A New Approach to the Analysis of Antibiotic Resistance Data from Hospitals

Abstract

We aimed to develop a new approach to the analysis of antimicrobial resistance data from the hospitals, which allows simultaneous analysis of both individual- and population-level determinants of bacterial resistance. This was a retrospective cohort study that included adult patients who stayed in the hospital >2 days. We analyzed data using shared frailty Cox models and tested our approach using a priori hypotheses based on biology and epidemiology of antibiotic resistance. For gram-negative bacteria, the use of the major selecting antibiotic by an individual was the main risk factor for acquiring resistant species. Hazard ratios (HRs) were strikingly high for ceftazidime-resistant Enterobacter species (HR=11.17; 95% confidence interval [CI]: 5.67–22.02), ciprofloxacin-resistant Pseudomonas aeruginosa (HR=4.41; 95% CI: 2.14–9.08), and imipenem-resistant P. aeruginosa (HR=7.92; 95% CI: 4.35–14.43). Ward-level use was significant for vancomycin-resistant enterococci (VRE) (HR=1.40; 95% CI: 1.07–1.83) and for imipenem-resistant P. aeruginosa (HR=1.40; 95% CI: 1.08–1.83). Previous incidence of infection in the same ward increased the risk of acquiring methicillin-resistant Staphylococcus aureus (HR=1.22; 95% CI: 1.15–1.30) and VRE (HR=1.53; 95% CI: 1.38–1.70). Our results were consistent with our hypotheses and showed that combining population- and individual-level data is crucial for the exploration of antimicrobial resistance development.

Introduction

Antimicrobial resistance is a major public health threat. This threat is particularly amplified in hospitals, and the past few decades have seen an alarming increase in the incidence of antibiotic-resistant hospital infections (ARHI). The evidence that heavy use of antimicrobial agents promotes dissemination of resistant bacteria is overwhelming; nevertheless, apart from avoiding unnecessary use, the best ways to extend the life of antibiotics are not well understood.^20,32

The majority of clinical studies on antibiotic resistance evaluate the effect of individual risk factors, such as underlying illness, length of stay, and antimicrobial use, on the individual's risk of developing a resistant infection. While individual-level approaches are traditionally the gold standard in noncommunicable disease epidemiology, they do not involve the transmissible aspect of ARHI. Although this approach has provided considerable information about the effect of antibiotic use on ARHI, it suffers from major limitations.

First, individual-level analyses may incompletely capture the population-level effects of antimicrobial agents.^5,21 Population-level effects are the consequences on individual risk of ARHI induced by the use of antimicrobial drugs by other members of a population.²¹ A patient in a unit with heavy use of a particular antibiotic may have a greater exposure to bacteria resistant to that antibiotic, even if the patient herself has not used it. Other types of population-level effects such as infection control measures and prevalence of resistant organisms within a hospital or a unit may also affect an individual's risk of ARHI.⁷ Such population-level effects disturb the ecological balance between antibiotic-resistant and antibiotic-susceptible strains and change the probability of unexposed patients acquiring a resistant strain.^3,19–21

Second, clustering affects statistical methods for distinguishing signal and noise. Individuals' outcomes are correlated: if one person is infected, it is more likely that a patient nearby in space and time is infected.¹⁰ Traditional statistical approaches generally assume that individual outcomes are uncorrelated and these approaches can lead to misleading conclusions, especially in settings where incidence and prevalence fluctuate substantially (e.g., when there are outbreaks).

Third, individual-level antibiotic effects may be confounded by population-level variables. Observed individual-level associations between use and ARHI may be, in part, confounded by population-level effects.^4,7,10 Therefore, for each pathogen-antibiotic combination, it is important to differentiate and estimate the direct and indirect effects of antibiotic consumption and transmission. Such an understanding can provide clues about how a particular antibiotic promotes resistance and helps in choosing appropriate prevention strategies.

We developed a new analytical approach that allowed us to combine and, therefore, explore the effect of individual- and population-level antibiotic use simultaneously. Additionally, we assessed if the risk of infection with a particular resistant species in an individual was increased during periods of increased incidence among other individuals; a pattern that suggests the importance of transmission for that species. Hypotheses were formulated and tested in a framework designed to separate individual- and population-level risk factors across distinct classes of organisms, according to the underlying biology and epidemiology of resistance.

Patients and Methods

We performed a retrospective cohort study using automated healthcare information from administrative, pharmacy, and laboratory databases of a 400-bed tertiary care hospital in Salt Lake City. This analysis was carried out with deidentified data and in full compliance with the Health Insurance Portability and Accountability Act of 1996 in the United States.

Study population

The study cohort consisted of patients who had been admitted to medical and surgical services from January 1995 to December 2000. All patients who were older than 15 years and stayed in the hospital >2 days were included in the study. The final dataset included each patient day as an observation, which allowed the use of time-varying adjustment for both exposure and confounding variables.

At the time of data collection, there was not an active screening policy for any of these microorganisms and no isolation precautions were in use for patients infected with these organisms.

Outcomes of interest

Main outcomes of interest for this study were isolation of five resistant bacterial species from clinical specimens. These were methicillin-resistant Staphylococcus aureus (MRSA), vancomycin-resistant enterococci (VRE), ceftazidime-resistant Enterobacter spp., ciprofloxacin-resistant Pseudomonas aeruginosa, and imipenem-resistant P. aeruginosa.

Exposures of interest and confounders

Variables were defined to account for individual history of infection with the susceptible form of the organism, exposure to antibiotics both at the individual and group level, as well as exposure to other infected patients. We chose the antibiotic or the antibiotic group, which exerted main selection pressure for the species that were the outcomes in this study, as our main exposure variables. These were vancomycin for VRE, β-lactams with antistaphylococcal activity for MRSA, carbapenems for imipenem-resistant P. aeruginosa, quinolones for ciprofloxacin-resistant P. aeruginosa, and third-generation cephalosporins (3GCs) for ceftazidime-resistant Enterobacter spp.

To capture the population-level effects of antibiotic use, we evaluated the recent total utilization in the wards. To compute this population-level variable, we divided the total person-days of use in a ward in the preceding 30 days by the total person-days in that ward for the same period. The final variable was expressed as the percentage of patient-days, on which the antibiotic was used in the last 30 days in the same ward. The use of all antibiotics other than the selecting was considered as potential confounders and both individual- and ward-level consumption were used to adjust for confounding. The isolation of a susceptible strain of the same bacteria, previous to the index culture date, was considered to be a confounder and an effect modifier. Diagnosis-related group (DRG) weights were used to adjust for the confounding effect of underlying disease. To capture the effect of transmission, we calculated the total number of new infections in a ward for each day in the preceding 30 days; these estimates were also adjusted for the total patient-days in the same ward during the same period. The metric for the final variable was (the natural log of ) the number of new cases per 100 patient-days.

Predicted associations based on biological mechanisms

Our predictions and the mechanisms underlying these predictions are summarized in Table 1.

Table 1.

Biological and Epidemiological Mechanisms Underlying Our Predictions

Individual use of the main selecting antibiotic (antibiotic A) confer resistance when

(i) Individuals who are infected with the susceptible form of the organism receive the selecting antibiotic AND resistance is conferred by a single mutation.

(ii) The selecting agent suppresses the endogenous flora that otherwise tends to block acquisition of the resistant organism AND individuals are exposed to infectious sources of the resistant organism during or shortly after the period of treatment.

(iii) Individuals are colonized by both resistant AND susceptible forms of bacteria AND treatment increases the load of resistant organisms by killing the competitive susceptible strains.

Individual use of other antibiotics (other than antibiotic A) confer resistance when

1. The individuals infected with A-susceptible strains receive the other antibiotic (B), and resistance to both A and B is conferred by the same mutation.

2. B suppresses the endogenous flora AND individuals are exposed to infectious sources of the A- resistant organism during or shortly after the treatment period.

3. Strains resistant to A tend also to be resistant to B, and treatment shifts the balance of colonizing organisms from mostly susceptible to mostly resistant.

4. Treatment with B leads to resistance to A by inducing hydrolyzing enzymes, target modifications, or efflux pumps.

Ward-level use of antibiotics confer resistance when

One of the mechanisms i–iii (for the same antibiotic) or 1–4 are present (for other antibiotics) AND secondary transmission of the resistant pathogen occurs at an appreciable frequency.

Effect of individual use of antibiotics on the ARHI risk

For many antibiotic-bacteria combinations, the previous use of antibiotics will cause an individual to be at a higher risk of resistant infection, even after adjusting for other individual risk factors and population-level use of the antibiotic. For the following, we refer to species S, resistant to antibiotic A, as the outcome of interest, and we consider how individual use of antibiotic A, or of other antibiotics, affects the risk of infection with A-resistant species S. We hypothesized that use of antibiotic A will increase the risk of A-resistant S infection when (i) individuals who are infected with the A-susceptible S are treated with antibiotic A AND resistance to A in species S is conferred by a single mutation; or (ii) antibiotic A suppresses the endogenous flora that otherwise tends to block acquisition of A-resistant S AND individuals are exposed to infectious sources of the A-resistant S during or shortly after the period of treatment; or (iii) individuals are colonized by both resistant AND susceptible forms of S AND treatment with A increases the load of resistant organisms by killing the competitive susceptible strains. Under scenarios (i) and (iii), use of the selecting antibiotic is expected to be an even stronger risk factor for resistant infection if the patient has previously had clinical infection with a drug-susceptible strain of the same organism.

Individual use of antibiotics other than A may also predispose to infection with A-resistant S. Such an association between infection with A-resistant S and individual use of antibiotics other than A is expected (1) if individuals infected with A-susceptible S receive the other antibiotic, and resistance to both A and the other antibiotic is conferred by the same mutation; (2) if other drug suppresses the endogenous flora AND individuals are exposed to infectious sources of the A-resistant organism during or shortly after the treatment period; (3) if strains resistant to A tend also to be resistant to the other drug, and treatment shifts the balance of colonizing organisms from mostly susceptible to mostly resistant; or (4) if treatment with the other antibiotic leads to resistance to A by inducing hydrolyzing enzymes, target modifications, or efflux pumps.

The association between individual use and resistant infection will be lowest for bacteria, in which resistance requires acquisition of one or several new genes such as methicillin resistance in Staphylococcus aureus and vancomycin resistance in Enterococcus spp. and transmission by healthcare workers is the main route for spread of resistance.¹³

Ward-level use of antibiotics on the ARHI risk

Antibiotic use promotes resistance also at the population level. For many antibiotic-bacteria combinations, an individual's risk of resistant infection will increase with the total use of antibiotic in the hospital or ward, during the period preceding their infection. This hypothesis is justified by previous ecological studies^14,18 We hypothesized that an association between population-level use and infection with a resistant organism would be strongest when one or more of the above-mentioned scenarios at the individual level are important AND secondary transmission of the resistant pathogen occurs at an appreciable frequency.

Effect of prior incidence of infection on ARHI risk

For antibiotic-bacteria combinations, such as VRE and MRSA, the development of resistance is mediated by the acquisition of one or several new genes; although these new resistance mechanisms arose and spread in large populations under antibiotic selection pressure, they are unlikely to occur de novo in any single person because of the multiple changes involved.^26,31 Organisms with these types of resistance must be acquired, generally as a consequence of cross transmission.^2,8,23,26 An association between prior incidence and infection with these pathogens is predicted to be strongest when pathogen is a common colonizer of human skin and/or gut AND transmission occurs with high frequency. For gram-negative infections, a combination of cross transmission and infection by the patient's own endogenous flora is believed to be responsible for most ARHI, suggesting a lower but perhaps nonzero association between prior incidence and risk.

Statistical analysis

Independent null hypotheses were developed for each species regarding the effects of antibiotic use and also for the effects of transmission. Therefore, we used two different sets of shared frailty Cox proportional hazards models for each species to assess the effect of antibiotic use and transmission. In models used to assess the effect of antibiotic use, we included age, gender, and DRG weights as baseline characteristics; individual- and ward-level use of antibiotics and prior isolation of a susceptible strain from the patient as time-varying variables. Antibiotic use data included vancomycin, β-lactamase inhibitor combinations, 3GCs, carbapenems, second-generation cephalosporins, aminoglycosides, macrolides, antistaphylococcal β-lactams, tetracylines, metronidazole, and quinolones, which were classified as the main selecting agent or confounder, depending on the species. For estimating the effect of transmission, previous incidence was the main exposure variable modeled as a time-varying variable. These transmission models also included age, gender, and DRG weights as baseline characteristics; individual-level antibiotic use and previous isolation of the susceptible strain of the bacteria. Infection with each resistant bacteria group was a mutually exclusive outcome and separate models were run for each group of bacteria.

In hospitals, patients staying in the same ward have some shared experiences, such as exposure to the same bacteria, same healthcare workers, similar antibiotics, and similar underlying diseases. This will cause a propensity or shared frailty among those patients, in the same ward, toward the occurrence of an event (ARHI in our case).^11,25 Furthermore, some gradual antibiotic policy changes occur in the hospitals, which take place over a long time in contrast to the day-to-day changes in the use. Since our dataset covered 6 years and our preliminary analyses showed long-term changes in antibiotic consumption, the Cox model assumption of independent timing of events could be violated. For these reasons, Cox models were run using shared frailty options, in which a multiplicative frailty was estimated for each year in each ward. For variable selection, we used a change-in-estimate method, requiring a 10% change in the hazard ratio (HR) of the main exposure variable (individual use of the selecting antibiotic), for a variable to be considered as a confounder. We tested interactions for the variables that we defined as effect modifiers a priori and used significance testing of the product term coefficient in deciding to keep them in the model. All analyses were performed using STATA 9.2 (StataCorp.).

Results

During the 6-year study period, there were 43,284 admissions that belonged to 29,291 patients, who contributed 382,785 person-days of observation time to the study cohort. Important characteristics of the study population are shown in Table 2.

Table 2.

Important Characteristics of the Study Population

Characteristic	Study population n=29,291
Age, mean years±SD	53±19
Gender, % for male/female	51.5/48.5
Length of stay, median days (min–max)	5 (3–190)
Patients with >1 hospitalization (%)	2,933 (10)
Most common diagnosis-related groups (%)
Major joint and limb reattachment procedure	1,965 (6.7)
Craniotomy	1,495 (5.1)
Major small and large bowel procedures	648 (2.2)
Intracranial hemorrhage and stroke with infarct	572 (2)
Respiratory tract infections	513 (1.8)
Location in the hospital (%)
Surgical wards	8,479 (30)
Internal medicine wards	6,034 (21)
Neurology wards	6,609 (23)
Special care wards^a	8,169 (28)
Number of patients receiving antibiotics >2 days (%)
Anti-staphylococcal β-lactams	850 (3.0)
Vancomycin	2,688 (9.2)
First-generation cephalosporins	6,281 (21.4)
Third-generation cephalosporins	3,127 (10.7)
Aminoglycosides	431 (1.47)
Quinolones	3,339 (11.4)
Carbapenems	1,541 (5.3)
β-Lactam-β-lactamase inhibitor combination	1,003 (3.4)

Includes burn, bone marrow transplant, and intensive care units.

Effect of individual use of the selecting antibiotics on the ARHI risk

Observed associations (Table 3) from our analyses were mostly consistent with our predictions that we based on a priori hypotheses that we summarized, as shown in Table 1. For gram-negative bacteria, individual antibiotic use was significantly associated with resistance; HR was 2.30 (95% confidence interval [CI]: 1.42–3.73) for ceftazidime-resistant Enterobacter spp., 2.08 (95% CI: 1.17–3.69) for ciprofloxacin-resistant P. aeruginosa, and 3.02 (95% CI: 1.8–5.06) for imipenem-resistant P. aeruginosa. HRs increased strikingly if the selecting antibiotic was used after isolation of the susceptible strain of the same bacteria (Table 3); HR was 11.17 (95% CI: 5.67–22.02) for ceftazidime-resistant Enterobacter spp., 4.41 (95% CI: 2.14–9.08) for ciprofloxacin-resistant P. aeruginosa, and 7.92 (95% CI: 4.35–14.43) for imipenem-resistant P. aeruginosa. For the two gram-positive bacteria, neither individual use of the selecting antibiotic nor prior recovery of the susceptible strain was associated with resistance.

Table 3.

Hazard Ratios Showing the Association Between the Acquisition of a Resistant Strain and Use of the Main Selecting Antibiotic at the Individual and Ward Level

	Selecting antibiotic
Type of organism	Use by a patient without previous susceptible culture HR (95% CI)	Use by a patient after a susceptible culture HR (95% CI)	Use in the ward HR (95% CI)
Gram-positive
MRSA	0.85 (0.39–1.84)	0.73 (0.32–1.67)	1.39 (0.87–2.22)
VRE	1.37 (0.87–2.17)	1.70 (0.81–3.57)	1.40 (1.07–1.83)
Gram-negative
Ceftazidime-resistant Enterobacter spp.	2.30 (1.42–3.73)	11.17 (5.67–22.02)	1.24 (0.98–1.57)
Ciprofloxacin-resistant Pseudomonas aeruginosa	2.08 (1.17–3.69)	4.41 (2.14–9.08)	1.05 (0.81–1.36)
Imipenem-resistant P. aeruginosa	3.02 (1.80–5.06)	7.92 (4.35–14.43)	1.40 (1.08–1.83)

Ward-level HR shows the increase in the hazard of having a resistant isolate for an additional 10% increase in use, during the last 30 days in the ward where the patient was hospitalized. All models were adjusted for age, gender, DRGs, year, location in the hospital, individual- and ward-level use of antibiotics, and prior isolation of sensitive strain.

CI, confidence interval; DRG, diagnosis-related group; HR, hazard ratio; MRSA, methicillin-resistant Staphylococcus aureus; VRE, vancomycin-resistant enterococci.

Effect of individual use of the antibiotics other than the selecting agent

Effects of individual use of antibiotics other than the selecting agent are shown in Table 4. The use of carbapenem and quinolone (which are both active against MSSA and wide spectrum) was associated with increased risk of MRSA infection, HRs were 1.45 (95% CI: 1.07–1.97) and 1.77 (95% CI: 1.30–2.42), respectively. HR for 3GC use, which is a known risk factor for VRE infection, was 1.80 (95% CI: 1.19–2.72). Carbapenem and quinolone use decreased the risk for ceftazidime-resistant Enterobacter spp. infections, HR was 0.47 (95% CI: 0.27–0.83) for carbapenems and 0.26 for quinolones (95% CI: 0.13–0.54). HR for the use of aminoglycosides was 1.95 (95% CI: 1.17–3.25) for imipenem-resistant P. aeruginosa infection, an expected finding due to well-known aminoglycoside resistance in the multidrug-resistant Pseudomonas species. We also observed an unexpected HR for vancomycin use in ciprofloxacin-resistant P. aeruginosa infections that showed an increased risk (HR=1.73, 95% CI: 0.95–3.04).

Table 4.

Association Between the Acquisition of a Resistant Strain and the Use of Antibiotics Other Than the Selecting Agent

	Antibiotics other than the main selecting agent
Type of organism	Antibiotics	Use by the patient HR (95% CI)	Use at the ward HR (95% CI)
Gram-positives
MRSA	Carbapenems	1.45 (1.07–1.97)	1.19 (0.99–1.43)
	Quinolones	1.77 (1.3–2.42)	0.95 (0.76–1.20)
VRE	Third-generation cephalosporins	1.80 (1.19–2.72)	1.36 (1.02–1.80)
Gram-negatives
Ceftazidime-resistant Enterobacter spp.	Carbapenems	0.47 (0.27–0.83)	1.32 (1.03–1.69)
	Quinolones	0.26 (0.13–0.54)	0.78 (0.55–1.12)
	BLI combinations	1.61 (0.81–3.22)	2.24 (1.12–4.49)
Ciprofloxacin-resistant P. aeruginosa	Vancomycin	1.73 (1.03–2.90)	1.06 (0.84–1.33)
Imipenem-resistant P. aeruginosa	Aminoglycosides	1.95 (1.17–3.25)	1.20 (0.92–1.57)

BLI, β-lactamase inhibitor.

Ward-level use of antibiotics on the ARHI risk

Ward-level use of the selecting antibiotic was associated with an increased risk for VRE (HR=1.40, 95% CI: 1.07–1.83) and for imipenem-resistant P. aeruginosa infections (HR=1.40, 95% CI: 1.08–1.83) (Table 3). For antibiotics other than the selecting agent (Table 4), a statistically significant association was observed between the ward-level use of 3GCs and VRE (HR=1.36, 95% CI: 1.02–1.80), carbapenems and ceftazidime-resistant Enterobacter spp. (HR=1.32, 95% CI: 1.03–1.69), and β-lactamase-inhibitor combinations (HR=2.24, 95% CI: 1.12–4.49) and ceftazidime-resistant Enterobacter spp.

Effect of prior incidence of infection on ARHI risk

Results of multivariable models exploring the effect of previous incidence of a resistant infection on patients' risk of acquiring a new infection with the same species are shown in Figure 1. For both MRSA and VRE, the incidence during the previous 30 days significantly increased the risk of acquiring a new infection; HR was 1.22 (95% CI: 1.15–1.30) for MRSA and 1.53 (95% CI: 1.38–1.70) for VRE. For gram-negatives previous incidence was positively associated with a patient's risk, but in no case was the association statistically significant.

FIG. 1.

Hazard ratios showing the effect of previous incidence of infection on a subsequent resistant infection. Filled squares represent effect estimates and lines show 95% confidence limits.

Discussion

Effective interventions are required to control the escalating problem of ARHI. It is increasingly recognized that the methodology of observational studies in this area is inadequate and led to conflicting results and recommendations.^6,7,28 Major methodological problems in these studies are the difficulty in accounting for indirect effects of antibiotic use and infectious nature of ARHI.^8,10,21 In this study, we developed a method that would allow estimation of the indirect and direct effects of antibiotic consumption, simultaneously. We used a similar approach to estimate the effect of transmission, using previous incidence in the ward as a proxy. We hypothesized that, for organisms in which resistance emerges due to mutational events and selection by drugs, individual antibiotic use and prior recovery of a susceptible strain would be predominant risk factors, and for organisms in which acquisition of new strains was required, group antibiotic use and hospital incidence of the resistant organism would take on a greater importance.^4,32

Our findings suggest that a patient's previous use of carbapenems and quinolones will increase the risk of acquiring P. aeruginosa infection resistant to these antibiotics, and ceftazidime use will significantly increase the risk of infection with a ceftazidime-resistant Enterobacter spp. These findings are consistent with prior studies^12,15 but are strengthened in this case because the associations remained significant and large, even after controlling for individual and ward-level consumption of 10 other antibiotics/antibiotic classes and other confounders. Moreover, we illustrated that if patients had a susceptible isolate previous to using these antibiotics, the risk of acquiring a resistant strain of the same bacteria increased remarkably. The highest risk was associated with ceftazidime use, in patients who had a previous susceptible strain of Enterobacter spp. (HR=11.17, 95% CI: 5.67–22.02). This finding probably demonstrates well-recognized selection pressure by 3GCs on mutant Enterobacter strains with hyperproduction of AmpC β-lactamase.²⁷ Molecular mechanisms of carbapenem and quinolone resistance in P. aeruginosa are well known, and the emergence of resistance after or during therapy has also been described.^15,22 Our markedly increased HRs indicate that emergence of resistance during treatment of P. aeruginosa infections with these antibiotics can be a serious therapeutic problem.

We found a protective effect of carbapenems and quinolone use in ceftazidime-resistant Enterobacter spp.; this finding is probably due to the efficacy of these antibiotics against β-lactamase-producing bacteria. Although it is surprising in the current context, where extended-spectrum β-lactamases (ESBL) carrying Enterobacter spp. possess changes that confer high-level resistance to quinolones, our study included cultures from a period (1995–2000) when ESBL carrying bacteria were very rare and ceftazidime resistance was mostly due to chromosomal β-lactamases. Our most unexpected findings was an increased risk (HR=1.73, 95% CI: 1.03–2.90) for vancomycin use in ciprofloxacin-resistant P. aeruginosa infections. We believe this is a spurious finding either due to chance or the confounding effect of the severity of the underlying disease, that is, patients who had more severe disease were more likely to receive vancomycin and also were at a higher risk infection with resistant P. aeruginosa.

Indirect effects of antibiotic use can be defined as the enhancement of risk for acquiring a resistant organism due to use of antibiotics by other individuals in a population.²¹ Measuring the indirect effects requires aggregating and averaging data on antibiotic use at the population level and it measures a different construct than individual-level use.²⁹ The necessity of combining population-level information with the individual-level information has been previously recognized in the infectious diseases area but mostly in the context of transmission dynamics.¹⁷ There have been only a few studies combining both levels of information for investigating antibiotic resistance.^18,24,30 In contrast to these studies, we chose to use more detailed information in our statistical models, which included multiple antibiotics and interaction with previous susceptible culture and calculated ward-level use for each patient-day in the hospital. Our results showed some important ward-level associations. The significant effect of 3GCs on VRE infections was consistent with our predictions since both individual-level effects and transmission were strong. Ward-level consumption of vancomycin was also significantly associated with VRE, even though it did not show a statistically significant individual-level effect. This may be due to a very strong transmission effect shown for this bacteria.^23,26 Particularly, vancomycin use might increase an individual's infectiousness with VRE (hence leading to a ward-level effect) but have a modest (unmeasurable) effect on the individual's own risk. Among gram-negatives ward-level antibiotic use was significantly associated with increased risk of imipenem-resistant P. aeruginosa and ceftazidime-resistant Enterobacter spp., which have also a strong tendency for transmission in the hospitals.

Molecular typing studies have shown that for MRSA and VRE, new cases arise by patient-to-patient transmission, rather than by acquisition during treatment. It is known that proximity to other patients who are colonized or infected with VRE is the single most important risk factor for VRE acquisition.⁹ In line with these findings, our results did not show a significant association between the individual use of the selecting antibiotic or harboring a susceptible strain with MRSA and VRE. Our exploration of the effect of previous incidence as a proxy for transmission was in line with our predictions and showed a significant association in gram-positive infections even after controlling for other risk factors (Fig. 1). Recently, efforts to account for the colonization pressure, as a predisposing factor, have increased, but classical regression methods at the individual-level have been utilized.^1,16 Our approach is unique and we believe will guide future studies trying to combine a measure of selection pressure with other population-level confounders.

Our study had a number of limitations. One limitation was that we had to evaluate the effect of antibiotics and transmission in separate models because according to our a priori model of causation (for which we used directed acyclic graphs—to conceptualize), ward-level use of antibiotics was a time-varying confounder in the causal pathway between the previous incidence in the ward and individual resistance. Another limitation is that our data have become old during the course of our research given the extended time period of our efforts. However, given that this study was a methodological work on how to analyze ARHI data, we are confident that the method we have developed is even more relevant today, considering the current condition of antibiotic resistance, that is, there are additional mechanisms of resistance, increased risk of horizontal transmission, and multiple drug resistance. All of these factors make the indirect effects of antibiotic use more important than ever and elucidate the importance of accounting for them in data analysis.

To conclude, our study was a comprehensive effort to account for both direct and indirect effects of antibiotic use, as well as the effects of transmission in a hospital. We included most of the relevant variables in our analysis especially all significant antibiotic classes and used five different resistant bacteria as the outcome measures, which allowed comparisons across different bacteria and antibiotics. Most important of all, instead of hunting for risk factors, our analyses were a test of our a priori hypotheses. Our results were mostly consistent with our hypotheses and showed the need to combine population- and individual-level data. However, integration of transmission and antibiotic consumption data at both levels requires further exploration of analytic methods such as G-estimation and dynamic models.

Footnotes

Disclosure Statement

No competing financial interests exist.

References

Ajao

A.O.

, Harris

A.D.

, Roghmann

M.C.

, Johnson

J.K.

, Zhan

, McGregor

J.C.

, and Furuno

J.P.

. 2011. Systematic review of measurement and adjustment for colonization pressure in studies of methicillin-resistant Staphylococcus aureus, vancomycin-resistant enterococci, and clostridium difficile acquisition. Infect. Control. Hosp. Epidemiol., 32:481–489.

Austin

D.J.

, and Anderson

R.M.

. 1999. Transmission dynamics of epidemic methicillin-resistant Staphylococcus aureus and vancomycin-resistant enterococci in England and Wales. J. Infect. Dis., 179:883–891.

Austin

D.J.

, Kakehashi

, and Anderson

R.M.

. 1997. The transmission dynamics of antibiotic-resistant bacteria: the relationship between resistance in commensal organisms and antibiotic consumption. Proc. Biol. Sci., 264:1629–1638.

Baquero

2011. The 2010 Garrod Lecture: the dimensions of evolution in antibiotic resistance: ex unibus plurum et ex pluribus unum. J. Antimicrob. Chemother., 66:1659–1672.

Baquero

, and Coque

T.M.

. 2011. Multilevel population genetics in antibiotic resistance. FEMS Microbiol. Rev., 35:705–706.

Baquero

, Coque

T.M.

, and de la Cruz

. 2011. Ecology and evolution as targets: the need for novel eco-evo drugs and strategies to fight antibiotic resistance. Antimicrob. Agents Chemother., 55:3649–3660.

Bonten

M.J.

, Austin

D.J.

, and Lipsitch

. 2001. Understanding the spread of antibiotic resistant pathogens in hospitals: mathematical models as tools for control. Clin. Infect. Dis., 33:1739–1746.

Bonten

M.J.

, and Mascini

E.M.

. 2003. The hidden faces of the epidemiology of antibiotic resistance. Intensive Care Med., 29:1–2.

Bonten

M.J.

, Slaughter

, Ambergen

A.W.

, Hayden

M.K.

, van Voorhis

, Nathan

, and Weinstein

R.A.

. 1998. The role of “colonization pressure” in the spread of vancomycin-resistant enterococci: an important infection control variable. Arch. Intern. Med., 158:1127–1132.

10.

Cooper

, and Lipsitch

2004. The analysis of hospital infection data using hidden Markov models. Biostatistics, 5:223–237.

11.

Dunson

D.B.

, and Chen

. 2004. Selecting factors predictive of heterogeneity in multivariate event time data. Biometrics, 60:352–358.

12.

Hancock

R.E.

1998. Resistance mechanisms in Pseudomonas aeruginosa and other nonfermentative gram-negative bacteria. Clin. Infect. Dis., 27 Suppl 1:S93–S99.

13.

Harbarth

, Cosgrove

, and Carmeli

2002. Effects of antibiotics on nosocomial epidemiology of vancomycin-resistant enterococci. Antimicrob. Agents Chemother., 46:1619–1628.

14.

Harbarth

, Harris

A.D.

, Carmeli

, and Samore

M.H.

2001. Parallel analysis of individual and aggregated data on antibiotic exposure and resistance in gram-negative bacilli. Clin. Infect. Dis., 33:1462–1468.

15.

Hooper

D.C.

2001. Emerging mechanisms of fluoroquinolone resistance. Emerg. Infect. Dis., 7:337–341.

16.

Johnson

J.K.

, Smith

, Lee

M.S.

, Venezia

R.A.

, Stine

O.C.

, Nataro

J.P.

, Hsiao

, and Harris

A.D.

. 2009. The role of patient-to-patient transmission in the acquisition of imipenem-resistant Pseudomonas aeruginosa colonization in the intensive care unit. J. Infect. Dis., 200:900–905.

17.

Koopman

J.S.

, and Lynch

J.W.

. 1999. Individual causal models and population system models in epidemiology. Am. J. Public Health, 89:1170–1174.

18.

Leibovici

, Berger

, Gruenewald

, Yahav

, Yehezkelli

, Milo

, Paul

, Samra

, and Pitlik

S.D.

. 2001. Departmental consumption of antibiotic drugs and subsequent resistance: a quantitative link. J. Antimicrob. Chemother., 48:535–540.

19.

Levin

B.R.

2001. How can we predict the ecologic impact of an antimicrobial: the opinions of a population and evolutionary biologist. Clin. Microbiol. Infect., 7 Suppl 5:24–28.

20.

Lipsitch

2001. The rise and fall of antimicrobial resistance. Trends Microbiol., 9:438–444.

21.

Lipsitch

, and Samore

M.H.

. 2002. Antimicrobial use and antimicrobial resistance: a population perspective. Emerg. Infect. Dis., 8:347–354.

22.

Livermore

D.M.

. 2002. Multiple mechanisms of antimicrobial resistance in Pseudomonas aeruginosa: our worst nightmare?. Clin. Infect. Dis., 34:634–640.

23.

Martinez

J.A.

, Ruthazer

, Hansjosten

, Barefoot

, and Snydman

D.R.

. 2003. Role of environmental contamination as a risk factor for acquisition of vancomycin-resistant enterococci in patients treated in a medical intensive care unit. Arch. Intern. Med., 163:1905–1912.

24.

Muller

, Mauny

, Talon

, Donnan

P.T.

, Harbarth

, and Bertrand

. 2006. Effect of individual- and group-level antibiotic exposure on MRSA isolation: a multilevel analysis. J. Antimicrob. Chemother., 58:878–881.

25.

O'Quigley

, and Stare

2002. Proportional hazards models with frailties and random effects. Stat. Med., 21:3219–3233.

26.

Patel

2003. Clinical impact of vancomycin-resistant enterococci. J. Antimicrob. Chemother., 51 Suppl 3:iii13–21.

27.

Paterson

D.L.

2006. Resistance in gram-negative bacteria: enterobacteriaceae. Am. J. Med., 119:S20–S28; discussion S62–S70.

28.

Patrick

D.M.

, and Hutchinson

. 2009. Antibiotic use and population ecology: how you can reduce your “resistance footprint”. CMAJ, 180:416–421.

29.

Schwartz

, and Carpenter

K.M.

. 1999. The right answer for the wrong question: consequences of type III error for public health research. Am. J. Public Health, 89:1175–1180.

30.

Vellinga

, Tansey

, Hanahoe

, Bennett

, Murphy

A.W.

, and Cormican

. 2012. Trimethoprim and ciprofloxacin resistance and prescribing in urinary tract infection associated with Escherichia coli: a multilevel model. J. Antimicrob. Chemother., 67:2523–2530.

31.

Walsh

2000. Molecular mechanisms that confer antibacterial drug resistance. Nature, 406:775–781.

32.

zur Wiesch

P.A.

, Kouyos

, Engelstadter

, Regoes

R.R.

, and Bonhoeffer

. 2011. Population biological principles of drug-resistance evolution in infectious diseases. Lancet Infect. Dis., 11:236–247.