Multivariable Analysis of the Association Between Antimicrobial Use and Antimicrobial Resistance in Escherichia coli Isolated from Apparently Healthy Pigs in Japan

Sample collection and questionnaire administration

Fecal samples were collected from apparently healthy pigs at 250 farms (one pig per farm) by the Livestock Hygiene Service Centers (LHSCs) in 46 of 47 prefectures across Japan between 2004 and 2007 under the JVARM program. This time period was selected because the sampling procedure was consistent, although this is not the most up-to-date time period. Six to eight farms were randomly selected in each prefecture and two E. coli isolates from each fecal sample per farm were collected. Data on the use of veterinary antimicrobial products within the previous 6 months as a binary response were collected from interviews with farmers using a structured questionnaire administered by veterinarians in the prefectural LHSCs. ³⁰ Antimicrobial agents used in these farms were categorized into eight classes: aminoglycosides, beta-lactams, tetracyclines, macrolides (including lincosamides), phenicols, trimethoprim–sulfamethoxazole combination drugs, fluoroquinolones, and colistin. Feed additives used were also categorized into eight classes: polypeptides, tetracyclines, macrolides, aminoglycosides, streptogramins, polyethers, synthetic antimicrobials, and others. The questionnaire also included questions relating to husbandry factors such as the age of pigs sampled, the type of farming (grow-finisher, reproduction, and farrow-to-finish), and the number of pigs reared at the farm.

Isolation and susceptibility test for E. coli

E. coli was isolated on desoxycholate–hydrogen sulfate–lactose agar (DHL agar; Eiken, Japan) at the LHSCs and sent to the National Veterinary Assay Laboratory for further investigation. ³⁵ Antimicrobial susceptibility tests were performed using agar dilution methods according to the Clinical Laboratory Standards Institute guidelines. ¹³ E. coli ATCC 25922, Staphylococcus aureus ATCC 29213, Pseudomonas aeruginosa ATCC 27853, and E. faecalis ATCC 29212 were used as reference strains for quality control. The minimum inhibitory concentration (MIC) breakpoints established by the Clinical Laboratory Standards Institute were adopted for gentamicin, kanamycin, ampicillin, cefazolin, chloramphenicol, trimethoprim, and nalidixic acid. ¹³ For the remaining antimicrobials, dihydrostreptomycin, ceftiofur, oxytetracycline, enrofloxacin, colistin, and bicozamycin, the MIC breakpoints were set as the midpoint of the two modes of susceptible and resistant MIC distributions. ⁴⁹ E. coli is intrinsically resistant to macrolide class drugs due to the limited permeability of the outer membrane to the drug ⁴⁰ and therefore resistance to macrolides was not tested, although they were used as antibiotics at farms in the study. To remove environmental effects from the same animal/farm, one of the two isolates was selected randomly and used in the statistical analysis.

Statistical analysis

In the univariate analysis, the associations between selection of antimicrobial resistance and antimicrobial use (including feed additives) and husbandry factors were tested. The association between antimicrobial use and resistance was tested using the odds ratio (OR) with a 95% confidence interval in the epitools package on statistical software, R. For the calculation of p-values, Chi-square test-based p-value was used unless at least one of the four cells had an expected frequency of less than five, where Fisher's exact test-based p-value was used. Although E. coli is intrinsically resistant to macrolide and lincosamide drugs, previous studies have reported that clindamycin promotes the establishment of intestinal colonization with antimicrobial-resistant gram-negative bacteria.^15,37 Therefore, the use of macrolides was included in the above analysis.

Associations between the age of pigs and the use of antimicrobials, between the number of pigs on a farm and the use of antimicrobials, and between the age of pigs and the antimicrobial resistance were examined using the Wilcoxon rank sum test. The associations between the type of pig farming and antimicrobial use and between the type of pig farming and resistance rates were examined using generalized linear models (GLMs) with binomial errors, choosing antimicrobial use and resistance as response variables and the type of farming as an explanatory variable.

Potential risk factors with p-values < 0.2 in these tests were used for multivariable analysis. Robust variance estimation was performed using generalized estimating equations (GEEs) ⁴⁷ with binomial errors to control the intraclass correlation at the district level caused by the effects of potential geographical homogeneity in the environment, the level of urbanization, and local trends in the use of antimicrobials on farms. ⁸ In fact, although not described in this article, there were significant differences in antimicrobial use between regions in Japan. In the model, the status of antimicrobial resistance was selected as a response variable and the use of antimicrobials and husbandry factors were selected as explanatory variables. When the use of antimicrobials had an association with husbandry factors with p-value < 0.2, their interaction terms were also included in the model. Stepwise model simplification was performed, removing the variable with the highest p-value sequentially until the p-value of all the remaining factors became < 0.05 and when further simplification showed significant change of deviance by analysis of variance (p < 0.05).

Tests for confounding were further performed for structurally unrelated associations using GLMs with binomial errors, selecting resistance as a response variable and uses as explanatory variables. Confounding can be described as the mixing together of the effects of two or more factors. Thus, when confounding is present, we might think we are measuring the association between an exposure factor and an outcome, but the observed association measure also includes the effects of one or more extraneous factors. ¹⁶ The change of logit of the use of structurally related antimicrobial, by removing the use of structurally nonrelated antimicrobials from the model, was monitored. These analyses were performed using statistical software R version 3.0.2.

Summary statistics

Table 1 shows the prevalence of resistance in 250 E. coli isolates to the antimicrobials tested. The most prevalent resistance in E. coli isolates was to oxytetracycline (62.4%, 156 isolates) and dihydrostreptomycin (44.8%, 112), followed by trimethoprim (28.8%, 72) and ampicillin (24.8%, 62). No resistance to colistin was observed.

Resistance patterns

Table 2 shows the resistance patterns of 188 E. coli isolates with resistance to at least one antimicrobial. The remaining 62 isolates (24.8% of 250 isolates) were susceptible to all of the antimicrobial agents tested. Multiantimicrobial-resistant E. coli isolates were highly prevalent; 97 isolates (51.6% of 188 resistant isolates) showed multiresistance to at least three antimicrobials. Of these, further 25 isolates (12.8% of 188 isolates) were resistant to five antimicrobials, nine isolates (4.8%) to six, two isolates (1.1%) to seven, and two isolates (1.1%) to eight antimicrobials.

Patterns of antimicrobial use

Table 3 shows the patterns of antimicrobial use at the 250 pig farms studied. No antimicrobials were used at 99 farms (39.6%) and at least one antimicrobial agent was used in the remaining farms. Among therapeutic uses of antimicrobials, the most commonly used drugs were tetracyclines (18.0%, 45 farms) and macrolides (15.2%, 38 farms). Fluoroquinolones were used at only five farms (2.0%). As for the use of feed additives, 28.4% (71 farms) used polypeptide class feed additives, 21.6% used synthetic feed additives, and 20.8% used feed additives other than polypeptides, macrolides, aminoglycosides, and synthetic antimicrobials. Tetracyclines, streptogramins, and polyethers were not used as feed additives at any of the farms in the study.

Risk factor analysis for antimicrobial resistance

Table 4 shows the results of univariate analysis for the associations between antimicrobial use and antimicrobial resistance whose p-values were < 0.2. Strong associations between the same class of antimicrobial use and antimicrobial resistance (the associations without asterisks in Table 4) were commonly observed. Associations between antimicrobial use and structurally unrelated antimicrobial resistance (p < 0.05) were also commonly observed: the use of beta-lactams and resistance to dihydrostreptomycin; use of tetracyclines and resistance to chloramphenicol; use of macrolides and resistance to ampicillin; use of fluoroquinolones and resistance to dihydrostreptomycin; and use of colistin and resistance to kanamycin. There were three negative associations between use and resistance to antimicrobials (OR < 1), indicating that the use of antimicrobials prevents selection of resistant E. coli: between use of feed additives other than polypeptides, macrolides, aminoglycosides, and synthetic antimicrobials and resistance to dihydrostreptomycin (OR = 0.5, p = 0.02), between use of synthetic feed additives and resistance to ampicillin (OR = 0.5, p = 0.055), and between use of colistin and resistance to trimethoprim (OR = 0, p = 0.19).

To investigate confounding for the above relationships between antimicrobial use and structurally unrelated antimicrobial resistance, the associations between the use of the antimicrobial and use of structurally related antimicrobials with resistance were examined. There were significant associations between the use of aminoglycosides and beta-lactams (OR = 6.0, p < 0.01), aminoglycosides and macrolides (OR = 4.4, p < 0.01), beta-lactams and macrolides (OR = 5.3, p < 0.01), tetracyclines and macrolides (OR = 3.9, p <0.01), and tetracyclines and phenicols (OR = 5.0, p = 0.02). None of the five farms using fluoroquinolones and six farms using colistin used aminoglycosides.

In the univariate analysis of the effects of age, pigs carrying ampicillin (median 4 months old) and oxytetracycline (4.6 months old)-resistant E. coli were significantly younger than pigs carrying susceptible E. coli (5 months, p = 0.02 and 5 months, p = 0.04, respectively, Table 5). In terms of antimicrobial usage, tetracyclines, fluoroquinolones, and polypeptide class feed additives were used in significantly younger pigs (medians 4, 1.5, and 4 months) compared with those where they were not used (5 months, p = 0.01; 5 months, p = 0.01; and 5 months, p = 0.03, respectively, Table 6).

In terms of the association between the farm size and use of antimicrobials, beta-lactams (p < 0.001), tetracyclines (p = 0.04), and phenicols (p < 0.001) were used for treatment in significantly larger scale farms than those where they were not used (Table 6).

The resistance rates were significantly different between farm types only for enrofloxacin (p < 0.01, Table 7), and the resistance rate in reproduction farms (8.3%) was significantly higher than in farrow-to-finish operation farms (0%, p = 0.003, Fisher's exact test). The proportions were not significantly different between reproduction farms and grow-finisher farms (0%, p = 0.06, Fisher's exact test). Differences in antimicrobial use between different farming systems were observed for aminoglycosides (p = 0.01) and polypeptide class feed additives (p = 0.04) (Table 8); both agents were most commonly used at reproduction farms (18.8% and 40.0%), followed by farrow-to-finish operations (6.5% and 28.8%; p-values in GLM: 0.02 and 0.16, with reference to the reproduction farms) and fattening farms (2.0% and 16.3%; p-values in GLM: 0.03 and 0.01, with reference to the reproduction farms).

In the multivariable analysis, 14 associations remained in the final GEE models (Table 9). Six of them were associated with exposure to drugs in the same antimicrobial class: the use of aminoglycosides in reproduction farms and resistance to kanamycin; use of beta-lactams and resistance to ampicillin (p-value is slightly above 0.05, but retained as structurally related); use of tetracyclines in the large scale farms (slope of logit = 0.001, p = 0.005) and resistance to oxytetracycline; use of phenicols and resistance to chloramphenicol; and use of fluoroquinolones and resistance to nalidixic acid and enrofloxacin. In contrast, seven types of resistances were associated with exposure to structurally unrelated classes of antimicrobials; the use of beta-lactams was associated with resistance to dihydrostreptomycin; use of colistin was associated with resistance to kanamycin; use of macrolides, especially at a younger age (slope of logit = −0.7), was associated with resistance to ampicillin and oxytetracycline (p = 0.01 and 0.02); and use of tetracyclines was associated with resistance to chloramphenicol. One negative association between the use of feed additives other than polypeptides, macrolides, aminoglycosides, and synthetic antimicrobials and resistance to dihydrostreptomycin remained in the multivariable model (slope of logit = −0.8, p = 0.02). This negative association was observed only in farrow-to-finisher farms (OR = 0.4 [0.2–0.9], p = 0.03). ORs were 0.6 (0.1–2.8, p = 0.7) in grow-finisher farms and 0.7 (0.2–2.5, p = 0.5) in reproduction farms. When pigs in farrow-to-finisher farms were categorized into sows and weaners to finishers, none of the sows were fed with this type of feed additive, while 22.3% (31/139) of weaners to finishers were (p = 0.07, Fisher's exact test). On the contrary, the resistance rate to dihydrostreptomycin was significantly higher in sows (71.4%, 10/14) than in weaners to finishers (40.3%, 56/139, x ²  = 3.8, df = 1, p = 0.05).

A GLM with resistance to dihydrostreptomycin as a response variable and uses of aminoglycosides and beta-lactams as explanatory variables was modeled to test for confounding. Removal of the use of beta-lactams from the model changed the logit of the use of aminoglycosides by 27.0% (from 0.708 to 0.899), suggesting moderate confounding. Thus, the association between the use of beta-lactams and resistance to structurally unrelated dihydrostreptomycin shown in the multivariable model result might be caused by simultaneous use of beta-lactams and aminoglycosides at farms. Similarly, a GLM with resistance to ampicillin as a response variable and use of beta-lactams and macrolides as explanatory variables was modeled, and removal of use of macrolides from the model changed the logit of the use of beta-lactams by 27.8% (from 0.713 to 0.911). This suggested that the statistical association between the use of macrolides and resistance to structurally unrelated ampicillin might be due to confounding. Removal of the use of macrolides and interaction term between the use of macrolides and age of pigs from the GLM with resistance to oxytetracycline as a response variable changed the logit of use of tetracyclines by 15.1% (1.049–1.207), suggesting the association between the use of macrolides and resistance to structurally unrelated oxytetracycline might be caused by a weaker confounding. In the GLMs with resistance to chloramphenicol as a response variable and use of phenicol and tetracycline as explanatory variables, and resistance to kanamycin as a response variable and use of aminoglycosides and colistin as explanatory variables, the changes of logits of the use of homologous antimicrobials by the removal of use of structurally unrelated antimicrobials were even smaller (14.3% change, 1.618–1.845, and 11.2% change, 1.183–1.050, respectively).

Direct selection and cross selection

Significant associations between antimicrobial use and selection of resistance to the same class of drugs were observed between the use of aminoglycosides and kanamycin resistance, use of beta-lactams and ampicillin resistance, use of tetracyclines and oxytetracycline resistance, use of phenicols and chloramphenicol resistance, and use of fluoroquinolones and nalidixic acid and enrofloxacin resistance.

In the present study, three aminoglycoside class drugs were investigated for resistance: dihydrostreptomycin, gentamicin, and kanamycin. Although the resistance rate for dihydrostreptomycin was high (44.8%), the association between the use of aminoglycosides and resistance to dihydrostreptomycin was not significant. This was probably due to the large number (88.4%) of dihydrostreptomycin-resistant isolates from nonaminoglycoside-using farms. The resistance rate for gentamicin was low (2.0%), and gentamicin resistance was not associated with the use of aminoglycosides. In contrast, kanamycin (14.4%) resistance was significantly associated with the use of aminoglycosides in reproduction farms, which matches with the higher proportion of use in reproduction farms shown in Table 8. In E. coli, the ant(2")-I, aadA,B, aphA, aac(6′)-Ib, and aac(3)-III genes on either transposons or integrons are the most important mechanisms responsible for kanamycin resistance.^38,42,43

In this study, resistance to ampicillin was associated with the use of beta-lactams. Beta-lactams, including amoxicillin, ampicillin, and benzylpenicillin, have been used for swine production in Japan.^6,23,30 The major mechanism of resistance to beta-lactam in gram-negative bacteria is the production of beta-lactamases. Most beta-lactamases are encoded on the plasmid-mediated gene. ³² Beta-lactam administration to pigs has led to the selection of E. coli harboring beta-lactamase genes in their digestive tracts. ⁹

Tetracyclines are a group of very broad-spectrum antibiotics and are one of the most common antimicrobial classes used in Japanese pig production.^6,23,30 In our study, tetracyclines were most commonly used in therapy (18.0% of the farms investigated) and the resistance rate for oxytetracycline (62.4%) was the highest among the antimicrobials investigated. A number of reports, including Harada et al. ²⁵ in Japan, have shown the statistical association between the use of tetracyclines and oxytetracycline resistance. The mechanism of oxytetracycline resistance is known to be due to the acquisition of resistance genes (tetA, tetB, etc.) on the R plasmid.^10,11,21,31

Chloramphenicol resistance was associated with the use of phenicols. The use of chloramphenicol in production animals was banned in 1998 in Japan. ²⁴ However, 20.3% of isolates collected from 2004 to 2007 in our study were resistant to chloramphenicol. The maintenance of chloramphenicol resistance in E. coli is suggested to be due to the routine use of phenicols such as thiamphenicol and florfenicol.^23,24 Resistance to chloramphenicol is mainly caused by chloramphenicol acetyltransferase (CAT) or chloramphenicol efflux pump (CmlA) and cat1 and cmlA genes, which are located on transposons or integrons.^24,33

Fluoroquinolones are a new class of synthetic antimicrobial agents, which have broad activity against both gram-positive and gram-negative bacteria. Quinolones inhibit bacterial DNA replication by inhibition of DNA gyrase and topoisomerase IV enzymes. The main quinolone resistance mechanisms are chromosomal mutations affecting the quinolone resistance-determining regions of those enzymes, although there are several other resistance mechanisms such as decreased intracellular accumulation due to efflux pumps or decreased membrane uptake and plasmid acquisition.^20,44 The acquisition of endogenous resistance genes on mobile genetic elements such as plasmids, transposons, or integrons is frequently responsible for resistance associated with the use of structurally related antimicrobials. It is likely that dissemination of these genes in the presence of antimicrobial selective pressure occurs between commensal E. coli and a wide range of microorganisms in the animal's gut. ⁴⁵ We found an association between the use of fluoroquinolones and resistance to nalidixic acid and enrofloxacin. Nalidixic acid is not used for pigs in Japan ²⁸ ; however, nalidixic acid and enrofloxacin are structurally related, and the resistance to these antimicrobials might be induced by the use of homologous antimicrobials.

Coselection of resistance

We found statistical associations between dihydrostreptomycin resistance and the use of beta-lactams. Aminoglycosides used in the pig farms investigated in this study were either kanamycin (n = 15) or streptomycin (n = 5). Four of the farms, which used streptomycin, used a streptomycin–penicillin combination drug, and only one farm used both noncombined streptomycin and penicillin (data not shown), which might have caused bias: penicillin (beta-lactam) might not select dihydrostreptomycin resistance, but it is possible that streptomycin might. A significant association between the use of aminoglycosides and beta-lactam in our results supports this as a confounding factor.

The association between the use of tetracyclines and chloramphenicol resistance was indicated by the multivariable analysis. This association was also reported in C. coli. ³⁶ The mechanisms of coresistance to chloramphenicol and tetracycline in E. coli are still unknown in pigs, although coselection for resistance to chloramphenicol induced by the use of tetracyclines was reported previously by Harada. ²² A study in the United States of America on drug resistance in E. coli from humans and food animals reported that over 90% of chloramphenicol-resistant E. coli isolates were also resistant to tetracycline and proposed that this could be due to coselection of mobile resistance elements. ⁴⁶ In fact, cocarriage of tetracycline resistance genes tet(A) and tet(B), and floR, a phenicol resistance gene on E. coli plasmids of swine origin, which shows cross-resistance to chloramphenicol, has been reported. ²⁷ Therefore, the selective pressure imposed by the use of tetracyclines may contribute to the selection of chloramphenicol-resistant E. coli. However, we found a statistical association between the use of tetracyclines and phenicols and thus the association between use of tetracyclines and resistance to chloramphenicol might be at least partially due to this confounding.

As discussed above, plausible coresistance was the use of tetracyclines and chloramphenicol resistance, although there was a moderate statistical confounding by the use of these antimicrobials at the same farms. The association between the use of beta-lactams and resistance to dihydrostreptomycin was concluded to be confounded by the combined use of drugs (streptomycin and penicillin).

Association between use and resistance to antimicrobials without direct and cross selections

The test for confounding in GLMs and ORs showed that the statistical association between the use of macrolides and resistance to structurally unrelated ampicillin was confounded by the simultaneous use of macrolides and beta-lactams at the same farms. A 30% change of logit of an explanatory variable in GLM by removing the other explanatory variable from the model is interpreted as a moderate confounding, ¹⁶ which distorted the relationship between the explanatory variable monitored and the response variable. Although the degree was weaker, similar confounding was observed also in the association between the use of macrolides and resistance to oxytetracycline. However, other reports of on-farm antimicrobial use and resistance also found that macrolide use was a risk factor for various E. coli resistances in pigs.^3,39 E. coli isolated from patients receiving erythromycin carried transferable plasmid encoding resistance to ampicillin, erythromycin, gentamicin, and streptomycin. ⁴ Another study reported that E. coli isolated from a human clinical specimen carried a transferable plasmid encoding resistance to erythromycin, chloramphenicol, and tetracycline. ⁵ From these findings, Rosengren et al. ³⁹ indicated that administration of erythromycin at a higher concentration than the MIC of E. coli may cause coselection of highly resistant E. coli. It has been reported that clindamycin promotes the establishment of intestinal colonization with antimicrobial-resistant gram-negative bacteria,^17,37 indicating that macrolides may select resistant E. coli indirectly by disrupting the intestinal microflora. Although confounding was suggested in the associations between the use of macrolides and resistance to ampicillin and oxytetracycline, a hypothesis that these associations were caused by unknown biological mechanisms cannot be rejected. Further bacteriological investigation is needed to determine the mechanism of resistance selection by macrolides in pigs.

The weak level of confounding (11.2%) in our test is not sufficient to conclude that the significant association between the use of colistin and resistance to kanamycin was due to bias. This also provided another hypothesis that the use of colistin selects resistant E. coli to kanamycin by an unknown mechanism.

Husbandry factors associated resistance prevalence with antimicrobial use

According to the results of the univariate analyses, the use of tetracyclines, fluoroquinolones, and polypeptide class feed additives was significantly associated with younger pigs. However, these were not associated with antimicrobial resistance in multivariable models and instead the use of macrolides in younger pigs was associated with resistance to ampicillin and oxytetracycline. In fact, the status of use of a drug was investigated by asking whether the drug was used in pigs within a 6-month period at the time of the survey. Considering the short production cycle (finishers will be sold at 6 months of age to slaughterhouses), the effect of timing of drug administration could not be measured in our study. Moreover, the univariate analysis on age dealt with sows, growers, and finishers altogether. The effect of macrolides on selecting ampicillin and oxytetracycline-resistant E. coli is thus still unknown, but should be considered in future experimental studies. The other husbandry factors remaining in the final models were the use of aminoglycosides in reproduction farms associated with kanamycin resistance and use of tetracyclines in large-scale farms associated with oxytetracycline resistance. Finding these associations showed the usefulness of including husbandry factors in multivariable analysis for a better understanding of the associations between the use and resistance of antimicrobials.

One negative association between the use of feed additives other than polypeptides, macrolides, aminoglycosides, and synthetic antimicrobials and resistance to dihydrostreptomycin remained in the multivariable model. Additional statistics found that this was due to the difference in the proportions of use of this type of feed additive and resistance rates to dihydrostreptomycin between sows and weaners to finishers in farrow-to-finisher farms.

In conclusion, consideration of multidrug-resistant isolates and confounding factors is essential when analyzing multivariate relationships of the resistance prevalence with antimicrobial use. Our results suggest that the use of antimicrobials on farms could select antimicrobial-resistant strains by direct selection, cross selection, or coselection, and even hypothetical unknown or indirect mechanisms. We hope that such hypothetical mechanisms can be used to plan future bacteriological experiments. Our study, however, used multivariable statistics, and future studies on E. coli in pigs should focus on understanding the multivariate relationships of multiple antimicrobial resistances, hopefully with more quantitative data.