The role of “spillover” in antibiotic resistance

Antibiotic use is a key driver of antibiotic resistance. Understanding the quantitative association between antibiotic use and resulting resistance is important for predicting future rates of antibiotic resistance and for designing antibiotic stewardship policy. However, the use-resistance association is complicated by “spillover”, in which one population’s level of antibiotic use affects another population’s level of resistance via the transmission of bacteria between those populations. Spillover is known to have effects at the level of families and hospitals, but it is unclear if spillover is relevant at larger scales. We used mathematical modeling and analysis of observational data to address this question. First, we used dynamical models of antibiotic resistance to predict the effects of spillover. Whereas populations completely isolated from one another do not experience any spillover, we found that if even 1% of interactions are between populations, then spillover may have large consequences: the effect of a change in antibiotic use in one population on antibiotic resistance in that population could be reduced by as much as 50%. Then, we quantified spillover in observational antibiotic use and resistance data from US states and European countries for 3 pathogen-antibiotic combinations, finding that increased interactions between populations were associated with smaller differences in antibiotic resistance between those populations. Thus, spillover may have an important impact at the level of states and countries, which has ramifications for predicting the future of antibiotic resistance, designing antibiotic resistance stewardship policy, and interpreting stewardship interventions.


INTRODUCTION 23
Antibiotic resistance is a major threat to public health (1). Outpatient antibiotic use, 24 which accounts for approximately 80% of human antibiotic use (2,3), is considered a 25 principal driver of antibiotic resistance in the community (4). Understanding the 26 relationship between use and resistance is important because it allows accurate 27 predictions of the future of antibiotic resistance and goal-oriented antibiotic stewardship 28 policy. The use-resistance association has been previously characterized in many 29 ecological studies at the level of US states (5-7) and European countries (8,9). 30 However, antibiotic resistance is a complex, temporally dynamic phenomenon (10-13), 31 and many factors complicate the use-resistance association, making what should be an 32 "obvious" connection sometimes difficult to identify and quantify (11). Even when 33 detected, observed use-resistance associations are sometimes weaker than might be 34 expected (14). One factor that could account for the difficulty in detecting use-resistance 35 associations in ecological studies and for the apparent weakness of such associations 36 is "spillover". 37 38 "Spillover" is a consequence of the fact that antibiotic-resistant and -susceptible bacteria 39 can be transmitted from person to person. Thus, one person's risk of an antibiotic 40 resistant infection depends on their own antibiotic use (15,16) as well as the rates of 41 antibiotic use among their contacts (17). For example, one person's use of antibiotics 42 increases the risk of an antibiotic resistant infection among their family members (18-43 21). As another example, hospitalized patients with no recent antibiotic use can have a 44 higher risk of resistance than people in the community with high antibiotic use (22)  45 because antibiotic use and resistance in other hospitalized patients are high. 46 47 Spillover is important for three reasons. First, it means antibiotic resistance is not merely 48 a localized problem. It is well-understood that new resistance determinants can emerge 49 in one geography and spread globally (23,24), but the role of spillover in determining the 50 levels of resistance in a given locale is not well-quantified. To what degree, for example, 51 can one US state expect that its antibiotic resistance levels are due to antibiotic use 52 within its borders, rather than in surrounding states? Second, spillover makes it difficult 53 to design antibiotic stewardship interventions and understand their results. For example, 54 if antibiotic use in one hospital changes, resistance might not change as expected 55 because of spillover, from the community or other hospitals into that hospital's patients. 56 Finally, spillover makes it difficult to interpret the results of controlled antibiotic 57 interventions, such as the effect of mass drug administration on antibiotic resistance 58 (25,26) The effects of spillover also depend on population sizes. For example, two large 68 populations will have most interactions within themselves, rather than between each 69 other. Spillover should therefore be most pronounced when considering small 70 populations and become less important for large populations. As mentioned above, a 71 single individual's risk of resistance is modulated by antibiotic use in their family or in 72 their healthcare facility. Spillover is also observed at the level of hospitals, as the level 73 of resistance in one hospital appears to be affected by resistance levels in nearby 74 hospitals as well as by antibiotic use rates in the surrounding communities (27)(28)(29). 75 Presumably, when examining ever larger populations, such as US Census tracts (30), 76 US states, or European countries, the effect of spillover will become less important. 77 However, the relationship between population size and spillover effects is not well 78 understood. 79

80
We hypothesized that US states and European countries, which are large populations 81 with relatively independent public health policies, may be subject to substantially lower 82 levels of antibiotic resistance spillover than family-or hospital-sized populations. This 83 hypothesis, if true, would mean that individual states or countries could act as 84 independent "laboratories" of antibiotic use and resistance. If not, it means that 85 outpatient antibiotic resistance policy must be national or international in order to 86 achieve its full effect. To evaluate this hypothesis, we first use mathematical models of 87 antibiotic use and resistance to make quantitative predictions about the effect of 88 spillover between populations as a function of their amount of mutual interaction. Then, 89 we search for signals of spillover in observational data of antibiotic use and resistance 90 in US states and European countries. To examine how interactions between populations could theoretically affect the 96 association between antibiotic use and resistance, we used the within-host neutrality 97 (WHN) mathematical model presented by Davies et al. (31) and described in the 98 Supplemental Methods. Briefly, the model predicts the prevalence ρ of antibiotic 99 resistance that results from an antibiotic use rate τ in a single, well-mixed population. To 100 verify that conclusions drawn from the WHN model are not specific to the model 101 structure, we also repeated all analyses with the "D-types" model of use and resistance 102 (32). We selected these two models because they demonstrate coexistence between 103 sensitive and resistant strains at equilibrium over a wide parameter space. Parameter 104 values and simulation methodology for both models are in the Supplemental Methods. 105 In the simulations, antibiotic use is measured as monthly treatments per capita and 106 resistance as the proportion of colonized hosts carrying resistant strains. 107 108 To conceptually frame and clarify the question of spillover, we simulated an antibiotic 109 stewardship intervention experiment using a structured host population approach 110 inspired by Blanquart et al. (33). We considered pairs of an intervention population with 111 antibiotic use rate τint and a control population with use rate τcont. To determine how 112 spillover affects the intervention's measured outcome, we modulated the proportion ε of 113 each population's contacts that are in the other population. For ε = 0%, the populations 114 are completely separate. For ε = 50%, contacts across populations are just as likely as 115 contacts within populations (Supplemental Methods). We varied ε between 0% and 116 50%, and we varied the difference in use Δτ = τcont -τint between 0 and 0.15 treatments 117 per person per month while keeping the average use 0.5 × (τcont + τint) fixed at 0.125, 118 reflecting the range of antibiotic use rates in the original model presentations. 119 120 Observational data 121 We examined antibiotic use and resistance for 3 pathogen-antibiotic combinations: S. 122 pneumoniae and macrolides, S. pneumoniae and β-lactams, and Escherichia coli and 123 quinolones. We considered these 3 combinations because they are the subject of many 124 modeling (31,32) and empirical studies (5,15). 125 126 Observational data were drawn from 3 sources. First, we used MarketScan (34) and 127 ResistanceOpen (35) as previously described (7) To test the theoretical prediction that the same difference in antibiotic use will be 151 associated with smaller differences in antibiotic resistance when two populations (US 152 states or European countries) have stronger interactions, we tested whether the use-153 resistance association is weaker in adjacent pairs of populations, which presumably 154 have more cross-population contacts, compared to non-adjacent populations. Two 155 populations were considered adjacent if they share a land or river border (40-42). 156 adjacent pairs and non-adjacent pairs of populations using the median value. Because 161 use-resistance associations between pairs of populations are correlated, we used the 162 jackknife method to compute confidence intervals on the difference in medians between 163 groups. We used the Mann-Whitney U test to compute statistical significance. Because 164 our theoretical results suggested the ratio of percentage points difference in resistance 165 divided by difference in antibiotic use rates was a predictable function of the degree of 166 population mixing, we considered only this functional form for the use-resistance 167 association. 168 169

Use-resistance associations by interactions 170
Because adjacency might be too coarse measure of populations' interactions to detect 171 spillover, we performed a similar analysis as above, but predicting the use-resistance 172 association between populations using transportation data. For US states, we used 173 inter-county commuting statistics from the US Census (43). For European countries, we 174 used inter-country passenger flight data from Eurostat (44). Rather than trying to infer a 175 precise mathematical relationship between transportation statistics and epidemiological 176 contacts, we used a nonparametric approach: we assumed that pairs of populations 177 with relatively little inter-population transportation also have relatively few inter-178 population contacts, but we infer only the rank ordering of inter-population contacts, not 179 their magnitudes. Specifically, we first converted the matrix of the number of counts 180 (workers in the commuting data and passengers in the flight data) from each population 181 to every other population into the proportion of counts moving from one population to 182 another (i.e., divided each row by its sum), then symmetrized the resulting matrix (taking 183 the elementwise average of the matrix and its transpose), and finally converted the 184 resulting values into ranks. We assumed that intra-population interactions outnumber 185 inter-population interactions and so set diagonal entries, which represent within-186 population interactions, to the highest rank. We measured the association between 187 ranked interactions and use-resistance associations using the nonparametric 188 Spearman's correlation, computed confidence intervals using the jackknife method, and 189  Figure 1). With increasing interaction strength, the same intervention, that is, the same 202 difference in antibiotic use between the populations, was associated with a smaller 203 difference in antibiotic resistance. The difference in resistance between populations 204 increases with the difference in antibiotic use (Figure 1d), but the use-resistance 205 association, measured as the ratio of the difference in resistance to the difference in 206 use, depends strongly on the interaction strength ( Figure 1e). Thus, spillover between 207 populations attenuates the measured use-resistance association. 208

209
The precise relationship between ε, the proportion of each population's contacts that are 210 in the other population, and the attenuation of the use-resistance association depended 211 on the choice of mathematical model (Supplemental Table 1, Supplemental Figure 1). 212 For ε = 1%, the use-resistance declined by approximately 30% in the WHN model and 213 more than 60% in the "D-types" model. In other words, the models predict that as few as 214 1% of contacts need to be across populations, rather than within populations, to cause 215 the observed effect of an antibiotic stewardship intervention to shrink by one-third, or 216 even half. 217

218
To test whether spillover is important at the scale of US states or European countries, 219 we measured use-resistance associations between pairs of populations in 6 220 combinations of pathogen species, antibiotic class, and data source ( Figure 2). We 221 reasoned that if spillover is relevant at these scales, then pairs of states or countries 222 with stronger interactions would have detectably weaker use-resistance associations. 223

224
We first tested whether pairs of physically adjacent populations (e.g., Massachusetts 225 and Connecticut) had weaker use-resistance associations than non-adjacent 226 populations (e.g., Massachusetts and Alaska). In 5 of 6 pathogen/antibiotic/dataset 227 combinations, the median use-resistance association was smaller among adjacent 228 populations than among non-adjacent populations ( Figure 3), but in no case was the 229 difference statistically significant after multiple hypothesis correction (Supplemental 230 Table 2). Point estimates of the relative difference in median use-resistance 231 associations between adjacent populations were 18% to 50% weaker than between 232 non-adjacent populations (excepting S. pneumoniae with β-lactams, which was an 233 outlier), consistent with the theoretical modeling results showing a 50% reduction in the 234 use-resistance association for populations with approximately 1% of interactions across 235 populations (Supplemental Table 1). 236 237 Next, to account for the possibility that adjacency was too coarse a measure for 238 interactions between populations, we instead used a rank-ordered estimate of 239 interactions using US commuting and European airline passenger flows (Supplemental 240 Figure 2). In 4 of 6 dataset/pathogen/antibiotic combinations, the nonparametric 241 association between increased inter-population interactions and decreased use-242 resistance associations was statistically significant, thus confirming the general trend 243 observed in the adjacency analysis (Figure 4, Supplemental Table 3). The weakest 244 significant result was for S. pneumoniae and macrolides in the 245 MarketScan/ResistanceOpen dataset (Spearman's ρ 0.07, 95% jackknife confidence 246 interval -0.03 to 0.18; p = 0.028, Mantel test), and the strongest was for E. coli and 247 quinolones in the Xponent/NHSN dataset (ρ = 0.13, 95% jackknife confidence interval 248 0.007 to 0.25; p = 0.001). The correlation for S. pneumoniae and macrolides in the 249 ECDC data has a point estimate suggesting spillover but was not statistically significant, 250 while the correlation for S. pneumoniae and β-lactams in the ECDC data had the 251 opposite point estimate, consistent with the adjacency analysis ( Figure 3). 252 253 Finally, to quantify the effect of increased interactions on the observed use-resistance 254 associations, we compared the use-resistance associations in pairs of populations 255 within the lowest decile of interactions against those in the highest decile, using the 256 same approach as for the adjacency analysis above (Supplemental Table 4). In 5 of 6 257 dataset/pathogen/antibiotic combinations, the point estimate for the different in use-258 resistance associations was consistent with spillover, with a weaker association among 259 pairs of populations with greater interactions. In those 5 cases, the point estimates 260 ranged from a 18% reduction up to a 75% reduction in use-resistance associations 261 among the highest-interacting pairs of populations, compared to the lowest-interacting 262 populations. 263 264 265

DISCUSSION 266
We used theoretical models to show that interactions between two populations can 267 attenuate the observed use-resistance association. In simulations, the quantitative 268 relationship between inter-population interactions and the attenuation of the use-269 resistance association was dependent on the theoretical model used. However, we 270 found that, in two models of the use-resistance association, having on the order of 1% 271 of interactions between a control and intervention population was sufficient to attenuate 272 the observed effect of theoretical stewardship intervention by 50%, relative to a situation 273 where the two populations were completely isolated. These theoretical results suggest 274 that even small numbers of interactions could lead to substantial spillover. 275 276 When examining observational antibiotic use and resistance data from US states and 277 European countries, we did not detect a robust signal of spillover among pairs of 278 adjacent populations, as opposed to non-adjacent pairs, even across 3 pathogen-279 antibiotic combinations in 3 separate datasets. However, when using more fine-grained 280 transportation data to estimate the relative ranking of epidemiological contacts between 281 those populations, we found a correlation between increased interactions and 282 attenuated use-resistance associations. Pairs of populations in the highest decile of 283 inter-population interactions, that is, those most subject to spillover, had use-resistance 284 associations on the order of 50% weaker than pairs in the lowest decile of interactions. 285 The 2 pathogen/antibiotic dataset combinations with data not indicative of spillover, 286 namely S. pneumoniae and β-lactams and macrolides in the ECDC data, may have not 287 shown the same signal as other cases because of the smaller number of populations in 288 those cases (27, versus 28 to 50 in the other cases) led to insufficient statistical power 289 or potentially because the biology or epidemiology of S. pneumoniae resistance in these 290 cases is somehow different and does not exhibit spillover. 291 292 These theoretical and empirical results suggest that spillover is relevant at the level of 293 US states and European countries. This finding has important ramifications. First, 294 attempts to attribute changes in a population's level of antibiotic resistance to changes 295 in that population's rates of antibiotic use may lead to inaccurate conclusions unless use 296 and resistance in surrounding populations is accounted for. Second, state-or country-297 level antibiotic stewardship pilot studies may substantially underestimate the potential 298 reduction in antibiotic resistance that would follow from a reduction in antibiotic use if 299 that reduction were implemented at a larger scale. Third, mass drug administration trials 300 may lead to elevated levels of antibiotic resistance in the control populations if those 301 populations are not entirely separated from the intervention population. Finally, spillover 302 can at least partly explain why use-resistance associations at the level of US states or 303 European countries are sometimes difficult to detect and, when they are detected, are 304 sometimes weaker than expected (5,11,14). Furthermore, spillover means that 305 theoretical models of antibiotic use and resistance that treat US states or European 306 countries as epidemiologically independent populations will not accurately represent the 307 dynamics of resistance (33). 308 309 Our study has several limitations. First, we interpreted the theoretical results and 310 ecological data as if the association between antibiotic use and resistance were causal 311 and deterministic. However, decreases in the use of an antibiotic may not necessarily 312 lead to declines in resistance to that antibiotic in a target pathogen (12,47-49). We do 313 not address co-resistance and cross-selection (50,51), and we assumed that resistance 314 equilibrates on a timescale comparable to an intervention. Previous research has shown 315 that resistance among E. coli, S. pneumoniae, N. gonorrhoeae and other organisms can 316 respond to changes in antibiotic use on the timescale of months (52-55), but the 317 expected delay between a perturbation to antibiotic use and the resulting change in 318 resistance remains a subject of active study (13,52,56,57). Nevertheless, the use of 319 ecological data was essential to addressing our hypothesis, as data from multiple 320 controlled, state-or country-wide experiments are not available. 321 322 Second, our analyses attributed all differences in antibiotic resistance between 323 populations to differences in use across those populations and to interactions between 324 them. In fact, antibiotic resistance is associated with factors beyond antibiotic use 325 (6,58), and those factor are likely spatially correlated. In other words, closely interacting 326 populations might have more similar use-resistance associations because they tend to 327 be more similar with respect to other determinants of antibiotic use. Our estimates of the 328 correlation between inter-population interactions and the attenuation of use-resistance 329 relationships may therefore be overestimates. A more careful quantification of the 330 relative roles of spillover versus other spatially-correlated determinants of resistance is 331 required. 332 333 Third, our analysis only considered pairs of populations, when in fact spillover is 334 happening between all pairs of populations in our analysis simultaneously. We used the 335 pairs approach because it allowed for a simple theoretical model and a straightforward 336 comparison of theory with the observational data. However, more sophisticated 337 approaches that account for the network of spillover interactions will likely lead to more 338 refined characterizations of spillover. 339 340 Finally, analyses based on administrative entities like US states or European countries, 341 although logistically attractive "laboratories" of antibiotic stewardship, will always be 342 difficult to interpret because administrative entities average over important dimensions 343 of population structure like age (59), sexual networks (60), and race/ethnicity (61). Thus, 344 use-resistance associations measured across states and countries may be different 345 from those that appear among geographically-proximate populations with dissimilar 346 antibiotic use rates, such as the sexes (62)   The views and opinions of the authors expressed herein do not necessarily state or 568 reflect those of the ECDC. The accuracy of the authors' statistical analysis and the 569 findings they report are not the responsibility of ECDC. ECDC is not responsible for 570 conclusions or opinions drawn from the data provided. ECDC is not responsible for the 571 correctness of the data and for data management, data merging and data collation after 572 provision of the data. ECDC shall not be held liable for improper or incorrect use of the 573 data.