Allergic to Confounding: Leveraging Penicillin Allergy as an Instrumental Variable to Estimate the Causal Effect of Antibiotic Use on Resistance ================================================================================================================================================ * Yaki Saciuk * Daniel Nevo * Michal Chowers * Uri Obolski ## Abstract **Importance** Antibiotic resistance is a growing concern due to overuse and misuse of antibiotics. Quantitative assessments of the risk of future resistance caused by different antibiotics are limited, although essential for improved guidelines for responsible antibiotic prescription. **Objective** To employ instrumental variable (IV) analysis using reported penicillin allergy to estimate the causal effects of prior penicillin use on the development of antibiotic resistance. **Design** Retrospective cohort study spanning from 2013 to 2022, with a follow-up period extending to 12 months. **Setting** Primary Care **Participants** The study analyzed 36,351 adults who underwent an AMR test prompted by a positive urine culture, representing 2.5% of the initial cohort of eligible adults who purchased antibiotics in primary care. Exclusion criteria included recent antibiotic use, hospitalization, nursing home residence, being homebound, or purchasing multiple antibiotic types. **Main Outcome(s) and Measure(s)** Binary results of antibiotic susceptibility tests at follow-up times of up to 90, 180, 270, and 365 days. We focused on resistance to ampicillin, amoxicillin/clavulanic acid, piperacillin/tazobactam, and ciprofloxacin, using gentamicin as a negative control outcome. **Results** The rate difference (RD) of resistance due to previous penicillin use compared to other antibiotics within 12 months was 15.2% (95% CI: 9.8%, 21.4%) for amoxicillin/clavulanic acid, 11.8% (95% CI: 5.5%, 20.3%) for ampicillin, and 1.0% (95% CI: 0.3%, 1.9%) for piperacillin/tazobactam, with NNH values of 6.6, 8.5, and 98, respectively. Gentamicin, used as a negative control, showed no effect. Direct comparison of penicillins and quinolones within 180 days post-exposure revealed that penicillins increased resistance to amoxicillin/clavulanic acid by an RD of 14.9% (95% CI: 2.9%, 24.6%; NNH: 6.7), while quinolones increased resistance to ciprofloxacin by 34.7% (95% CI: 25.3%, 45.9%; NNH: 2.9). **Conclusions and Relevance** The study provides causal estimates of how prior use of penicillins and quinolones in outpatient settings impacts future antibiotic resistance, with implications for treatment guidelines and clinical practice. Additionally, it introduces a novel methodological approach to quantify antibiotic-induced resistance, which can be applied to other antibiotic groups known to cause allergic reactions. ## Introduction Antibiotics are vital in combating infections, yet their effectiveness is diminishing due to increased resistance, partly stemming from overuse and misuse.1,2 To enhance responsible prescription of antibiotics, guidelines have been developed that weigh the anticipated pathogens, resistance profiles, and related healthcare costs of different antibiotics. In recent years, there has been a growing focus on the potential of different antibiotics to induce future resistance. For example, quinolones, previously the primary choice for urinary tract infections and community-acquired pneumonia, are now often disfavored due to their high potential for resistance induction.3 A significant gap in current research is the lack of quantitative assessments of the risk of future resistance caused by different antibiotics. Randomized control trials, the benchmark for causal effect estimation, present logistical difficulties in this context, including limited patient follow up and the requirement for numerous studies to address various clinical scenarios and antibiotic combinations. Consequently, ranking of antibiotics by their collateral resistance ‘cost’ is often reached by expert opinion.4–6 Alternatively, cohort studies aimed at evaluating the effects of antibiotics are susceptible to numerous confounding mechanisms that undermine the causal validity of their results.7,8 Potential confounders, such as the indication for antibiotic use, illness severity at presentation, and patient demographics, present substantial challenges for retrospective studies in this field; researchers often struggle to access these variables. However, advances in causal inference theory have substantially improved our ability to discern causal effects from observational data. In particular, the use of instrumental variables (IVs) can provide reliable estimates of causal effects, when the appropriate methodology and falsification tests are applied,9,10 emulating natural-experiment-like conditions. Here, we propose to employ reported allergy to penicillin as such an IV. Reported penicillin allergy satisfies the essential conditions needed for a valid IV analysis. First, it causally influences prescription practices,11 leading physicians to opt for non-penicillin antibiotics upon presentation of penicillin allergy. Second, reported penicillin allergy also inherently does not directly affect antimicrobial resistance (AMR) of bacterial populations. By controlling for measured covariates which have the potential to confound the relationship between the IV (allergy to penicillin) and the outcome (AMR), we can further bolster the assumption of the causal nature of this association. Moreover, the setting of antibiotic susceptibility testing provides a falsification test for this assumption. We can utilize resistance to antibiotics outside the penicillin group as a negative control outcome (NCO),10,12,13 which could offer additional support for the validity of this assumption. The essential causal relationships assumed in our study are concisely depicted in the directed acyclic graph (DAG),14 in Figure 1. ![Figure 1:](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/08/20/2024.08.19.24312204/F1.medium.gif) [Figure 1:](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/F1) Figure 1: Causal relations in the instrumental variable (IV) framework can be depicted using Directed Acyclic Graphs (DAGs). The interest lies in determining the effect of the exposure (X) on the outcome (Y). This effect may not be identifiable—i.e., causally estimable—in the presence of unmeasured confounders (U1). However, employing a valid IV (Z) may enable the identification of this effect. First, Z must influence the exposure X, a condition known as the Relevance condition. Furthermore, Z should affect Y solely through X—a condition termed Exclusion Restriction. Measured confounders (C) and unmeasured variables (U) may impact Z, X and Y, potentially confounding their association. Adjusting for C blocks the Z <- C -> Y path, strengthening the third requirement—the Independence condition. However, the presence of an unmeasured variable (U2) could still compromise this condition and hinder the identification of the causal effect of X on Y. Dashed and dotted lines represent violations of the Exclusion Restriction and Independence conditions, respectively. A negative control outcome (N) should share a similar set of measured common causes (C) with Y and Z. If the unobserved common causes of Z and Y overlap with those of Z and N (U2), then N is considered “U-comparable” to Y. Hence, if the independence condition is met, we would naturally expect Z to not be associated with N, when adjusting for C. In this study, we utilize a large and comprehensive medical records database and leverage IV analysis to estimate the effects of penicillin use, compared to other antibiotics, on the Risk Difference (RD) and Number Needed to Harm (NNH) of subsequent resistance to penicillins in *E. coli.* We estimate these effects for different time periods since antibiotic consumption, adjusting for multiple variables to eliminate potential IV-outcome confounding and perform inverse probability of censoring weighting (IPCW) to mitigate selection bias. Finally, we perform negative control analyses to ascertain the validity of our IV. ## Study design and statistical analysis ### 1. Data Maccabi Healthcare Services (MHS), the second-largest non-profit health fund in Israel, covers 26.7% of the population (∼2.7 million in 2022) nationwide, with less than 1% disengagement. Membership in one of the four health funds is mandatory and non-excludable. MHS maintains a long-term centralized database of electronic medical records (EMRs), including extensive demographic and clinical data from both outpatients and hospitals. ### 2. Study design and participants In this retrospective cohort study, data were collected on adult members (aged≥18 years) of MHS who purchased antibiotics in primary care between January 1, 2013, and December 31, 2022. Participants were eligible to contribute multiple entries to the study within the specified timeframe, provided they had not purchased antibiotics within 12 months prior to the entry date or been hospitalized for at least one night during the same period. The exposed group consisted of all penicillin-group antibiotic purchases, while the control groups included purchases from the following antibiotic families: cephalosporins, macrolides, trimethoprim-sulfamethoxazole, quinolones, tetracyclines, fosfomycin, nitrofurantoin, aminoglycosides, and other beta-lactam antibiotics (Table 1). View this table: [Table 1:](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/T1) Table 1: Characteristics of the Primary Research Population at First Eligible Antibiotic Use (2013-2022), stratified by reported Penicillin Allergy. The exposures to penicillins were comprised of 46% amoxicillin, 39% amoxicillin/clavulanic acid, and the remaining 15% other penicillin derivatives. For both groups, the index date was set as the date of antibiotic purchase, and participants were followed until either a positive urine culture test that prompted an AMR test (the outcome) or patient exclusion occurred. Censoring, in the sense of excluding the follow-up of patients to observe their outcomes, was applied in cases of subsequent antibiotic use, hospitalization, death, or reaching the conclusion of the follow-up period, which extended until 365 days after the index date or until December 31, 2022 - whichever came first. Potential selection bias due to such censoring was addressed using IPCW, as described below. We excluded cases where more than one antibiotic family had been purchased to negate the influence of other antibiotic exposures. Furthermore, we excluded individuals residing in nursing homes and those who were homebound to mitigate potential confounding attributable to healthcare-associated infections or distinct prescription practices prevalent outside of primary care settings. This study design permitted the inclusion of multiple AMR tests for a single entry, if the specified exclusion and censoring criteria were fulfilled. ### 3. Procedures Individual-level demographic data for the study population included sex, age at the index date, socio-economic status, and social sector indication. The socioeconomic status of residential areas was assessed using coded Geographical Statistical Areas (GSAs), the smallest units recognized by the Israeli census. Assigned by Israel’s Central Bureau of Statistics, these GSAs are rated on a scale from 1 (lowest) to 10, based on multiple parameters. Data on chronic conditions were obtained from MHS’ automated registries, covering cardiovascular diseases,15 hypertension,16 diabetes,17 chronic kidney disease (CKD),18 and immunocompromised states. Allergy to penicillin was identified using the ICD-9 V14.0 code, with the status determined on the index date. In Israel, the procurement of antibiotics is strictly regulated, requiring a prescription for purchase and all transactions related to the acquisition of antibiotics by members of MHS to be systematically recorded and digitized within the EMR system. ### 4. Outcomes Our outcomes were the binary results of antibiotic susceptibility tests, with resistance coded as 1, at various follow-up times (90, 180, 270, and 365 days). We considered antibiotic resistance results within the penicillin group, comprising ampicillin, amoxicillin clavulanic acid and piperacillin/tazobactam, as well as ciprofloxacin within the fluoroquinolones group. Gentamicin, an aminoglycoside antibiotic, served as an NCO for the IV independence assumption10,12,13 (explained below). We chose gentamicin due to its lack of association with the penicillin/beta-lactam family, its relative frequency in AMR test panels, and its infrequent use in outpatient settings. The bacterial identification and antibiotic susceptibility testing were conducted using the automated VITEK 2 system, following Clinical and Laboratory Standards Institute guidelines. Intermediate results were considered as resistant. Our analysis focused exclusively on *Escherichia coli* isolates, which accounted for 71.6% of the total samples. Utilizing IV analysis, we estimated the risk difference of antibiotic resistance between individuals exposed to penicillins and those exposed to alternative antibiotic groups at the various post-antibiotic use follow-up times. ### 5. Statistical Methods #### IV framework We assert that antibiotic allergies, particularly penicillin,19 can serve as an IV in studying antibiotic use and future AMR. Although often self-reported, misdiagnosed, or resolved,20 perceived allergies influence doctors’ treatment choices, meeting the conditions for a valid IV. To be valid, IVs need to fulfill three essential conditions.9 The first, known as the Relevance condition, stipulates that the IV must be associated with exposure. The second, the Exclusion Restriction condition, asserts that the IV’s impact on the outcome occurs exclusively through exposure, and not via alternative paths. The third, referred to as the Independence condition, requires that the association between the IV and the outcome is not confounded by external factors. Provided that the linear model specification employed for estimation is valid, and under the assumption of a homogeneous treatment effect across the population, these conditions enable the identification and consistent estimation of the average treatment effect (ATE).21 #### Estimation We estimated the RD using Two Stage least squares (2SLS) regression.21–23 To uphold the Independence condition, the estimation was adjusted for potential confounders of the IV (allergy to penicillin) and the outcome (AMR) association, including sex, age, socio-economic status, social sector, calendar time, and comorbidities such as diabetes mellitus (DM), chronic kidney disease (CKD), immunosuppression, hypertension and cardiovascular diseases. To address the potential influence of healthcare-seeking behavior on both penicillin allergy reporting and AMR, we used proxies like the count of previous antibiotic purchases and urine culture tests, adjusting for these variables in the model. All continuous variables were modeled using restricted cubic splines with five knots.24 Potential confounders were measured at baseline. A DAG depicting the assumed causal relationships between penicillin allergy, AMR, and other relevant covariates is presented in Figure S1. Model diagnostics showed adequate fit, with all instrument coefficients and fitted values within the expected range of [0,1]. To evaluate the relevance condition, we used a “first-stage” ordinary least squares (OLS) regression, modeling the exposure as a function of the IV and confounding covariates. The condition’s validity was confirmed by examining the IV coefficient and t-statistic. The exposure variable, indicating the purchase of a penicillin antibiotic, was coded as binary. To maintain consistent directionality, the IV for penicillin allergy was coded as ‘1’ for non-allergic and ‘0’ for allergic individuals. Finally, selection bias may be introduced into the study as not all participants are subjected to resistance testing after a positive urine culture, and those who do may not be provided with relevant antibiotic panels. To address this potential bias and account for interdependent observations, we calculated 95% confidence intervals (CIs) using bootstrap resampling at the individual level, with sampling weights proportionally applied to the IPCW.25 IPCW compensates for censored patients by giving more weight to similar non-censored patients, thus increasing their representation. The weights were obtained from a logistic regression model that estimated the likelihood of undergoing AMR testing. The model was adjusted for sex, age, socio-economic status, social sector, antibiotic group, calendar time, and comorbidities (CKD, immunosuppression, and cardiovascular diseases). Censoring probabilities and their distribution are in Table S1. Due to potential temporal variations, we used average weights in the bootstrap procedure. All analyses were conducted using R version 3.6. The package *ivmodel* was used to estimate the IV models. ## Ethics and data declarations This study received approval from the Institutional Review Board (IRB) of Maccabi Healthcare Services. Given the retrospective nature of the study, the IRB waived the requirement for informed consent since all identifying participant information had been removed before the analysis. ## Data availability statement According to the Israel Ministry of Health regulations, individual-level data cannot be shared openly. Specific requests for remote access to de-identified community-level data should be referred to KSM, Maccabi Healthcare Services Research and Innovation Center. ## Role of the funding source The funders of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of this article. ## Results Within the primary research population, encompassing 1,469,624 individuals who received at least one outpatient antibiotic prescription from January 2013 to December 2022, a subset of 36,351 individuals underwent at least one AMR test panel within a following 12-month period. In total, 41,647 AMR test panels were administered. The distribution and selection criteria for the research population are further illustrated in Figure 2. To verify the IV Relevance condition, we computed the first-stage adjusted OLS regression, which yielded IV statistically significant coefficients of approximately 23% - indicating a strong link between the IV and the exposure (Table S2). ![Figure 2:](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/08/20/2024.08.19.24312204/F2.medium.gif) [Figure 2:](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/F2) Figure 2: Study population flow chart. We analyzed varying post-exposure follow-up times for two main reasons. First, the impact of antibiotic exposure on AMR diminishes over time, with most associations likely decaying after a year26, making estimates at different time points valuable. Second, identifying this time-dependent effect (or its absence, in the NCO) further validates our analyses. In the initial up to 90 days post-exposure, penicillin users exhibited significantly higher RDs for AMR within the penicillin group compared to other treatments. Specifically, amoxicillin/clavulanic acid, ampicillin, and piperacillin/tazobactam had RDs of 30.0% (95% CI: 19.7%, 42.4%), 21.2% (95% CI: 11.3%, 33.8%), and 2.5% (95% CI: 1.1%, 3.7%), respectively. The corresponding NNH values were 3.3, 4.7, and 40.8. These RDs consistently remained statistically significant across various post-exposure outcome time points and exhibited a temporal decline, aligning with expected AMR reduction as time elapsed. In contrast, gentamicin’s RD, serving as a negative control, was close to zero and remained statistically non-significant at all time points and did not show a decreasing trend in the RD point estimate (Figure 3 and Table 2). It is worth emphasizing that the estimate within each interval encapsulates a weighted average of the RD over the entire period, rather than an RD estimated at a specific time point. ![Figure 3:](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/08/20/2024.08.19.24312204/F3.medium.gif) [Figure 3:](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/F3) Figure 3: IV-Estimated Risk Differences for AMR Over Time Post-Exposure to Penicillins Compared to Other Antibiotics. AMR outcomes include amoxicillin/clavulanic acid (a), ampicillin (b), piperacillin/tazobactam (c), and gentamicin (d). Bars represent 95% confidence intervals (CIs), with the x-axis displaying days after exposure (90, 180, 270, and 365 days). View this table: [Table 2:](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/T2) Table 2: IV-Estimated Risk Differences of AMR Rates: A Comparison between Exposure to Penicillins and Other Antibiotics Across Varying Time Points Post-Exposure. Furthermore, to address the composition of the penicillin group, we conducted a sensitivity analysis: We repeated the analysis above when limiting the exposure to either amoxicillin/clavulanic acid, amoxicillin, or amoxicillin with other penicillins. The consistent results from this analysis are displayed in Table S3. We directly compared the effects of penicillin and quinolone exposure on AMR, focusing on a 180-day follow-up period. This follow-up duration was chosen to balance the trade-off between effect size and sample size, as the subgroup sizes became smaller for longer follow-up periods. Evaluating quinolone resistance, we considered ciprofloxacin susceptibility results. The findings, summarized in Table 3 and Figure 4, indicate prior penicillin use increases resistance, relative to prior to fluoroquinolone use in three penicillin antibiotics, albeit the ampicillin resistance outcome was not statistically significant. Conversely, prior quinolone exposure, relative to penicillin exposure, presented a pronounced influence on subsequent quinolone resistance (with a negative RD due to quinolones being the control group). The RD for amoxicillin/clavulanic acid was 14.9% (95% CI: 2.9%, 24.6%), for ampicillin 8.7% (95% CI: -0.7%, 18.9%), for piperacillin/tazobactam 2.4% (95% CI: 1.4%, 3.5%), for gentamicin -4.5% (95% CI: -10.1%, 0.6%), and for ciprofloxacin - 34.7% (95% CI: -45.7, -25.3). ![Figure 4:](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/08/20/2024.08.19.24312204/F4.medium.gif) [Figure 4:](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/F4) Figure 4: IV-Estimated Risk Differences for AMR Between Penicillins and Quinolones Within 180 Days After Exposure. Bars represent 95% confidence intervals (CIs), with the x-axis displaying tested AMR outcomes. View this table: [Table 3:](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/T3) Table 3: IV-Estimated Risk Differences of AMR: Direct Comparison of Penicillin vs. Quinolones Within 180 Days Post-Exposure We conducted pairwise comparisons between penicillin and each antibiotic group within the control arm, excluding quinolones, aminoglycosides, and other beta-lactams, within the 180-day window post-exposure (Table S4). The RDs varied somewhat. Comparisons with cephalosporins, another beta-lactam, yielded statistically insignificant RDs for penicillins. Similarly, comparisons of penicillins versus other groups for gentamicin AMR also yielded statistically insignificant RDs. ## Discussion This study exploits a unique phenomenon, namely purported penicillin allergy, to estimate the causal effects of prior antibiotic use on antibiotic resistance, in real-world primary care settings. Our analysis quantifies the impact of previous use of penicillins on resistance rates of future infections to the three penicillins. The largest effect was observed in amoxicillin/clavulanic acid, followed by ampicillin and finally piperacillin/tazobactam. In line with both biological reasoning and previous knowledge, all three outcomes demonstrated a gradual decrease in the RD for antibiotic resistance with each subsequent time point, reflecting the diminishing impact of prior antibiotic use.26 In contrast, gentamicin, the negative-control antibiotic, displayed no such effect nor trend, further affirming the validity of our analysis. In our additional analysis, we compared past usage of penicillin with that of quinolones on corresponding AMR tests within the 180-day post-exposure. The past use of penicillins was linked to resistance rates for penicillins; but this this association was lower than the association between past use of quinolones, and higher resistance to ciprofloxacin. This finding is consistent with previous evidence that quinolones usage leads to substantial increase in resistance.26–29 However, the current study allows provides an unbiased, causally justified comparison between the antibiotic classes. As with any study attempting to infer causality from observational data, this study has its limitations. The IV analysis performed here requires several specific assumptions. The Relevance condition has been confirmed through the coefficients obtained from the first-stage regression models. With respect to the Independence condition, Table 1 contrasts differences in demographic characteristics between individuals with and without a penicillin allergy. While the differences are not pronounced, they are present: the allergic population is predominantly female, older, and of higher socioeconomic status, suggesting a possible behavioral rather than biological basis for the reporting of penicillin allergy.19 The integrity of this condition was kept by adjusting for these variables. Moreover, the near-zero effect observed in the NCO, and the lack of temporal trend, further reassures that this condition is maintained. To estimate the average treatment effect using IVs, we assume that treatment effects are constant across the population. If this assumption is violated, the estimated effect can still be interpreted if the Monotonicity assumption holds9. This assumption, also known as the ‘no defiers’ assumption, means that no individual will consistently choose a treatment contrary to their IV assignment. In our study, it implies that individuals who would receive penicillin if not allergic would also avoid it if allergic. Under this assumption and additional statistical assumptions, the estimated effect is an average of Complier Average Causal Effects (CACEs) or Local Average Treatment Effects (LATEs) 9. The CACE/LATE represents the impact on individuals who comply with the treatment indicated by the IV, known as ‘*compliers*’. The study’s final limitation concerns the generalizability of its estimated effects. We used a comprehensive dataset that well-represents antimicrobial resistance of *E. coli* urine isolates in Israel. However, the study outcomes are likely influenced by the local variations in resistance patterns of pathogens during the period and location of the study (Figure S2). This is a common occurrence in infectious diseases, the dynamics of which are dictated by local biological and behavioral factors. Nevertheless, our study also provides researchers with a novel approach and detailed methodology to estimate the effects of antibiotic exposure on future resistance, to apply it to specific locations and time periods. In conclusion, our findings have implications for both guidelines and clinical practice, and methodological approaches to the epidemiology of antibiotic resistance. Clinically, our results provide causal estimates of the additional risk and NNH impacted by prior use of penicillins and quinolones in the outpatient settings, over different periods since exposure. Such estimates should also play a substantial role in the cost-benefit considerations when determining empiric antibiotic treatment guidelines. Specifically, penicillins induce substantial future resistance to amoxicillin/clavulanic acid, ampicillin, and to a lower extent piperacillin tazobactam. Moreover, penicillins, compared to quinolones, appear to induce lower respective AMR, thereby enhancing their utility in subsequent treatments. Methodologically, we propose a novel approach, built on an original IV, that opens a venue for further research on antibiotic groups inducing allergic reactions, such as cephalosporins, fluoroquinolones, and sulfonamides. ## Data sharing According to the Israel Ministry of Health regulations, individual-level data cannot be shared openly. Specific requests for remote access to de-identified community-level data should be referred to KSM, Maccabi Healthcare Services Research and Innovation Center. ## Declaration of interests The authors have no competing interests to declare. ## Funding This work was supported by the Israel Science Foundation (ISF 1286/21). ## Supporting information Appendix [[supplements/312204_file09.pdf]](pending:yes) ## Data Availability According to the Israel Ministry of Health regulations, individual-level data cannot be shared openly. Specific requests for remote access to de-identified community-level data should be referred to KSM, Maccabi Healthcare Services Research and Innovation Center. ![Figure5](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/08/20/2024.08.19.24312204/F5.medium.gif) [Figure5](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/F5) ![Figure6](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/08/20/2024.08.19.24312204/F6.medium.gif) [Figure6](http://medrxiv.org/content/early/2024/08/20/2024.08.19.24312204/F6) * Received August 19, 2024. * Revision received August 19, 2024. * Accepted August 20, 2024. * © 2024, Posted by Cold Spring Harbor Laboratory The copyright holder for this pre-print is the author. All rights reserved. The material may not be redistributed, re-used or adapted without the author's permission. ## References 1. 1.Talebi Bezmin Abadi A, Rizvanov AA, Haertlé T, Blatt NL. World Health Organization Report: Current Crisis of Antibiotic Resistance. BioNanoScience. 2019;9(4):778–788. doi:10.1007/s12668-019-00658-4 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/s12668-019-00658-4&link_type=DOI) 2. 2.Lesho EP, Laguio-Vila M. The Slow-Motion Catastrophe of Antimicrobial Resistance and Practical Interventions for All Prescribers. Mayo Clin Proc. 2019;94(6):1040–1047. doi:10.1016/j.mayocp.2018.11.005 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.mayocp.2018.11.005&link_type=DOI) 3. 3.Bush NG, Diez-Santos I, Abbott LR, Maxwell A. Quinolones: Mechanism, Lethality and Their Contributions to Antibiotic Resistance. Molecules. 2020;25(23):5662. doi:10.3390/molecules25235662 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3390/molecules25235662&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=33271787&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) 4. 4.Weiss E, Zahar JR, Lesprit P, et al. Elaboration of a consensual definition of de-escalation allowing a ranking of β-lactams. Clin Microbiol Infect. 2015;21(7):649.e1–649.e10. doi:10.1016/j.cmi.2015.03.013 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.cmi.2015.03.013&link_type=DOI) 5. 5.Madaras-Kelly K, Jones M, Remington R, Hill N, Huttner B, Samore M. Development of an Antibiotic Spectrum Score Based on Veterans Affairs Culture and Susceptibility Data for the Purpose of Measuring Antibiotic De-Escalation: A Modified Delphi Approach. Infect Control Hosp Epidemiol. 2014;35(9):1103–1113. doi:10.1086/677633 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1086/677633&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=25111918&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) 6. 6.World Health Organization. The Selection and Use of Essential Medicines: Report of the WHO Expert Committee, 2017 (Including the 20th WHO Model List of Essential Medicines and the 6th Model List of Essential Medicines for Children). World Health Organization; 2017. Accessed April 1, 2024. [https://iris.who.int/handle/10665/259481](https://iris.who.int/handle/10665/259481) 7. 7.Chatterjee A, Modarai M, Naylor NR, et al. Quantifying drivers of antibiotic resistance in humans: a systematic review. Lancet Infect Dis. 2018;18(12):e368–e378. doi:10.1016/S14733099(18)30296-2 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/S1473-099(18)30296-2&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) 8. 8.Chowers M, Zehavi T, Gottesman BS, Baraz A, Nevo D, Obolski U. Estimating the impact of cefuroxime versus cefazolin and amoxicillin/clavulanate use on future collateral resistance: a retrospective comparison. J Antimicrob Chemother. 2022;77(7):1992–1995. doi:10.1093/jac/dkac130 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1093/jac/dkac130&link_type=DOI) 9. 9.Angrist JD, Imbens GW, Rubin DB. Identification of Causal Effects Using Instrumental Variables. J Am Stat Assoc. 1996;91(434):444–455. doi:10.1080/01621459.1996.10476902 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1080/01621459.1996.10476902&link_type=DOI) 10. 11.Centers for Disease Control and Prevention. Adult Outpatient Treatment Recommendations. Accessed March 23, 2024. [https://www.cdc.gov/antibiotic-use/clinicians/adult-treatment-rec.html](https://www.cdc.gov/antibiotic-use/clinicians/adult-treatment-rec.html) 11. 12.Shi X, Miao W, Tchetgen ET. A Selective Review of Negative Control Methods in Epidemiology. Curr Epidemiol Rep. 2020;7(4):190–202. doi:10.1007/s40471-020-00243-4 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/s40471-020-00243-4&link_type=DOI) 12. 13.Danieli O, Nevo D, Walk I, Weinstein B, Zeltzer D. Negative Controls for Instrumental Variable Designs. Published online 2023. doi:10.48550/ARXIV.2312.15624 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.48550/ARXIV.2312.15624&link_type=DOI) 13. 14.Greenland S, Pearl J, Robins JM. Causal Diagrams for Epidemiologic Research. Epidemiology. 1999;10(1). [https://journals.lww.com/epidem/fulltext/1999/01000/causal\_diagrams\_for\_epidemiologic\_research.8.aspx](https://journals.lww.com/epidem/fulltext/1999/01000/causal\_diagrams_for_epidemiologic_research.8.aspx) 14. 15.Shalev V, Chodick G, Goren I, Silber H, Kokia E, Heymann AD. The use of an automated patient registry to manage and monitor cardiovascular conditions and related outcomes in a large health organization. Int J Cardiol. 2011;152(3):345–349. doi:10.1016/j.ijcard.2010.08.002 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.ijcard.2010.08.002&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=20826019&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) 15. 16.Weitzman D, Chodick G, Shalev V, Grossman C, Grossman E. Prevalence and Factors Associated With Resistant Hypertension in a Large Health Maintenance Organization in Israel. Hypertension. 2014;64(3):501–507. doi:10.1161/HYPERTENSIONAHA.114.03718 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1161/HYPERTENSIONAHA.114.03718&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=24958503&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) 16. 17.Chodick G, Heymann AD, Shalev V, Kookia E. The epidemiology of diabetes in a large Israeli HMO. Eur J Epidemiol. 2002;18(12):1143–1146. doi:10.1023/B:EJEP.0000006635.36802.c8 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1023/B:EJEP.0000006635.36802.c8&link_type=DOI) 17. 18.Coresh J, Turin TC, Matsushita K, et al. Decline in Estimated Glomerular Filtration Rate and Subsequent Risk of End-Stage Renal Disease and Mortality. JAMA. 2014;311(24):2518. doi:10.1001/jama.2014.6634 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1001/jama.2014.6634&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=24892770&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000337757000025&link_type=ISI) 18. 19.Albin S, Agarwal S. Prevalence, and characteristics of reported penicillin allergy in an urban outpatient adult population. Allergy Asthma Proc. 2014;35(6):489–494. doi:10.2500/aap.2014.35.3791 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.2500/aap.2014.35.3791&link_type=DOI) 19. 20.Stone CA, Trubiano J, Coleman DT, Rukasin CRF, Phillips EJ. The challenge of de-labeling penicillin allergy. Allergy. 2020;75(2):273–288. doi:10.1111/all.13848 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1111/all.13848&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) 20. 21.Angrist JD, Pischke JS. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton University Press; 2009. 21. 22.Brookhart MA, Wang PS, Solomon DH, Schneeweiss S. Evaluating Short-Term Drug Effects Using a Physician-Specific Prescribing Preference as an Instrumental Variable: Epidemiology. 2006;17(3):268–275. doi:10.1097/01.ede.0000193606.58671.c5 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/01.ede.0000193606.58671.c5&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=16617275&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000236926500012&link_type=ISI) 22. 23.Wang PS, Schneeweiss S, Avorn J, et al. Risk of Death in Elderly Users of Conventional vs. Atypical Antipsychotic Medications. N Engl J Med. 2005;353(22):2335–2341. doi:10.1056/NEJMoa052827 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1056/NEJMoa052827&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=16319382&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000233574600005&link_type=ISI) 23. 24.Harrell FE. In: Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. Second edition. Springer series in statistics. Springer; 2015:24–28. 24. 25.Howe CJ, Cole SR, Lau B, Nirvanic S, Eron JJ. Selection Bias Due to Loss to Follow Up in Cohort Studies: Epidemiology. 2016;27(1):91–97. doi:10.1097/EDE.0000000000000409 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/EDE.0000000000000409&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) 25. 26.Baraz A, Chowers M, Nevo D, Obolski U. The time-varying association between previous antibiotic use and antibiotic resistance. Clin Microbiol Infect. 2023;29(3):390–e1. 26. 27.Cherny SS, Nevo D, Baraz A, et al. Revealing antibiotic cross-resistance patterns in hospitalized patients through Bayesian network modelling. J Antimicrob Chemother. 2021;76(1):239–248. doi:10.1093/jac/dkaa408 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1093/jac/dkaa408&link_type=DOI) 27. 28.Cherny SS, Chowers M, Obolski U. Bayesian network modeling of patterns of antibiotic cross-resistance by bacterial sample source. Commun Med. 2023;3(1):61. doi:10.1038/s43856-02300289-7 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1038/s43856-02300289-7&link_type=DOI) 28. 29.Gottesman BS, Carmeli Y, Shitrit P, Chowers M. Impact of Quinolone Restriction on Resistance Patterns of *Escherichia coli* Isolated from Urine by Culture in a Community Setting. Clin Infect Dis. 2009;49(6):869–875. doi:10.1086/605530 [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1086/605530&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=19686074&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F20%2F2024.08.19.24312204.atom) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000269145100006&link_type=ISI)