The potential for "spillover" in outpatient antibiotic stewardship interventions among US states

Abstract Antibiotic stewardship combats antibiotic resistance by reducing inappropriate antibiotic use. Stewardship policy should be guided by experimental stewardship interventions. However, the design and interpretation of stewardship interventions is subject to “spillover”, in which the transmission of microbes between the control and intervention population reduces the intervention’s measured effect. Small-scale stewardship experiments may therefore underestimate the effect of a larger-scale implementation. Here, we aimed to quantify the effect of spillover on state-level outpatient antibiotic stewardship interventions to determine if states are feasible “laboratories” for designing national policy. First, we used dynamical models of antibiotic resistance to predict the effects of spillover, finding that if even 1% of residents’ interactions are between, rather than within, US states, the measured effect of a state-wide stewardship intervention 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. We found that these cross-sectional data were insufficiently powered to detect even the large spillover effect sizes predicted by the mathematical models. We were unable to rule out the possibility that state-level changes in antibiotic use, either increases or reductions, may lead to substantially smaller changes in antibiotic resistance than if those changes took place nationwide. We suggest that well-designed, controlled interventions, couple with more sophisticated modeling and analysis, with could help determine spillover’s policy ramifications.


Introduction 46
Antibiotic resistance is a major threat to public health (1). Outpatient antibiotic use, which accounts for approximately 80% of human antibiotic use (2,3), is considered a 48 principal driver of antibiotic resistance in the community (4), and antibiotic stewardship 49 aims to mitigate antibiotic resistance (5-7) by reducing antibiotic use. US national 50 stewardship policy should be guided by evidence from experimental stewardship 51 interventions at smaller scales. For example, the results of interventions at the scale of 52 US states could be used to inform the design of national policy. However, antibiotic 53 resistance is a complex, temporally dynamic phenomenon (8-11), making the design 54 and interpretation of stewardship interventions challenging. 55

56
A key feature of antibiotic resistance is that it can be transmitted from person to person, 57 so that one person's risk of an antibiotic resistant infection depends on that person's 58 antibiotic use (12,13) as well as the rates of antibiotic use among that person's contacts 59 (14). The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint For example, if antibiotic use in one hospital changes, resistance might not change as 69 expected because resistant or susceptible bacteria can be transmitted, or "spill over", to 70 that hospital's patients in the community or in other hospitals. 71

72
The effect of susceptibility and resistance "spilling over" between populations during 73 stewardship interventions could theoretically be reduced by using larger populations. 74 Smaller populations tend to have more transmission with the surrounding populations 75 compared with larger populations, which tend to have more contacts within populations, 76 rather than between populations. Thus, the problem of "spillover" is mitigated when 77 studying larger populations. However, even hospitals are subject to spillover, as the 78 level of resistance in one hospital appears to be affected by resistance levels in nearby 79 hospitals as well as by antibiotic use rates in the surrounding communities (20-22). It is 80 therefore possible that even hospitals may be too small and too subject to spillover to 81 be accurate "laboratories" for stewardship. 82 83 We hypothesized that stewardship interventions at the level of US states, which are 84 large populations with relatively independent public health policies, may be subject to 85 substantially lower levels of spillover than individual-level or even hospital-level 86 interventions. We evaluated this hypothesis using mathematical models and cross-87 sectional data of antibiotic use and resistance. First, we use mathematical models of 88 antibiotic use and resistance to make quantitative predictions about the effect of 89 spillover between US states and European countries. Second, we search for signals of 90 spillover in observational data of antibiotic use and resistance in US states and 91 We adapted the WHN model, using a structured host population approach inspired by 115 Blanquart et al. (30), to simulate a stewardship experiment in which an intervention 116 population has a lower antibiotic use rate "#$ than a control population with use rate 117 %&#$ . To determine how spillover affects the intervention's measured outcome, we 118 In this study, we examined antibiotic use and resistance for 3 pathogen-antibiotic 127 combinations: S. pneumoniae and macrolides, S. pneumoniae and β-lactams, and 128 Escherichia coli and quinolones. We considered these 3 combinations because they are 129 the subject of many modeling (28,29) and empirical studies (12,23). 130 131 Observational data were drawn from 3 sources. First, we used MarketScan (31) and 132 ResistanceOpen (32) as previously described (25  To test the theoretical prediction that the same difference in antibiotic use will be 156 associated with smaller differences in antibiotic resistance when two populations (US 157 states or European countries) have stronger interactions, we tested whether the use-158 resistance association is weaker in adjacent pairs of populations, which presumably 159 have more cross-population contacts, compared to non-adjacent populations. Two 160 . CC-BY 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It In a sensitivity analysis, to account for the possibility that the use-resistance association 171 is not well-described using the simple difference in resistance proportions, we use the 172 log odds ratio of resistance as the numerator in the use-resistance association. 173 174

Use-resistance relationships by adjacency, accounting for confounders 175
We expected that analyzing use-resistance associations by adjacency might artificially 176 inflate the signal for spillover because determinants of antibiotic resistance aside from 177 antibiotic use are spatially correlated. For example, if temperature affects levels of 178 resistance (24), then the fact that adjacent populations tend to have similar climates 179 may cause those populations to have more similar resistances, mimicking spillover. To 180 partially account for these other determinants of resistance, we performed robust linear 181 regressions predicting the use-resistance relationship from adjacency (dichotomous 182 variable) as well as the differences in population density (38), per capita income (39), 183 . CC-BY 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint and mean temperature (24) between the two populations (Supplemental Methods). 184 Regressions were computed using the rlm function in the MASS package (40) in R. 185 Confidence intervals on the adjacency-use interaction coefficient were computed using 186 the jackknife method described above. 187 With increasing interaction strength, the same intervention, that is, the same difference 204 in antibiotic use between the populations, was associated with a smaller difference in 205 antibiotic resistance. The difference in resistance between populations increases with 206 . CC-BY 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint the difference in antibiotic use (Figure 1d), but the use-resistance association, 207 measured as the ratio of the difference in resistance to the difference in use, depends 208 strongly on the interaction strength ( Figure 1e). Thus, spillover between populations 209 attenuates the measured use-resistance association. 210

211
The use-resistance association was sensitive to ε, the proportion of each population's 212 contacts that are in the other population, but depended on choice of the mathematical 213 model of use-resistance association (Supplemental Table 1, Supplemental Figure 1). 214 For values as small as = 10 89 , a typical level of interaction between two US states or 215 European countries (Supplemental Figure 2), the use-resistance association declined 216 by less than 1% with the WHN model but up to 20% for the "D-types" model. For = 217 1%, the use-resistance declined by approximately 30% in the WHN model and more 218 than 60% in the "D-types" model. In other words, the models predict that as few as 1% 219 of contacts need to be across populations, rather than within populations, to cause the 220 observed effect of an antibiotic stewardship intervention to shrink by one-third, or even The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint

230
We first tested whether pairs of physically adjacent populations (e.g., Massachusetts 231 and Connecticut) had weaker use-resistance associations than non-adjacent 232 populations (e.g., Massachusetts and Alaska). In 5 of 6 pathogen/antibiotic/dataset 233 combinations, the median use-resistance association was smaller among adjacent 234 populations than among non-adjacent populations (Figure 3). In 2 cases, the confidence 235 interval on the ratio of use-resistance associations among adjacent populations, 236 compared to non-adjacent populations, did not include zero (Supplemental Table 2). 237 First, for S. pneumoniae resistance to macrolides in the MarketScan/ResistanceOpen 238 dataset, use-resistance associations were 27% weaker (95% CI 6% to 49%) among 239 adjacent states compared to non-adjacent states. Second, for E. coli resistance to 240 quinolones in the Xponent/NHSN dataset, use-resistance associations were 50% 241 weaker (95% CI 27% to 73%) among adjacent states compared to non-adjacent states. 242 Results were similar when using a different metric of the use-resistance association 243 (Supplemental Table 3). 244 245 We next checked that determinants of antibiotic resistance aside from antibiotic use 246 were not artificially amplifying spillover, making differences in the use-resistance 247 association between adjacent and non-adjacent pairs larger. We performed robust 248 regressions, predicting the use-resistance association from adjacency while controlling 249 for the differences in other covariates that are established determinants of resistance 250 levels. Results were almost identical when including these covariates (Supplemental 251 . CC-BY 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint Table 4), suggesting that these spatially-correlated covariates of resistance are not 252 driving the spillover signal we observed. 253 254 Finally, we checked whether adjacency was too coarse a measure for interactions 255 between populations by replacing the dichotomous adjacency variable with a 256 continuous variable, the "commuting fraction", defined as the proportion of residents of a 257 pair of populations that commute to the other population (Figure 4). In 2 258 dataset/pathogen/antibiotic combinations, the confidence interval around the spillover 259 signal did not include zero (Supplemental Table 5). First, for E. coli resistance to 260 quinolones in the MarketScan/ResistanceOpen dataset, a modest commuting fraction 261 comparable to the effect of adjacency (10 89 ) was associated with a 0.9% decrease 262 (95% CI 0.5% to 1.3%) in use-resistance associations, compared to pairs of states with 263 no inter-state commuters. Second, S. pneumoniae resistance to macrolides in the 264 ECDC dataset, that same commuting fraction was associated with a 12% decrease 265 (95% CI 3% to 20%) in use-resistance associations. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint models of the use-resistance association, having on the order of 1% of interactions 275 between a control and intervention population was sufficient to attenuate the observed 276 effect of theoretical stewardship intervention by 50%, relative to a situation where the 277 two populations were completely isolated. Thus, in theory, even small numbers of 278 interactions could lead to a substantial underestimation of the potential reduction in 279 antibiotic resistance that would follow from a reduction in antibiotic use, compared to the 280 same reduction in use implemented in a completely isolated population. 281

282
In observational antibiotic use and resistance data in 3 pathogen-antibiotic combinations 283 across 3 datasets, we found that point estimates of the spillover effect varied from as 284 small as 1% to as large as 50%. In general, however, the confidence intervals on these 285 estimates were wide, encompassing zero in most cases. We therefore did not find 286 strong evidence to support our hypothesis, that spillover would have minimal effects at   The copyright holder for this preprint (which was not peer-reviewed) is the author/funder.  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/536714 doi: bioRxiv preprint