key: cord-0321298-j88ky1y8
authors: Robert, A.; Kucharski, A. J.; Funk, S.
title: The impact of local vaccine coverage and recent incidence on measles transmission in France between 2009 and 2018
date: 2021-06-01
journal: nan
DOI: 10.1101/2021.05.31.21257977
sha: 2a4331a50efa33c3dbaed245d325b0a8e7b49b13
doc_id: 321298
cord_uid: j88ky1y8

Background Despite high levels of vaccine coverage, sub-national heterogeneity in immunity to measles can create pockets of susceptibility, which are hard to detect and may result in long-lasting outbreaks. The elimination status defined by the World Health Organization aims to identify countries where the virus is no longer circulating and can be verified after 36 months of interrupted transmission. However, since 2018, numerous countries have lost their elimination status soon after reaching it, showing that the indicators used to define elimination may not be predictive of lower risks of outbreaks. Methods and Findings We quantified the impact of local vaccine coverage and recent levels of incidence on the dynamics of measles in each French department between 2009 and 2018, using mathematical models based on the 'Epidemic-Endemic' regression framework. High values of local vaccine coverage were associated with fewer imported cases and lower risks of local transmissions. Regions that had recently reported high levels of incidence were also at a lower risk of local transmission, potentially due to additional immunity accumulated during these recent outbreaks. Therefore, all else being equal, the risk of local transmission was not lower in areas fulfilling the elimination criteria (i.e., low recent incidence). After fitting the models using daily case counts, we used the parameters' estimates to simulate the effect of variations in the vaccine coverage and recent incidence on future transmission. A decrease of 3% in the three-year average vaccine uptake led to a five-fold increase in the number of cases simulated in a year on average. Conclusions Spatiotemporal variation in vaccine coverage because of disruption of routine immunisation programmes, or lower trust in vaccines, can lead to large increases in both local and cross regional transmission. The association found between local vaccine coverage and incidence suggests that, although regional vaccine uptake can be hard to collect and unreliable because of population movements, it can provide insights into the risks of imminent outbreak. Periods of low local measles incidence were not indicative of a decrease in the risks of local transmission. Therefore, the incidence indicator used to define the elimination status was not consistently associated with lower risks of measles outbreak in France. More detailed models of local immunity levels or subnational seroprevalence studies may yield better estimates of local risk of measles outbreaks.

measles in each French department between 2009 and 2018, using mathematical models based on the 23 'Epidemic-Endemic' regression framework. High values of local vaccine coverage were associated with 24 fewer imported cases and lower risks of local transmissions. Regions that had recently reported high 25 levels of incidence were also at a lower risk of local transmission, potentially due to additional immunity 26 accumulated during these recent outbreaks. Therefore, all else being equal, the risk of local 27 transmission was not lower in areas fulfilling the elimination criteria (i.e., low recent incidence). After 28 fitting the models using daily case counts, we used the parameters' estimates to simulate the effect of 29 variations in the vaccine coverage and recent incidence on future transmission. A decrease of 3% in the 30 three-year average vaccine uptake led to a five-fold increase in the number of cases simulated in a year 31 on average. 32

To do so, we implemented an Epidemic-Endemic time-series model using hhh4, a framework developed 75 by Held, Höhle and Hofmann to study the separate impact of covariates on importation, cross-regional 76 transmission and local transmissions on aggregated case counts [14, 15] . We adapted this framework 77 to daily case counts and applied it to the daily number of measles cases per department (NUTS3 levels) 78

in France reported to the European Center for Disease Prevention and Control (ECDC) between January 79 2009 and December 2018. We computed the average values of vaccine uptake and the number of cases 80 per department in the past three years to mimic the timeframe used to define the elimination status, 81 and modelled their impact on the local risks of outbreaks. 82

Methods 83 Description of the hhh4 framework 84 We used the modelling framework implemented in the "hhh4" model, which is part of the R package 85 "surveillance" [15] , to analyse infectious disease case counts. All the notations are defined in Table 1 . The predictors , , , and , are independently impacted by different covariates, i.e., a covariate 109 may be associated with a reduction of importations, but have little impact on the spread of the virus 110 within the region. We assume that , , the number of observed cases at in , follows a negative 111 binomial distribution to allow for overdispersion [16] . The overdispersion parameter is estimated. 112

The predictors , , , and , are estimated using log-linear regressions. For each predictor, we 113 estimate: i) The intercept (identical across spatial units), and ii) the vector of coefficients associated 114 with , the vector of covariates at in included in each component. 115 The observed case counts , was computed from 14,461 cases (10,988 confirmed and 3,473 probable 121 cases) routinely collected in metropolitan France, and reported to the ECDC between January 2009 and 122

December 2018 ( Figure 1A ). This data was retrieved on The European Surveillance System (TESSy) on 123 22 January 2019. The cases were stratified by the metropolitan department they were reported in. The 124 department correspond to French NUTS3 regions. We excluded three cases where this information was 125 not available. We used the date of symptom onset reported for each case to compute the daily number 126 of cases from 2009 to 2018 per department. 127 step is close to the average serial interval, cases of the same generation of transmission can be assumed 141 to be roughly grouped together in the same time point [18] . Nevertheless, studying weekly (or 142 fortnightly) aggregated cases counts does not reflect the distribution of the serial interval (i.e., it ignores 143 overlapping generations of transmission because of shorter or longer delays between primary and 144 secondary cases). This can lead to directly connected cases being grouped in the same time step, or 145 separated by more than one time step. This aggregation also ignores the potential for unreported cases, 146 which may lead to cases causing transmission two to three weeks after their onset date via an 147 intermediate, unobserved case. Finally, the starting date of aggregation influences how cases are 148 grouped, which can lead to discrepancies in the parameter estimates. 149

Recent developments in the surveillance package included weight estimation to represent the relative 150 impact of previous time steps on the number of cases at [19] . Since we are using daily case counts, 151

we set the weights of the different time steps from the distribution of the serial interval. We computed 152 . Only a subset of measles 154 cases are reported to the surveillance system [20] , therefore we accounted for the risks of unreported 155 cases by computing a composite serial interval from three different transmission scenarios ( Figure 1B) : 156 1-In case of direct transmission between two cases and , the number of days between the two 157 cases 1 ( ) follows a Normal distribution truncated at 0: 1 ( )~(11.7, 2) [17]. 158 2-In case of unreported cases between and , the number of days between the two cases 2 ( ) 159 follows a Normal distribution truncated at 0: 2 ( )~(23.4, √8). This distribution corresponds 160 to the convolution of 1 ( ) with itself. 161 3-If and share the same unreported index case, the number of days between and follows a 162 half-Normal distribution (excluding 0) of standard deviation √8 days. This distribution 163 corresponds to the distribution of the difference of 1 ( ) with itself, excluding values below 1. 164

We added this last scenario to account for multiple concurrent importations stemming from 165 an unreported infector. 166

We considered that 50% of the composite serial interval reflected direct transmission (scenario 1, 167 without missing generations between cases), and 50% came from the two scenarios with unreported 168 cases (scenarios 2 and 3). The distribution of the composite serial interval is shown in Figure 1B . We ran 169 sensitivity analysis to estimate the parameters of the model using composite serial intervals computed 170 with different proportions of direct transmission, and observed it had little influence on the estimation 171 of each parameter (Supplement Section 1). 172

In the hhh4 framework, the average number of cases caused in the department at time by cases 174 from another department is quantified by the neighbourhood component. It is equal to ϕ i,t * * 175 , −1 (Equation 1). Therefore, the number of cases caused by cases from in in hhh4 is influenced by 176 three factors: 177

• The susceptibility of the department , quantified by the neighbourhood predictor ϕ i,t , defined 178 as ( , ) = ( ) + ( ) * ( ) . 179

• The number of connections from to , calculated using an exponential gravity model [21] , 180 whereby the number of connections between and is proportional to the product of the 181 number of inhabitants in the department of origin , the department of destination and 182 an exponential decrease in the distance between and . Therefore, the number of 183 connections from to was calculated as = e − . 184

• The proportion of the population in that is infectious. 185 Therefore, the average number of cases expected from department to department at can be 186 written as the product of these three factors: Since 26% of the entries in the coverage dataset were missing, we ran a beta mixed model to infer the 225 missing values. We used the time and squared time (in years) as covariates, and random effects 226 stratified by department. We used the average prediction to infer the missing values from the fitted 227 model and get the complete vaccine coverage dataset. More details on the regression, and the 228 sensitivity analyses that were run are presented in the Appendix (Supplement Section 2). All values of 229 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The level of recent incidence was a covariate in all three components. 256 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 1, 2021.

In the subsection "Connectivity between departments", we discussed the impact of the number of 258 inhabitants on the number of movements between departments. Furthermore, several studies have 259 indicated a potential association between the population density and the number of secondary 260 transmissions [31] [32] [33] . Therefore, we controlled for the impact of the number of inhabitants in each 261 department, and the surface area (i.e., the geographical size) on the number of local transmissions. 262

The log-number of inhabitants log ( , ) in the department at time was added as a covariate in all 263 three components. The log-surface of the department log ( , ) was added as a covariate in the 264 autoregressive component. We studied four scenarios: i) Using the latest local values of coverage (averaged over the past three 315 years), population and category of recent incidence, ii) Increasing the vaccination coverage in each 316 department by three percent, iii) Decreasing the vaccination coverage in each department by three 317 percent, and iv) setting the recent incidence in each department to minimal levels (i.e. conditions 318 fulfilling the WHO elimination status requirements). 319 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The association between the level of incidence over the past three years (parameters:

1 and 344 importations, but no difference was noticeable in cross-regional or within-region transmission. 357

The other covariates included in the model showed that the number of inhabitants in a department 358 had an important impact on both the endemic and neighbourhood components: departments with 359 more individuals were more likely to report background importations and cross-regional transmission. 

Using the mean parameter estimates, and the latest values of vaccination coverage, incidence, and 376 number of inhabitants per department, we computed the local predictors , , and in both models 377 to highlight the spatial heterogeneity of the transmission risks (Figure 3) . The predictors were computed 378 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. the North of France were also at risk of secondary transmission despite higher vaccination coverage. 387

As expected, the overall number of baseline importations in Model 1 was lower than in Model 2, which 388 was compensated by a higher number of cross-regional transmissions (Figure 3 ). This shows that some 389 of the cases that could not be linked to local transmission, or transmission between neighbouring 390 departments in Model 2, were classified as cross-regional transmissions in Model 1, which would 391 indicate long-distance transmission events. In both models, departments with a higher number of 392 inhabitants were most at-risk of cross-regional and baseline importations, which corresponds to the 393 strong association between the number of inhabitants and the endemic and neighbourhood 394 components highlighted in Figure Under the latest measures of coverage and incidence, the simulated outbreaks display a wide variation 454 in the number of cases in 2019 (minimum 100 cases, median 1,100 cases, maximum 11,100 cases). 455

Active transmission was generated in a wide range of departments. Indeed, across the simulation set, 456 44 of the 94 French departments reported more than 10 cases in at least 25% of the simulations. There 457 was noteworthy spatial heterogeneity in the levels of incidence. Indeed, in 12 departments, there was 458 no case generated in more than half of the simulations (Figure 5, top right panel) . The departments 459 most vulnerable to active transmissions were highly populated urban areas, such as Paris, the Bouches-460 du-Rhône, and the North of France. Because they are highly populated, these departments were 461 susceptible to repeated importations (they reported at least 1 case in more than 95% of the 462 simulations), which could then cause large transmission clusters. This was especially evident in the 463 South-East of France, where we highlighted that the number of secondary cases per case in the 464 department was among the highest in the country (Figure 3 and Figure 5 ). Numerous departments 465 were affected by large outbreaks in a subset of the simulated datasets: 27 departments reported more 466 than 50 cases in at least 5% of the simulations ( Figure 5) . Further, at least one major outbreak was 467 generated in the majority of the simulations: in 55% of the simulations, one department reported more 468 than 100 cases (the most commonly affected department were Paris and its surroundings, the Nord, 469

and Bouches-du-Rhône). 470

Decreasing the average three-year vaccine coverage by three percent led to an important increase in 471 the number of cases per outbreak (median 4,900 cases, more than 95% of the simulations resulted in 472 more than 1,000 cases). This was first due to an increase in importations and cross-regional 473 transmission: all 94 departments had at least one case in more than half of the simulations, 77 in at 474 least 90% of the simulations. Furthermore, the decrease in vaccination coverage resulted in higher 475 chances of uncontrolled transmissions in many departments ( Figure 5, third row) . On the other hand, 476 increasing the vaccine coverage by three percent caused an important drop in the number of cases 477 (median 605 cases, 80% of the simulations generated less than 1,000 cases), caused by both a decrease 478 in the number of importations, and in the potential for secondary transmission following importations. 479

Although outbreaks were still punctually generated, these events are much rarer than in the other two 480 simulation sets: in 25.8% of the simulations, at least one department generated more than 100 cases 481 (54.1% with the baseline scenario, 95.4% when we reduced the local vaccine coverage). 482 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 1, 2021. ; https://doi.org/10.1101/2021.05.31.21257977 doi: medRxiv preprint Finally, setting the local recent incidence to the minimum level in each department, which would fulfil 483 the elimination guidelines, had two opposite effects: it led to a decrease in the number of importations 484 and cross-regional transmission, and an increase in the number of infections within each department 485 ( Figure 2 ). In this simulation set, the number of departments where no cases were generated in more 486 than half of the simulations was similar to when the vaccine coverage was increased (24 departments 487 in this simulation set, 29 when the vaccine coverage was increased, Figure 5) , which shows the 488 reduction in the number of cross-regional transmission and background importations. Conversely, the 489 number of large outbreaks was only marginally inferior to the reference simulation set: in 44% of the 490 simulations, there were more than 100 cases generated in at least one department (54% in the 491 reference dataset). The geographical distribution of the risks of large outbreaks was almost identical to 492 the reference simulation set ( Figure 5) . Therefore, although the number of importations was reduced, 493 changing the level of recent incidence did not have a clear impact on the risks of active transmission. 494

More departments became vulnerable to secondary transmission, and despite importations in these 495 departments being rarer, they were more likely to lead to large outbreaks when they happened. The 496 two opposing effects recent incidence had on importation and transmission therefore created a 497 different dynamic of transmission observed in the simulation set, without strongly reducing the risks of 498 outbreaks. 499

Each of these simulation sets highlighted the wide range of scenarios that could be generated using the 500 parameter distributions inferred by our model. In order to gain more understanding on the spatial 501 spread and consequences of importations, we then explored the impact of localised repeated 502 importations on overall transmission. 503 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 1, 2021. ; https://doi.org/10.1101/2021.05.31.21257977 doi: medRxiv preprint February, which corresponds to the peak period of the other components, and would therefore be 514 more likely to cause secondary transmissions (Supplement Section 3). We simulated one year of 515 transmission following ten importations in December 2018 to illustrate: i) the potential for local 516 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 1, 2021. ; https://doi.org/10.1101/2021.05.31.21257977 doi: medRxiv preprint outbreaks, and ii) the spatial spread of transmission following repeated local importations. We selected 517 four departments to compare the impact of repeated importations in a range of settings: Paris (many 518 inhabitants, 91% vaccine coverage, surrounded by urban areas), Bouches-du-Rhône (many inhabitants, 519 84% vaccine coverage), Haute Garonne (many inhabitants, 91% vaccine coverage but high levels of 520 recent incidence, surrounded by rural areas with lower vaccine coverage), and Gers (Rural area, 79% 521 vaccine coverage) (Figure 6 ). 522

Firstly, major local outbreaks in the department of importation were generated in all four simulation 523 sets, and especially in Paris and Bouches-du-Rhône, where the proportion of simulations that yielded 524 more than 100 subsequent cases in the department was 40% and 39%, respectively. In the Bouches-525 du-Rhône, large outbreaks were mostly due to the low vaccination coverage, whereas in Paris, 526 outbreaks were mostly linked to the connectivity to nearby areas and the high number of inhabitants, 527 which meant the department was likely to attract cross-regional transmissions. Major local outbreaks 528 were rarer in the other two scenarios (9% of simulations above 100 in Haute Garonne, 10% in Gers). 529

The lower proportion of large outbreaks resulted from different factors: recent large outbreaks in Haute 530

Garonne reduced the autoregressive predictor, lowering the number of secondary cases per case 531 imported; whereas since Gers is a rural department, with a low number of inhabitants, almost all the 532 local cases were due to local transmission (auto-regressive component), with very few cross-regional 533 transmissions into Gers. 534

Conversely, the simulations where cases were imported in Gers yielded the largest spatial spread 535 throughout the country: the median number of departments that reported at least 1 case was 53 (16 536 when the importations were generated in Haute Garonne; 15 in Bouches-du-Rhône; 39 in Paris). As 537 stated in the method, the number of cross-regional transmissions is the product of the predictor and 538 the connectivity matrix, divided by the number of inhabitants in the department of origin, to represent 539 that only a fraction of commuters will be infected. Therefore, populous areas are more likely to attract 540 cross-regional transmissions, whereas more rural departments are more likely to seed outbreaks in 541 other areas. The relatively high spatial spread when cases were imported in Paris is due to the short 542 distance between Paris and its suburbs, which is then more likely to cause cross-regional transmission 543 in the northern departments. Despite the cross-regional spread observed in both of these simulations 544 sets, outbreaks remained local, and occurrences of nation-wide outbreaks were almost null. The 545 departments most at risk of outbreak following cross-regional spread were some of the direct 546 neighbours of the department of importations, or the large urban areas (Figure 6 ). To further explore 547 this, we ran the same simulations decreasing the vaccine coverage by three percent, which greatly 548 increased the number of departments exposed in each simulation set, and increased the risk of local 549 transmission (Supplement Section 6). Therefore, although repeated importations could cause active 550 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. outbreaks despite high national vaccine uptake [1, 2, 4, 25] . Our study showed that local values of vaccine 562 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted June 1, 2021. coverage were linked to lower transmission, whereas lower levels of recent incidence were not 563 associated with lower risks of local transmission. Furthermore, we highlighted that a drop of 3% in the 564 three-year vaccine coverage triggered a five-fold increase in the number of cases simulated in a year. 565

The fact that higher vaccine coverage was associated with a lower number of secondary cases is 566 consistent with prior expectations, and would confirm that the local values of first dose vaccine 567 coverage are a good indicator of the actual immunity in the population and risks of future transmission. 568

Reporting accurate values of local vaccine coverage is challenging, for instance because the vaccination 569 status of people moving regions can be hard to track and lead to measurement errors. Furthermore, 570

we did not have access to complete data on the coverage of the second MMR dose, which would be a 571 better indicator of vulnerable areas. Therefore, detecting the association between recent vaccine 572 uptake and incidence is encouraging. The impact of local vaccination coverage on transmission may 573 also be muddled by sub-regional vaccine heterogeneity. For instance, pockets of susceptibles within a 574 region, i.e. areas within the region where the vaccine coverage is substantially lower than the regional 575 average, may be at high risk of transmission and would not be observable in regional coverage [39] . 576

This phenomenon can only be hypothesised here, and could be explored using local data on incidence 577

and vaccine uptake at a sub-regional scale. 578

Variations in vaccine coverage had a noticeable impact on the number of cases generated in the 579 simulation study. We showed the effects of a three percent increase and decrease of the three-year 580 average vaccine coverage on the number of cases, which highlighted the risks of uncontrolled 581 transmission in the event of a decrease of vaccine-induced protection. Events such as the disruption 582 caused by the SARS-COV-19 pandemic on routine measles vaccination campaigns could therefore highly 583 increase the risks of uncontrolled measles transmission in the years to come [40, 41] . 584

The departments that reported few cases per million in the past three years were associated with 585 higher risks of local transmission (autoregressive component). Therefore, according to our model, 586 regions eligible for elimination status were not associated with lower risks of onwards transmission. 587

Conversely, high levels of recent transmission were associated with a lower number of cross-regional 588 transmissions and importations, although we cannot methodologically establish the causality of this 589 association. The impact on the simulation study was clear: when we set the category of recent incidence 590 to the lowest level, departments were less exposed to cases, and spatial spread was rarer, whilst there 591 was little change in the risks of major outbreaks. The simulations showed an 'all-or-nothing' situation: 592 departments tended to report very few to no cases, whilst also being more likely to be affected by 593 outbreaks. These results would indicate that looking into the level of incidence to quantify the future 594 risks of outbreaks can be deceptive, and importations in a department with low recent incidence would 595 result in large transmission clusters. 596

We proposed a new framing of the Epidemic-Endemic model implemented in hhh4 by adapting it to 597 daily count data using the distribution of the serial interval to compute the local transmission potential. 598

Using daily case counts allowed us to avoid biases associated with aggregated case counts, such as the 599 influence of the arbitrary aggregation date, by accounting for the impact of variation in the serial 600 intervals. We also accounted for the risks of unreported cases by computing a composite multimodal 601 serial interval, thus allowing for transmission with a missing generation, or an unreported ancestor. The 602 model was able to capture the dynamic of transmission better than the 10-day aggregated model, as 603 shown by the calibration study (Supplement Section 7). Nevertheless, our framing of the hhh4 model 604 introduced new biases: we used a distribution of the serial interval based on previous studies rather 605 than estimating the weights during the fitting procedure and set the proportion of missing generations 606 in the composite serial interval. We explored the impact of the proportion of missing generations by 607 fitting the model with different composite serial intervals and concluded that the impact of each 608 covariate was robust to these changes (Supplement Section 1). We also integrated a potential day-of-609 the-week effect, and observed that although it had an impact on the auto-regressive component, it did 610 not change the estimates of the other parameters, and therefore did not change the conclusions of the 611 study (Supplement Section 8). 612 Using the hhh4 model allowed us to analyse the different impact of various covariates on local and 613 cross-regional transmission, and background importation of cases. According to the models we 614 implemented, an overwhelming majority (>90%) of the transmission came from the cross-regional and 615 local components of the regression. This indicates that in the models, the endemic component only 616 corresponds to rare background cases that could not be linked to concurrent transmission events. This 617 could point towards model misspecifications, for example, connecting unrelated importations to 618 concurrent local transmission. Since endemic transmission tends to refer to cases otherwise 619 unexplained by the mechanistic components, the seasonality of the endemic component is decoupled 620 from the other components, i.e. endemic cases are likely when local and cross-regional transmission 621 are lower. 622

Since the endemic component accounted for such a small minority of the cases, group importations of 623 cases in a given department were rarely observed in the simulations. However, tourism, and local 624 events lead to large gatherings and can increase the risks of group importations in a limited period of 625 time [36, 37] . We simulated the spatial spread following repeated importations in a given department, 626 and highlighted that although large outbreaks in the department of importations were common, 627 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted June 1, 2021. ; https://doi.org/10.1101/2021.05.31.21257977 doi: medRxiv preprint nation-wide transmission following these importations was very rare. Only the departments where all 628 cases had been imported, and its neighbours, were at risk of uncontrolled outbreaks. Decreasing the 629 level of vaccination by three percent was associated with a large increase in the level of exposure of all 630 departments, and in the number of departments where large outbreaks were generated (Supplement 631 Section 6 and 7). The high levels of transmission observed in recent years in France suggest that 632 importations are frequent, and even a small drop in vaccination could dramatically increase measles 633 transmission in the country. France, and in Savoie (East). This could be due to importations and cross-regional transmission that are 640 under-estimated by our model. Although the model captured the dynamics seen in the data, the 641 calibration study showed it was only able to predict short-term transmission up to one week. The PIT 642 histogram associated to the 14-day calibration displayed signs of bias, which shows that the model was 643 not able to consistently predict variations in the future number of cases in the next two weeks. We 644 identify several factors that could explain the discrepancies observed for longer term predictions: i) the 645 indicator of local immunity we used was flawed: two-dose coverage would be a better indicator of the 646 proportion of the population that is protected; ii) The sub-regional heterogeneity in coverage and past 647 incidence within the department that could be concealed by NUTS3 aggregated data: because of social 648 groups that rarely mix with one another, or large NUTS regions, large outbreaks in a given community 649 would not be a good indicator of the overall level of immunity in a region. Nevertheless, we believe that

A Measles Outbreak in an 705

Underimmunized Amish Community in Ohio

Large measles epidemic in the Netherlands

World Health Organization. Framework for verifying elimination of measles and rubella

European Region loses ground in effort to eliminate measles

Measles outbreak in the Americas

From the measles-free status to the current outbreak in Brasil

Factors associated with measles resurgence in the 733

United States in the post-elimination era

Measles 736 outbreak linked to insufficient vaccination coverage in Nouvelle-Aquitaine region

A statistical framework for the analysis of multivariate infectious 740

Endemic-epidemic models with discrete-time serial interval distributions for 751 infectious disease prediction

The tip of the iceberg : incompleteness 753 of measles reporting during a large outbreak in The Netherlands in 2013 -2014

Systematic comparison of trip distribution laws and 756 models

Estimation de la population au 1 er 758 janvier

Package 764 "raster

Heterogeneity in vaccination coverage explains the size and 766 occurrence of measles epidemics in German surveillance data

Données départementales 2007-2012 de couverture vaccinale rougeole, 769 efforts and 2008-2011 outbreak

Données de déclaration obligatoire de la rougeole

Use of a geographic 785 information system to map cases of measles in real-time during an outbreak in

Geospatial characteristics of measles 788 transmission in China during

Spatiotemporal 791 distribution and determinants of measles incidence during a large outbreak

Assessing the performance of 795 real-time epidemic forecasts: A case study of Ebola in the Western Area Region of Sierra Leone

Predictive Model Assessment for Count Data

independent concurrent chains. Thirdly, since this is not a transmission model, some extreme values 662 could trigger unlikely behaviour. For instance, if the vaccination rate would be 100%, we would still 663 expect sporadic transmission. Although this would not be entirely implausible given that only the 664 vaccination coverage in the past three years was taken into account in the models (i.e. even if it was 665 100% coverage, there could be susceptible individuals in different age groups). Finally, the impact of 666 the different covariates on the number of cases was constant through time. For instance, the impact of 667 seasonality may depend on factors such as the weather which may vary each year, which would not be 668accounted for in the model we developed. 669We used variables collected in a wide range of settings (regional vaccine coverage, incidence, number 670 of inhabitants, surface), therefore this analysis can be reproduced in other countries to analyse the 671 potential for local transmission as well as the impact of recent incidence and vaccine-induced immunity. 672Since the case counts data are not publicly available, we share the code used to generate the analysis 673 applied to a simulated dataset on a Github repository: (https://github.com/alxsrobert/measles-674 regional-transmission). 675Data availability 676 The daily case counts data came from the European Surveillance System -TESSy, provided by Santé 677 Publique France and released by ECDC. The data cannot be shared publicly. To make this study as 678 reproducible as possible, we generated simulated case counts in France over the same timespan as the 679 main analysis. The code used to generate the simulated dataset, and all the figures presented in the 680 paper is shared in a Github repository (https://github.com/alxsrobert/measles-regional-transmission). 681