Long-term use of antibiotics and risk of type 2 diabetes in women: a prospective cohort study

Abstract

Background

Accumulating evidence suggested that long-term antibiotic use may alter the gut microbiome, which has, in turn, been linked to type 2 diabetes. We undertook this study to investigate whether antibiotic use was associated with increased risk of type 2 diabetes.

Methods

This prospective cohort study included women free of diabetes, cardiovascular disease and cancer in the Nurses’ Health Study (NHS 2008–2014) and NHS II (2009–2017). We evaluated the overall duration of antibiotics use in the past 4 years and subsequent diabetes risk with Cox proportional-hazards regression adjusting for demography, family history of diabetes and lifestyle factors.

Results

Pooled analyses of NHS and NHS II (2837 cases, 703 934 person-years) revealed that a longer duration of antibiotic use in the past 4 years was associated with higher risk of diabetes [Trend-coefficient = 0.09, 95% confidence interval (CI) 0.04 to 0.13]. Participants who received antibiotics treatment for a medium duration of 15 days to 2 months [hazard ratio (HR) 1.23, 95% CI 1.10 to 1.39] or long duration of >2 months (HR 1.20, 95% CI 1.02 to 1.38) had higher risk of type 2 diabetes as compared with non-users. Subgroup analyses suggested that the associations were unlikely to be modified by age, family history of diabetes, obesity, smoking, alcohol drinking, physical activity and overall diet quality.

Conclusions

A longer duration of antibiotic use in recent years was associated with increased risk of type 2 diabetes in women. Physicians should exercise caution when prescribing antibiotics, particularly for long-term use.

Key Messages

• Accumulating evidence suggested that long-term antibiotic use may alter the gut microbiome, which has, in turn, been linked to type 2 diabetes.

• This prospective cohort study showed that a longer duration of antibiotic use in recent years was associated with increased risk of type 2 diabetes in women.

• Physicians should exercise caution when prescribing antibiotics particularly for long-term use.

Introduction

Type 2 diabetes has become a global epidemic over the past few decades. The global prevalence of diabetes has nearly doubled, rising from 4.7% in 1980 to 8.5% in 2014.¹ The aetiology of type 2 diabetes is complex, involving multiple genetic, behavioural and environmental factors.² In recent years, researchers have turned their attention to the role of human gut microbiota, which is essential for key metabolic processes such as energy harvesting, production of neurotransmitters and hormones, and nutrient storage, in the development of diabetes.^3,4 Accumulating studies support a causal role for the gut microbiota in metabolic diseases.^5,6 The latest findings from the Integrative Human Microbiome Project (iHMP) have revealed insights into pathways and responses that differ between glucose-dysregulated and healthy individuals during health and disease.⁷

Antibiotics are one of the most widely prescribed medications, with a global consumption of 54 billion standard units in 2000 and 73 billion units in 2010.⁸ Despite the irreplaceable role, antibiotics may alter the gut microbiome and host health. A previous study showed that broad-spectrum antibiotics could affect the abundances of 30% of the bacteria in the gut community, causing rapid drops in taxonomic richness, diversity and evenness.⁷ The alterations can persist for years; in some cases, the initial composition is never completely recovered.^9,10 Considering the impact on gut microbiota, a number of studies have investigated antibiotics use and the development of metabolic disorders, and suggested that antibiotic use could increase the risk of obesity^11,12 and diabetes.^13–15

Although a number of observational studies have investigated antibiotics use and type 2 diabetes risk,^13–16 the relationship remains unclear. First, the results of these studies were inconsistent, showing either a positive^13–15 or a null association.¹⁶ Second, since previous studies were carried out in the general population, who have a high rate of undiagnosed diabetes,¹⁷ they cannot rule out the possibility that the positive association shown in these studies was because the participants seeking antibiotic treatment were more likely to have been previously unrecognized with diabetes.^14,15 Finally, a large population-based, prospective cohort study is the most appropriate study design for investigating diabetes risk factors, but such a study investigating the effect of anbiotics is still lacking. In order to address these limitations, we evaluated antibiotic use and risk of type 2 diabetes among women enrolled in the Nurses’ Health Study (NHS) and Nurses’ Health Study II (NHS II).

Methods

Study population

NHS is an ongoing prospective cohort study, beginning in 1976 and enrolling 121 700 female nurses from 11 US states aged 30–55 years. NHS II, established in 1989, originally included 116 430 younger female registered nurses who were between the ages of 25 and 42 years from 14 states in the USA. The participants receive a biennial questionnaire since baseline to collect data on demographics, health-related behaviours, medical history and newly diagnosed diseases, with a follow-up rate of >90% for each questionnaire cycle. The recruitment and data collection in NHS and NHS II are detailed previously.¹⁸ NHS and NHS II were approved by the Human Research Committee at the Brigham and Women’s Hospital, Boston, MA. The study protocol was approved by the institutional review board (IRB) of the Brigham and Women's Hospital, and the IRB allowed participants’ completion of questionnaires to be considered as implied consent.

Previous studies have investigated antibiotic use and risk of colorectal adenoma,¹⁹all-cause and cause-specific mortality²⁰ using the NHS or NHS II data set. For the current analysis, we restricted the participants to those without a diagnosis of diabetes (including type 1, type 2 and gestational diabetes). We also excluded the participants who had missing data on antibiotic use (see Figure 1). The baseline was 2008 for NHS and 2009 for NHS II when data about antibiotic use in the past 4 years were collected. The end of follow-up was 2014 for NHS and 2017 for NHS II when the latest ascertainment of diabetes was carried out. Those who did not respond to any of the questionnaire cycles between 2008 and 2017 were considered as being lost to follow-up.

Figure 1.

Flowchart of participant selection from the Nurses’ Health Study and Nurses’ Health Study II.

Assessment of antibiotic use

In 2008 and 2009, NHS and NHS II participants were asked about their total amount of time using antibiotics (excluding skin creams, mouthwash or isoniazid) during the past 4 years (none, 1–14 days, 15 days–2 months, 2–4 months, 4 months–2 years, 2–3 years or >3 years). The most common reasons for using an antibiotic (respiratory infection, urinary-tract infection, acne/rosacea, chronic bronchitis, dental and other reason) were collected in both cohorts.

Ascertainment of type 2 diabetes

On the subsequent biennial questionnaires, the participants were asked whether they had ever been diagnosed with diabetes. Participants reporting physician-diagnosed type 2 diabetes were mailed a validated Supplementary questionnaire, available as Supplementary data at IJE online, regarding symptoms, diagnostic tests and hypoglycaemic therapy to confirm self-reported diagnoses. Confirmed diabetes requires at least one of the following reported on the Supplementary questionnaire, available as Supplementary data at IJE online, in accordance with the American Diabetes Association criteria: (i) one or more classic symptoms (excessive thirst, polyuria, weight loss, hunger, pruritus or coma) plus fasting plasma glucose (PG)  ≥ 126 mg/dl (7.0 mmol/L) or random PG ≥ 200 mg/dl (11.1 mmol/L); (ii) at least two elevated PG levels on different occasions (fasting PG ≥ 140 mg/dl and/or random PG ≥200 and/or 2 hr PG ≥ 200 on oral glucose tolerance testing) in the absence of symptoms; or (iii) treatment with hypoglycaemic medication (insulin or oral hypoglycaemic agent). Two separate studies have shown high validity of self-reported type 2 diabetes diagnosis (>97%) through medical-record reviews.^21,22

Assessment of covariates

We selected covariates that may confound the association between antibiotics use and type 2 diabetes risk based on a literature review. In the baseline and follow-up questionnaires, we obtained updated information on age, ethnicity, marital status, living status, family history of diabetes, body mass index (BMI), smoking status, alcohol intake, menopausal status and postmenopausal hormone use and history of hypertension and hypercholesterolemia. We calculated the 2010 Alternative Healthy Eating Index (AHEI-2010) to assess overall diet quality. Physical activity was measured by weekly expenditure of metabolic equivalents (METs), which has been validated in a previous study.²³

Statistical analysis

This study calculated person-years from the date of return of the baseline questionnaire to the date of diagnosis of diabetes, death or the end of follow-up, whichever came first. We evaluated the hazard ratios (HRs) and 95% confidence intervals (CIs) with multivariable time-dependent proportional-hazards models accounting for potential time-varying effects in covariates. We stratified the analyses jointly by age (in months) and the year that the questionnaire was returned. The participants were classified into non-use, short-term use (1–14 days), medium-term use (15 days–2 months) and long-term use (2+ months) groups, taking the non-use group as the reference. We considered the antibiotic-use durations as categorical variables in the primary analysis and tested for trend by modelling the duration as an ordinal variable (non-use: 1, 1–14 days: 2, 15 days–2 months: 3 and 2+ months: 4). To present the association in a clinically useful way, we calculated number needed to harm (NNH) with the method described by D.G. Altman.²⁴

We generated missing covariate data (BMI, physical activity, Alternate Health Eating Index, Menopausal status and postmenopausal hormone use, smoking and days with alcohol drinking per week) with multiple imputation. Five imputations were generated by the fully conditional specification method.²⁵ We evaluated potential non-linear associations between continuous covariates and diabetes risk with cubic splines.²⁶ In the multivariable-adjusted model 1, we controlled for ethnicity (White or other), marital status (married, divorced, widowed or others), living status (alone, with families or other), history of diabetes in a first-degree relative (yes or no), BMI (continuous), menopausal status and postmenopausal hormone use (premenopausal, postmenopausal (never, past or current menopausal hormone use)). To control the potential confounding in lifestyle factors, we further adjusted smoking status (never, past or current), days with alcohol drinking per week (0, 1–3 or >3 days), physical activity (continuous or MET-h/week) and overall diet quality (continuous or AHEI score) to the multivariable-adjusted model 2. We carried out regression in NHS and NHS II separately and then pooled the effects with inverse variance-weighted, random-effect meta-analyses. We tested the heterogeneity of associations between the two cohorts using the Cochrane Q statistic and the I² statistic.

To verify potential interaction effects, we undertook subgroup analysis according to age, family type 2 diabetes history, BMI, smoking, alcohol drinking, physical activity and AHEI score. We tested interactions by including interaction terms in the fully adjusted Cox-regression model. To control potential confounding by the indication of antibiotics use, we did sensitivity analyses by additionally adjusting for two surrogate indicators, including rheumatoid arthritis and depression, which have been shown to be associated with increased risk of infection.^27,28 To investigate the potential bias from healthcare utilization (i.e. the participants with better healthcare utilization are likely to have a better access to antibiotics and lower chance to be undiagnosed if they had diabetes), we adjusted physical examination in the previous 2 years (yes or no) as a surrogate indicator. As hypertension and elevated cholesterolemia may partially mediate or confound the effect of antibiotic on diabetes, we adjusted for these covariates to evaluate the controlled direct effect or adjusted effect. We also adjusted parity to investigate potential residual confounding effects. We performed the analyses using SAS software, version 9.4 (SAS Institute, Cary, NC, USA).

Results

This study included 114 210 participants, of whom 50 810 were from NHS and 63 400 were from NHS II (Figure 1). Table 1 presents the baseline participant characteristics according to the duration of antibiotic use in the past 4 years. The participants with longer antibiotic use were more likely to have a higher BMI and higher rate of current postmenopausal hormone use.

Table 1.

Age-adjusted baseline characteristics according to the antibiotic-use duration in the past 4 years

	Duration of antibiotic use in the past 4 years
	None	1–14 days	15 days–2 months	2+ months
NHS, 2008 (n = 50 810)
Number of participants	13 110	21 237	12 481	3982
Mean (SD) age, years	73.7 (6.9)	72.9 (6.7)	72.33 (6.7)	73.1 (7.0)
White, %	97	98	98	98
Married, %	64	65	65	64
Live alone, %	28	27	26	26
Family history of diabetes, %	19	20	20	20
Mean(SD) BMI, kg/m2	25.5 (4.7)	26.0 (4.9)	26.3 (5.1)	26.2 (5.3)
Current smoker, %	6	6	5	5
Median (IQR) days of alcohol drink per week	2 (1)	2 (1)	2 (1)	2 (1)
Mean (SD) physical activity time, MET-hours/week	29.5 (95.8)	27.9 (91.9)	25.2 (86.9)	29.9 (114.5)
Mean (SD) Alternate Health Eating Index	61.3 (12.3)	63.1 (12.0)	62.8 (11.8)	62.3 (12.1)
Premenopausal, %	0	0	0	0
Current postmenopausal hormone use, %	10	12	15	17
NHS II, 2009 (n = 63 400)
Number of participants	14 596	26 703	17 296	4805
Mean(SD) age, years	54.7 (4.6)	54.7 (4.6)	54.7 (4.7)	54.7 (4.6)
White, %	96	96	97	97
Married, %	78	78	77	75
Live alone, %	11	11	11	11
Family history of diabetes, %	34	36	37	39
Mean (SD) BMI, kg/m2	26.2 (5.4)	27.1 (5.9)	27.9 (6.4)	28.5 (7.0)
Current smoker, %	5	6	7	7
Median (IQR) days of alcohol drink per week	2 (1)	2 (1)	2 (1)	2 (1)
Mean (SD) physical activity time, MET-hours/week	34.2 (92.2)	32.8 (93.4)	31.3 (96.6)	31.7 (102.6)
Mean (SD) Alternate Health Eating Index	62.1 (13.1)	61.0 (12.9)	60.1 (12.8)	59.0 (12.7)
Premenopausal, %	28	25	22	20
Current postmenopausal hormone use, %	19	23	27	32

SD, standard deviation; BMI, body mass index; IQR, interquartile range; MET, metabolic equivalent.

We documented 1031 cases of type 2 diabetes during 260 765 person-years of follow-up in NHS and 1806 cases during 443 169 person-years in NHS II. The annual incidence of type 2 diabetes was 4.07/1000 person-years in NHS and 3.95/1000 person-years in NHS II. In NHS, the adjusted HRs were 1.14 (95% CI 0.96 to 1.37) for the medium-term use group (15 days–2 months) and 1.17 (95% CI 0.91 to 1.50) for the long-term use group (2+ months) as compared with non-users (Table 2). In NHS II, we observed a dose–response relationship between antibiotic-use duration and type 2 diabetes risk (Trend-coefficient = 0.10, 95% CI 0.05 to 0.15). The HR in the fully adjusted multivariable models was 1.30 (95% CI 1.13 to 1.49) for the medium-term use group and 1.20 (95% CI 1.00 to 1.45) for the long-term antibiotic-use group, as compared with the non-use group. In the meta-analysis of the two cohorts, the duration of antibiotic use was positively associated with diabetes risk (Trend-coefficient = 0.09, 95% CI 0.04 to 0.13) with no major heterogeneity (I² < 25%). Overall, the medium-term use group (HR 1.23, 95% CI 1.10 to 1.39) and long-term use group (HR 1.20, 95% CI 1.02 to 1.3) had higher risk of type 2 diabetes as compared with non-users. In the analyses limiting participants in those with complete covariates data, we observed no substantial changes in primary results.

Table 2.

Antibiotic use during the past 4 years and risk of type 2 diabetes in the Nurses’ Health Study and Nurses’ Health Study II

	Duration of antibiotic use in the past 4 years				Trend-coefficient
	None	1–14 days	15 days–2 months	2+ months
NHS
No. of cases	220	429	290	92
No. of participants	13 110	21 237	12 481	3982
Person-years of follow-up	67 518	109 578	64 041	19 628
Incidence, n/1000 person-years	3.26	3.92	4.53	4.69
Hazard ratio [95% confidence interval]
Age-stratified model	1.00 (referent)	1.17 [0.99, 1.37]	1.31 [1.10, 1.57]	1.35 [1.06, 1.73]	0.11 [0.04, 0.18]
Multivariable-adjusted model 1a	1.00 (referent)	1.07 [0.91, 1.26]	1.15 [0.97, 1.38]	1.18 [0.92, 1.51]	0.06 [−0.01, 0.13]
Multivariable-adjusted model 2b	1.00 (referent)	1.07 [0.90, 1.26]	1.14 [0.96, 1.37]	1.17 [0.91, 1.50]	0.06 [−0.01, 0.13]
NHS II
No. of cases	307	694	627	178
No. of participants	14 596	26 703	17 296	4805
Person-years of follow-up	103 720	186 788	119 865	32 796
Incidence, n/1000 person-years	2.96	3.72	5.23	5.43
Hazard ratio [95% confidence interval]
Age-stratified model	1.00 (referent)	1.25 [1.01, 1.47]	1.68 [1.28, 2.20]	1.71 [1.19, 2.44]	0.24 [0.19, 0.29]
Multivariable-adjusted model 1a	1.00 (referent)	1.08 [0.94, 1.23]	1.35 [1.18, 1.55]	1.28 [1.06, 1.55]	0.12 [0.07, 0.18]
Multivariable-adjusted model 2b	1.00 (referent)	1.06 [0.93, 1.21]	1.30 [1.13, 1.49]	1.20 [1.00, 1.45]	0.10 [0.05, 0.15]
Pooled
No. of cases	527	1123	917	270
No. of participants	27 706	47 940	29 777	8787
Person-years of follow-up	171 238	296 366	183 906	52 424
Incidence, n/1000 person-years	3.08	3.79	4.99	5.15
Hazard ratio [95% confidence interval]
Age-stratified model	1.00 (referent)	1.19 [1.05, 1.35]	1.45 [1.15, 1.84]	1.46 [1.18, 1.82]	0.18 [0.06, 0.30]
Multivariable-adjusted model 1a	1.00 (referent)	1.07 [0.97, 1.19]	1.26 [1.08, 1.47]	1.26 [1.08, 1.47]	0.10 [0.04, 0.16]
Multivariable-adjusted model 2b	1.00 (referent)	1.06 [0.96, 1.18]	1.23 [1.10, 1.39]	1.20 [1.02, 1.38]	0.09 [0.04, 0.13]

^aMultivariable-adjusted model 1: additionally adjusted for ethnicity (White or other), marital status (married, divorced, widowed or others), living status (alone, with families or others), family history of diabetes (yes or no), body mass index (continuous), menopausal status and postmenopausal hormone use (premenopausal, postmenopausal (never, past or current menopausal hormone use)).

^bMultivariable-adjusted model 2: additionally adjusted for smoking status (never, past, current or missing), days with alcohol drinking per week (0, 1–3 or >3 days), physical activity (metabolic equivalents-h/week or continuous) and multiple dietary factors (Alternative Healthy Eating Index score or continuous).

To present the associations in a clinically useful way, we calculated NNHs based on the estimated HRs and incidence rates of type 2 diabetes in the control group (Supplementary Table 1, available as Supplementary data at IJE online). Over 5 years, 102.6 (95% CI 79.3 to 207.4) persons were needed to use antibiotics for 15 days–2 months and 114.7 (95% CI 80.8 to 551.8) persons were needed to use antibiotics for 2+ months, to cause one case of type 2 diabetes.

To investigate the potential influence of the underlying infection for using antibiotics, we evaluated the HRs according to the primary reasons for using antibiotics (Supplementary Table 2, available as Supplementary data at IJE online). Taking respiratory infections as the reference, urinary-tract infection (HR 1.01, 95% CI 0.70 to 1.44), dental infection (HR 1.06, 95% CI 0.92 to 1.23) and other infections (HR 1.00, 95% CI 0.90 to 1.11) did not associate with different risk of type 2 diabetes risk. To confirm the independent effect of antibiotic use on type 2 diabetes risk, we additionally adjusted the primary reasons for using antibiotics in the regression model (Supplementary Table 3, available as Supplementary data at IJE online). Medium antibiotic use (HR 1.20, 95% CI 0.99 to 1.46) and long-term antibiotic use (HR 1.22, 95% CI 1.06 to 1.40) were associated with higher risk of diabetes as compared with short-term antibiotic users.

Table 3 shows the stratified analyses of antibiotic use in the past 4 years and type 2 diabetes risk. The associations between antibiotic use and risk of diabetes were unlikely modified by age, family diabetes history, BMI, smoking, alcohol drinking, physical activity and AHEI score.

Table 3.

Subgroup analyses of antibiotic use in the past 4 years and subsequent type 2 diabetes risk

Subgroup	Event/participant	HR [95% CI] by the duration of antibiotic use
		None	1–14 days	15 days–2 months	2+ months	P-interaction
Age
<70 years	2234/82 736	1.00 (referent)	1.04 [0.92, 1.17]	1.24 [1.08, 1.38]	1.08 [0.91, 1.28]	0.57
≥70 years	603/31 474	1.00 (referent)	1.06 [0.86, 1.31]	1.11 [0.88, 1.40]	1.18 [0.86, 1.62]
Family history of diabetes
Yes	1496/81 310	1.00 (referent)	1.06 [0.96, 1.18]	1.25 [1.12, 1.39]	1.20 [1.04, 1.40]	0.25
No	1341/32 900	1.00 (referent)	1.09 [0.95, 1.26]	1.19 [1.02, 1.38]	1.26 [1.03, 1.55]
BMI
<30 kg/m2	1062/87 712	1.00 (referent)	1.08 [0.97, 1.20]	1.28 [1.14, 1.42]	1.22 [1.05, 1.42]	0.75
≥30 kg/m2	1775/26 498	1.00 (referent)	1.23 [0.87, 1.74]	1.38 [0.96, 2.00]	1.45 [0.85, 2.45]
Non-smoker
Yes	1656/65 306	1.00 (referent)	1.07 [0.96, 1.18]	1.25 [1.12, 1.39]	1.20 [1.04, 1.40]	0.50
No	1181/48 904	1.00 (referent)	1.13 [0.83, 1.56]	1.37 [0.99, 1.90]	1.01 [0.64, 1.59]
Alcohol drinking days/week
0 day	1492/43 480	1.00 (referent)	1.07 [0.96, 1.18]	1.25 [1.12, 1.40]	1.21 [1.04, 1.40]	0.27
1 day or over	1345/70 730	1.00 (referent)	1.04 [0.75, 1.45]	1.26 [0.90, 1.77]	0.86 [0.54, 1.37]
Physical activity
<9 MET-hours/week	2363/81 884	1.00 (referent)	1.06 [0.96, 1.18]	1.25 [1.12, 1.40]	1.21 [1.04, 1.40]	0.33
≥9 MET-hours/week	474/32 326	1.00 (referent)	1.39 [0.99, 1.94]	1.43 [1.01, 2.03]	1.22 [0.75, 1.97]
Alternate Health Eating Index
<60	1511/51 107	1.00 (referent)	1.07 [0.96, 1.18]	1.25 [1.12, 1.39]	1.21 [1.04, 1.40]	0.27
≥60	1326/63 103	1.00 (referent)	0.98 [0.70, 1.35]	0.97 [0.69, 1.36]	1.05 [0.67, 1.66]

HR, hazard ratio; CI, confidence interval; SD, standard deviation; BMI, body mass index; MET, metabolic equivalent.

The estimates were based on fully adjusted models (see the covariates in the footnote to Table 2).

In the sensitivity analyses by additional adjustment for rheumatoid arthritis and depression, we observed no substantial changes in primary results (Supplementary Table 4, available as Supplementary data at IJE online). The results were stable after additional adjustment for physical examination in the past 2 years and number of parity.

Discussion

In this prospective cohort of women, recent medium- and long-term antibiotic use was associated with an increased risk of type 2 diabetes, independently of traditional diabetes risk factors including demographic factors, family history of diabetes, obesity and lifestyle factors. The association between antibiotic use and type 2 diabetes risk was unlikely to be modified by age, family history of diabetes, obesity and lifestyle factors.

We did a systematic literature search and identified four observational studies that evaluated antibiotics and risk of diabetes (see the summary of these studies in Supplementary Table 5, available as Supplementary data at IJE online), of which three studies found a positive association.^13–15 A nested case–control study of 208 002 diabetic cases and 815 576 matched controls showed an association between past antibiotic exposure and diabetes risk.¹⁴ In a nationwide register-based study of 170 504 cases and 1 364 007 matched controls from Denmark, the authors found an increased risk of type 2 diabetes with increasing intake of antibiotics. Increased use of antibiotics in type 2 diabetes patients was found ≤15 years before diagnosis of diabetes.¹⁵ Despite the large sample size, the results of these two studies might be confounded by lifestyle factors and family history of diabetes. A retrospective cohort study of 14 361 veterans showed that one or more prescriptions of antibiotics was associated with increased risk of diabetes (HR = 1.13) when compared with the non-prescription group. The individual class of antibiotics with significantly increased risk included cephalosporin, macrolide and penicillin.¹³ In a relatively small nested case–control study from the Alberta's Tomorrow Project (1676 cases and 13 401 controls), the authors found no significant association after adjustment for clinical and lifestyle factors.¹⁶ Important confounding factors in the Alberta's Tomorrow Project study have been controlled in our study and we still observed a positive association. Possible explanations for the inconsistency included different study design, participant characteristics and study sample size.

One concern for the positive association is that the participants with undiagnosed type 2 diabetes have increased demand for antibiotics because they are more prone to developing infections.²⁹ Lack of universal blood-glucose screening may also influence the confirmation of the study outcome. However, the influence would be minor because: (i) the study population are nurses who have rich medical knowledge and high awareness of diabetes symptoms for themselves; (ii) US nurses have a high frequency of blood-glucose tests, since most of them have a primary-care provider; and (iii) we adjusted physical examination in the previous 2 years and found no major change in the primary results.

The mechanism between antibiotic use and type 2 diabetes risk is still unclear. Accumulating evidence suggests that long-term use of antibiotics can have a profound impact on gut microbiota^1,9,10,30 and gut microbiota play a causal role in metabolic-disease development;^4–6,31 thus, it is reasonable to link antibiotics use with type 2 diabetes. Animal models have shown that antibiotics may alter insulin sensitivity and glucose tolerance, as well as enhance lipid deposition by altering the gut-microbiota composition.^32–34 In humans, a randomized trial has shown that oral vancomycin significantly decreases intestinal microbiota diversity, bile-acid dihydroxylation and peripheral insulin sensitivity in subjects with metabolic syndrome.³⁵ Furthermore, many antibiotics, such as fluoroquinolones and pentamidine, may increase risk of dysglycemia independently of the gut microbiota.^36–38

One of our strengths is that this study was based on two well-established prospective cohorts with reasonably large sample sizes and high follow-up rates. The participants were nurses who were able to provide complete and accurate health information. Additionally, we comprehensively controlled for traditional diabetes risk factors, which minimized the potential confounding effect. Robust sensitivity analyses and the clear dose–response relationship also increased our confidence in the results.

This study has limitations. First, our results may be influenced by immortal time bias because, at study entry, eligible participants should be free of diabetes during the past 4 years. However, the potential influence would be minor because (i) the rates of type 2 diabetes in the past 4 years also increased with total duration of antibiotic use (NHS: None 2.33%, 1–14 days 2.93%, 15 days–2 months 3.28%, 2+ months 3.79%; NHS II: None 1.62%, 1–14 days 1.98%, 15 days–2 months 2.86%, 2+ months 3.82%), (ii) the start of follow-up was after the end of the immortal time period, which is a proposed approach to prevent immortal time bias.³⁹ Second, as an observational study, this study could not confirm a causal effect of antibiotic use on type 2 diabetes. Third, the association between antibiotic use and diabetes risk may be confounded by the underlying infection. However, our analyses suggested that the risk of type 2 diabetes was similar among different types of infections and the risk still increased with the duration of antibiotic use after additionally adjusting for the primary reason for using antibiotics. Forth, long-term antibiotic exposures may differ qualitatively from short-term exposures. Though we have adjusted for the reasons given and compared the risk of type 2 diabetes among different types of infection, we cannot rule out this concern because (i) the exact reason for using antibiotics was not collected and (ii) we cannot differentiate whether the participant used antibiotic once for >2 months or used antibiotics multiple times for difference reasons with an accumulating duration of >2 months. Fifth, our conclusion may be influenced by recall bias. However, the influence would be minor, as the participants are health professionals who are well educated and familiar with personal antibiotic use. In addition, report of antibiotics use is expected to be non-differential to diabetes diagnosis, because the participants were unlikely to know that there was association between antibiotics use and risk of diabetes [the first report of this association was published in 2015, whereas the time for collecting data about antibiotics use was much earlier (2008–2009)]. Sixth, we were unable to evaluate the effect of specific classes of antibiotics, as such data were not collected. However, we did not find major differences according to the primary reasons for antibiotic use, which are closely associated with the type of prescription. Last, all included participants were women, so the findings may be not applicable to men.

Conclusions

Overall, this prospective cohort study indicated that long-term antibiotic use in recent years was associated with increased risk of type 2 diabetes in women, independently of traditional diabetes risk factors. Long-term use of antibiotics has been linked with a series of health problems like obesity,^11,12 colorectal adenoma¹⁹ and all-cause and cardiovascular mortality.²⁰ Our results add further evidence that physicians should exercise caution when prescribing antibiotics, particularly for long-term treatment. Further prospective studies are still required to confirm our findings.

Supplementary data

Supplementary data are available at IJE online.

Author contributions

J.Y., Y.P. and J.T. had the initial concept and designed this study. J.Y. and Y.P. obtained the funding. J.Y. performed data analysis. J.Y. and Y.H. drafted the manuscript. All authors commented on and revised the manuscript. Prof. Jinling Tang (jltang@cuhk.edu.hk) and Prof. Yihang Pan (panyih@mail.sysu.edu.cn) contribute equally to this study and should be considered as co-corresponding authors.

Funding

This work was supported by the start-up grant for the 100 Top Talents Program, The Seventh Affiliated Hospital, Sun Yat-sen University (392012), the National Key Research and Development Program (2018YFA0902801), and Vice-Chancellor Fellowship from the University of Bristol and the UK Medical Research Council Integrative Epidemiology Unit (MC_UU_00011/1 and MC_UU_00011/4). NHS and NHS II were supported by the National Cancer Institute (UM1 CA176726 and UM1 CA186107) and the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) (DK58845 and DK112940). The funders had no role in the design, conduct or data interpretation of the study.