2017 for clade 2.3.2.1 and 2013–2018 for clade 2.3.4.4. The dependent variable, predictor variables, and units used in the model are summarized in Figure S1b.

Effect size (with lower and upper 95% highest posterior density interval estimates) is a measure of the size of mean differences or strength of associations. BF adjusted of 25 was considered as cutoff for significance. After this step, collinearity between predictors with BF_adjusted equal to or higher than 25 was tested with VIF, and no collinearity issue was detected. A positive effect size indicated a positive association (when this variable increases, the spread of HPAI increases), and a negative effect size indicated a negative association (when this variable increases, the spread of HPAI decreases).

2.7. Counts of Quarters, Countries, and Number of Sequences for Each Model and Clade

Counts of quarters, countries, and number of sequences for each model and clade are presented in Table S5. For the GLMM, the number of observations was nearly equivalent for clades 2.3.2.1c and 2.3.4.4b. For the phylogeography-informed GLM, the total number of sequences considered for the two clades were comparable. Finally, from the counts of countries sequences by world region, we see that for clade 2.3.2.1c, Asia is the region most represented in the clade distribution (with more than 50% of countries and sequences from this region), and Europe for 2.3.4.4b.

3.1. GLMM for Disease Introduction and Reintroduction

3.2. Phylogeographic Inference Coupled With GLM With Time-Dependent Coefficients for Viral Spread

As explained above, HPAI viral migration was only analyzed for countries associated with sufficient viral sequences after the subsampling step by Fusaro et al. [49]. The evolutionary relationship between viral lineages from these countries is reported through the spatially annotated time-scaled phylogenetic tree obtained through the DTA and displayed in Figure S2.

For the GLM, results for the factors that influence HPAI spread between countries for each of the clades of interest are presented in Table S4.

Comparative results of (a) the GLMM for disease introduction and reintroduction and (b) phylogeography-informed GLM are presented in Table S6.