Results of the search

The combined results of all literature searches are described in the study flow diagram (Figure 1).

Figure 1

---

---

Figure 1 Study flow diagram.

Included studies

The Characteristics of included studies table lists 221 studies, of which 211 used the following designs to evaluate the intended effect of interventions: 138 ITS studies, 58 RCTs (14 cluster RCTs), 6 CBAs, and 8 NRTs. The remaining 11 studies were designed to identify unintended consequences of interventions and used the following designs: 8 cohort (Connor 2007; Duvoisin 2014; Friedberg 2009; Kanwar 2007; LaRosa 2007; Linkin 2007; Welker 2008; Winters 2010), 1 case control (Calfee 2003), and 1 qualitative (semi‐structured interviews) (Baysari 2013) and 1 ITS (Bell 2014).

Excluded studies

We excluded 32 unique studies from the review because they did not contain relevant or interpretable data (Selection of studies). For details of each study, see Characteristics of excluded studies.

Allocation

Most of the RCTs had high risk of selection bias because of problems with concealment of allocation (Figure 2). The RCTs with low risk of selection bias were either cluster RCTs or interventions with circumstantial reminders, for which concealment of allocation is relatively straightforward.

Blinding

Most of the RCTs also had high risk of performance and detection bias because RCTs in single hospitals were often single‐blind and it was difficult to conceal the allocation of participants in these trials (Figure 2).

Incomplete outcome data

The RCTs used data collected specifically for the trial, and all provided convincing evidence about lack of attrition bias. Most of the ITS studies used data from routine systems for prescribing (pharmacy) and microbial (microbiology) outcomes; we assessed these sources as having low risk of attrition bias (Figure 2). Examples of high risk of attrition bias in routine data are changes in the number of participants who did not have serum creatinine measure preoperatively during the study period, which may have biased ascertainment of postoperative kidney injury (Bell 2014), and use of surveillance data about surgical‐site infection that did not include information about infections arising after discharge from hospital (Dua 2014).

Selective reporting

We also assessed routine data systems as being at low risk of reporting bias (Figure 2). Most of the ITS studies used computerised pharmacy systems to measure drug consumption.

Other potential sources of bias

Less than 25% of RCTs provided clear information about baseline outcome; most of these were cluster RCTs (Figure 2). The most common single risk of bias for ITS studies was that the intervention was not independent of other changes (Figure 2). For ITS studies, the main risks of bias were that there were insufficient data to account for seasonal variation or that one or more of the microbial 'Risk of bias' criteria were present (Figure 2).

Studies included in evidence synthesis and 'Summary of findings' tables

Outcomes from 49 (84%) of the 58 RCTs and110 (80%) of the 138 ITS studies were used in at least one meta‐analysis or meta‐regression or are summarised in text or Additional tables. The contribution that each RCT made can be found in Appendix 4. One ITS study contributed data about unintended consequences (Bell 2014). The contribution of 109 ITS studies to meta‐regression of prescribing outcomes is summarised in Appendix 5. Reasons for exclusion of 10 RCTs and 28 ITS studies from evidence synthesis can be found in Appendix 6.

The 10 case control, cohort, or qualitative studies of unintended consequences all contributed evidence about adverse effects.

None of the 6 CBAs or 8 NRTs included evidence about adverse effects of interventions, and there were not enough studies for evidence synthesis.

Intended prescribing outcomes for RCTs and ITS studies included in evidence synthesis

Interventions were targeted at antibiotic treatment for 46 (94%) of 49 RCTs and 101 (92%) of 110 ITS studies. The remaining 11 studies targeted surgical antibiotic prophylaxis (Bell 2014; Dull 2008; Gulmezoglu 2007; Kritchevsky 2008; Meyer 2010; Perez 2003; Schwann 2011; Sun 2011; Van Kasteren 2005; Wax 2007; Weinberg 2001).

For the 148 interventions targeted at antibiotic treatment, the intended outcome of 137 (93%) interventions was to decrease excessive use of antibiotics: 45/46 (98%) RCTs and 93/102 (91%) ITS studies. The only RCT that was primarily intended to increase effective treatment targeted dosing of gentamicin (Burton 1991). Two RCTs with antibiotic choice as the primary outcome did include time to first antibiotic dose for participants with community‐acquired pneumonia as a secondary outcome (Schouten 2007; Yealy 2005). The only other evidence about increasing effective treatment of sepsis came from six ITS studies that aimed to reduce time to first antibiotic dose (Barlow 2007; Hitti 2012; Jobson 2015; Marwick 2013; Volpe 2012; Weiner 2009).

In contrast, reduction in excessive use of antibiotics was the intended outcome of only 3 (25%) of the 12 interventions targeted at surgical antibiotic prophylaxis (Bell 2014; Sun 2011; Van Kasteren 2005). The remaining nine interventions were all intended to increase effective use of antibiotics by increasing the number of participants who received prophylaxis or reducing the time to first antibiotic dose.

Effectiveness and adverse effects of interventions

Effectiveness of interventions in RCTs

Interventions were associated with an increase in compliance with desired practice by 15% (95% confidence interval (CI) 14% to 16%) in 29 RCTs (Analysis 1.1; Figure 3). We obtained similar results in sensitivity analyses for unit of analysis errors (Analysis 1.2) or risk of bias (Analysis 1.3). Interventions were associated with a reduction in duration of total antibiotic treatment by ‐1.95 days (95% CI ‐2.22 to ‐1.67) in 14 RCTs (Analysis 1.4; Figure 4). We obtained similar results in sensitivity analyses for unit of analysis errors (Analysis 1.5) or risk of bias (Analysis 1.6).

Figure 3

---

---

Forest plot of comparison: 1 Prescribing: RCTs of all interventions to reduce unnecessary prescribing, outcome: 1.1 Dichotomous outcomes, increase in desired practice.

Figure 4

---

---

Forest plot of comparison: 1 Effectiveness: Prescribing outcomes from RCTs of interventions to reduce unnecessary antibiotic use, outcome: 1.4 Continuous outcomes, duration of all antibiotic treatment (days).

In four RCTs the prescribing outcome was the consumption of targeted antibiotics measured in different units (cost, days, or defined daily dose), so results were expressed as standardised mean reduction (Analysis 1.7.).

Adverse effects of interventions

Evidence from RCTs

Interventions were not associated with any increase in mortality (95% CI 1 to 0 fewer deaths per 100 participants) in 28 RCTs (Analysis 2.1; Figure 5). We obtained similar results in sensitivity analyses for unit of analysis errors (Analysis 2.2) or risk of bias (Analysis 2.3). Interventions were associated with reduction in length of stay by ‐1.12 days (95% CI ‐1.54 to ‐0.70) in 15 RCTs Analysis 2.4; Figure 6). We obtained similar results in sensitivity analyses for unit of analysis errors (Analysis 2.5) or risk of bias (Analysis 2.6). We found no evidence of a difference in results for interventions that targeted antibiotic exposure (decision to treat or duration of all antibiotic treatment) versus the choice of antibiotic prescribed (Analysis 3.1; Analysis 3.2; Analysis 4.1; Analysis 4.2).

Figure 5

---

---

Forest plot of comparison: 2 Adverse effects: Clinical outcomes from RCTs of interventions to reduce unnecessary antibiotic use, outcome: 2.1 Mortality, all RCTs.

Figure 6

---

---

Forest plot of comparison: 2 Adverse effects: Clinical outcomes from RCTs of interventions to reduce unnecessary antibiotic use, outcome: 2.4 Length of stay, all RCTs.

One RCT measured clinical outcome as potentially harmful delay in essential treatment (Strom 2010). The outcome was ascertained by the Trial Monitoring Committee, who stopped the trial prematurely when four participants were found to have potentially harmful delay in treatment with trimethoprim‐sulphamethoxazole or warfarin. This was a restrictive intervention intended to prevent interactions between these drugs.

Evidence from NRS

ITS studies

Clinical outcome data were measured as mortality in four ITS studies (Table 2) and length of stay in one ITS study (Table 3). However, we could only calculate 95% CI for three of these studies (Lee 2014; Popovski 2015; Skaer 1993), and the outcome data came from all participants in the hospital rather than just the participants who were the targets of the interventions.

Open in table viewer

Table 2. Unintended consequences of ITS studies: mortality*

Study	Prescribing target	Restriction	Design of analysis	Effect estimate	95% CI
Lee 2014	Choice of drug	No	Cohort	Incidence rate ratio 1.1	0.9 to 1.5
Popovski 2015	Choice of drug	No	Cohort	Increase by 1.4%	‐1.2% to 4.1%
Wang 2014	Choice of drug	Yes	ITS, segmented regression	Change in slope ‐0.0172	No data
Yoon 2014	Choice of drug	Yes	Cohort	+0.43 per 1000 OBD	No data

*Mortality was measured in all patients in the hospital rather than just those patients who were the targets of the interventions.

CI: confidence interval
ITS: interrupted time series
OBD: occupied bed day

Open in table viewer

Table 3. Unintended consequences of ITS studies: length of stay*

Study	Prescribing target	Restrictive	Design of analysis	Effect estimate	95% CI
Mittal 2014	Exposure, % treated	No	Cohort	‐0.5 days	No data
Skaer 1993	Choice of drug	No	Cohort	‐0.1 days	‐0.49 to +0.29

*Length of stay was measured in all patients in the hospital rather than just those patients who were the targets of the interventions.

CI: confidence interval
ITS: interrupted time series

Three ITS studies reported other clinical outcomes that provided more direct evidence about unintended consequences of the interventions (Table 4). An intervention to promote gentamicin for prophylaxis was intended to reduce risk of CDI but was associated with a large increase in acute kidney injury in the participants undergoing target operations, and as a consequence the antibiotic policy change was reversed (Bell 2014). An intervention designed to shorten time to first antibiotic dose for people with sepsis was not associated with any increase in the time left without being seen for all other participants in the emergency department (Volpe 2012). An intervention to reduce the duration of surgical antibiotic prophylaxis was not associated with increased surgical‐site infection (Van Kasteren 2005).

Open in table viewer

Table 4. Unintended consequences of ITS studies: other

Study	Prescribing target	Design of analysis	Effect measure	Effect estimate	95% CI
Bell 2014	Antibiotic choice	ITS, segmented regression	Risk of postoperative acute kidney injury	Increase 98%	93.8% to 94.2%
Van Kasteren 2005	Exposure, duration	Cohort	Surgical‐site infection	Decrease 0.8%	‐2.2% to 0.6%
Volpe 2012	Time to first antibiotic dose	Cohort	Left without being seen rate	Decrease 0.4%	No data

CI: confidence interval
ITS: interrupted time series

Case control, cohort and qualitative studies

Ten studies investigated unintended consequences of interventions to change antibiotic choice with cohort (n = 8), case control (n = 1), or qualitative case study (n = 1) designs (Table 5).

Open in table viewer

Table 5. Unintended consequences studies (case control, cohort, or qualitative)

Study	Design	Patients	Intended target	Unintended consequence	Effect estimate	95% CI
Interventions with a restrictive component
Baysari 2013	Qualitative	36 physicians	Reduce unnecessary use of restricted antibiotics	Inaccurate feedback	Not quantified; qualitative study
Calfee 2003	Case control	Not clear		Increase in physician‐based diagnosis of nosocomial infection	No denominator data
Connor 2007	Cohort	120		Failure to warn prescribers about discontinuation	—	—
Duvoisin 2014	Cohort	222	Reduce unnecessary laboratory tests	Delay in TFAD (HR > 1 shows delay less likely in intervention period)	Multivariate HR 1.56	1.17 to 2.07
LaRosa 2007	Cross‐sectional	15,440	Reduce unnecessary use of restricted antibiotics	Orders for restricted antibiotics (% all orders) from 10 to 11 pm vs all other hours	—	—
	Cohort	360		% appropriate orders 10 to 11 pm vs 9 to 10 pm	‐23.7%	‐31.8% to ‐15.5%
Linkin 2007	Cohort	200		Risk of inaccurate information in orders judged inappropriate vs appropriate	OR 2.2	1.0 to 4.4
Winters 2010	Cohort	3251		Risk of 1‐hour delay in TFAD	OR 1.5	1.2 to 1.8
				Risk of 2‐hour delay in TFAD	OR 1.8	1.4 to 2.2
Interventions with no restrictive component
Friedberg 2009	Cohort	13,042	Reduce time to first antibiotic dose for patients with community‐acquired pneumonia	% CAP diagnoses	1% increase	No denominator data
Kanwar 2007	Cohort	518		% correct CAP diagnoses	‐7.9% decrease	‐15.4% to ‐0.4%
Welker 2008	Cohort	548		% correct CAP diagnoses	‐16.0% decrease	‐7.6% to ‐24.4%

CAP: community‐acquired pneumonia
CI: confidence interval
HR: hazard ratio
OR: odds ratio
TFAD: time to the first antibiotic dose

There was a restrictive component to the intervention in seven studies. One study showed that restriction of laboratory tests of inflammation (C‐reactive protein and white blood cell count) was not associated with an increase in time to first antibiotic dose (Duvoisin 2014). The remaining six studies all revealed unintended consequences of interventions that restricted antibiotic choice by requiring prior approval, as follows.

• Negative professional culture through breakdown in trust and communication (Baysari 2013; Calfee 2003; Connor 2007; Linkin 2007).

• Delay in time to first antibiotic dose (LaRosa 2007; Winters 2010). Evidence of delay in essential treatment was also seen in one RCT (Strom 2010).

In three studies (Friedberg 2009; Kanwar 2007; Welker 2008), the intervention was a national financial incentive in the USA that was intended to reduce time to first antibiotic dose for people admitted to hospital with community‐acquired pneumonia (CAP). In all three studies, the unintended consequence was misdiagnosis of pneumonia, which could lead to an increase in unnecessary antibiotic treatment. In two single‐centre studies, there was a decrease in the percentage of participants with correct diagnosis of CAP based on prespecified criteria (Kanwar 2007; Welker 2008). In contrast, a large, multicentre study reported no evidence of an overall increase in the diagnosis of CAP (Friedberg 2009); however, this study was at high risk of bias.

Explaining heterogeneity in the intended effect of interventions

Meta‐regresson of RCTs

We performed meta‐regression on 29 RCTs with dichotomous prescribing outcomes (Analysis 1.1; Figure 3). Outcomes for all of these trials could be expressed as number of participants where treatment was compliant with policy divided by total participants. We did not perform meta‐regression on 15 RCTs with continuous prescribing outcomes because the outcomes were heterogeneous (Analysis 1.4; Analysis 1.7) and because none of the interventions included restriction or feedback, and only two did not include enablement (Danaher 2009; Kerremans 2008).

Meta‐regression results for 29 RCTs with dichotomous outcomes

In the meta‐regression, enablement, restriction, targeting antibiotic choice versus exposure and high risk of bias were significantly associated with greater intervention effect in univariate analysis, and they all remained significant in multivariate analysis (Figure 7).

Figure 7

---

---

Meta‐regression by effect modifier for 29 RCTs. A positive value for Beta indicates enhanced intervention effect. One RCT had both enabling and restrictive components in the intervention (Strom 2010).

Of the 23 RCTs of enabling interventions, four also included feedback (Camins 2009; Schnoor 2010; Schouten 2007; Yealy 2005). All four of these RCTs targeted antibiotic choice, so we have compared their effects with seven RCTs of enabling interventions without feedback that also targeted antibiotic choice. The mean risk difference for interventions with feedback was 19% (95% CI 16% to 22%) (Figure 8) compared with 13% (95% CI 9% to 17%) (Figure 9) for interventions with no feedback. Only two of the feedback RCTs also included action planning (Schouten 2007; Yealy 2005).

Figure 8

---

---

Forest plot of comparison 5: RCTs of enablement with and without feedback, outcome: 5.1 Enablement plus feedback.

Figure 9

---

---

Forest plot of comparison 5: RCTs of enablement with and without feedback, outcome: 5.2 Enablement without feedback.

Meta‐regression of ITS studies

Do interventions that involve enablement have greater initial effect?

There were 107 ITS studies with data that could be used for meta‐regression of prescribing outcomes at one, six, or 12 months' postintervention. We used multivariable meta‐regression to identify effect modifiers in 91 ITS studies including data about prescribing at six months' postintervention. As with the RCTs (Figure 7), both enablement and restriction were independently associated with increased effect in ITS studies (Figure 10). Of 29 ITS studies with restrictive interventions, 13 (45%) also had enablement, and this independently enhanced intervention effect (Figure 11). In comparison with interventions targeting antibiotic exposure, those targeting choice were associated with greater effect in RCTs (Figure 7), but not in ITS studies (Figure 10). The number of studies in each category only allowed analysis of the effects of setting in ITS studies (Figure 10), and intention could only be included in meta‐regression of ITS studies of enabling intervention (Figure 12). The limited evidence suggests that intention and setting were not effect modifiers (Figure 7; Figure 10).

Figure 10

---

---

Meta‐regression by effect modifiers of intervention for 91 ITS studies. Outcome is effect on prescribing six months' postintervention. There are 16 studies with both enabling and restricting intervention components (Figure 11).

Figure 11

---

---

Meta‐regression of prescribing outcome by effect modifiers for 29 ITS studies of interventions that included restriction.

Figure 12

---

---

Meta‐regression by effect modifier for 43 ITS studies of interventions that included enablement but not restriction. Outcome is effect on prescribing six months' postintervention. Note that four studies with feedback were not included in this analysis because they also included restriction.

Are interventions that include feedback more effective than those that do not?

Feedback was included in 4 (17%) of 23 RCTs (Figure 8) and 20 (47%) of 43 ITS studies (Figure 12) of enabling interventions that did not include restriction. The intervention was audit and feedback alone in three RCTs and 10 ITS studies. In one RCT and 11 ITS studies, audit and feedback was combined with review and recommend change or circumstantial reminders. Interventions that included feedback were more effective than those that did not. However, there were too few studies with goal setting or action planning to assess their effect in addition to feedback.

There were only two ITS studies with enough data to analyse the effect of adding an additional component to an effective intervention. However, the second intervention component did not include goal setting, feedback, or action planning in either study (Mol 2005; Po 2012)

Summary of interventions for the studies included in meta‐regression

In comparison with RCTs, the ITS studies were more likely to have multiple intervention components: 35 (38%) of 91 ITS studies versus 5 (17%) of 29 RCTs, odds ratio 3.00 (95% CI 1.05 to 8.59) (Table 6). There were also differences in the components for enabling interventions (review and recommend change was included in 53% of ITS studies versus 25% of RCTs) and restrictive interventions (removal of target drugs from clinical areas was included in 34% of ITS studies but in no RCTs) (Table 6). Educational meetings or distribution of educational materials was the most common intervention in studies that did include enablement or restriction (75% of RCTs and 89% of ITS studies) (Table 6).

Open in table viewer

Table 6. Summary of intervention components for 29 RCTs (Analysis 1.1; Figures 3 and 7) and 91 ITS studies (Figure 10)

Intervention function and components	RCT	ITS
Enablement	24 studies	59 studies
Number of enabling or restrictive intervention components	27	76
Studies with > 1 Enabling intervention component	2 8%*	19 32%*
Audit and feedback	4 17%	24 41%
Computerised decision support	1 4%	3 5%
Circumstantial reminders	16 67%	18 31%
Review and recommend change	6 25%	31 53%
Restriction	2 studies	29 studies
Number of Restrictive intervention components	3	41
Studies with > 1 Restrictive intervention component	1 50%	10 34%
Expert approval	1 50%	18 62%
Compulsory order form	1 50%	7 24%
Removal	0	10 34%
Review and make change	1 50%	6 21%
No Enablement or Restriction	4 studies	18 studies
Number of intervention components	6	25
Studies with > 1 intervention component	2 50%	6 33%
Educational materials or meetings	3 75%	16 89%
Educational outreach (academic detailing)	1 25%	6 33%
Physical reminders	1 25%	2 11%
Structural intervention	1 25%	1 6%

*The denominator for all percentages is the number of studies for each intervention function. One RCT, Strom 2010, and 16 ITS studies (Figure 11) included both enabling and restrictive intervention components.

ITS: interrupted time series
RCT: randomised controlled trial

Sustainability of intervention effect

Sustainability was assessed in 64 of 91 ITS studies, with prescribing outcome data at both 6 and 12 months' postintervention. Intervention effect was sustained at 12 months' postintervention in 55 (86%) of these studies (95% CI 77% to 94%). There were 13 interventions with neither enablement nor restriction; intervention effect was sustained in 11 (85%) (95% CI 65% to 100%). Consequently, it was unlikely that either enablement or restriction would be associated with greater sustainability. However, the results suggest that restrictive interventions were less likely to have sustained effect if they did not include enablement: 5/8 (62%) versus 12/13 (92%) with enablement, risk difference 30% (95% CI ‐7% to 66%).

Five ITS studies with data about removal of interventions provided additional information about sustainability of interventions (Table 7). Three of these studies also provided data about the effect of the intervention. The intended effect of all interventions was decrease in the use of target antibiotics. Removal of the intervention was associated with increase in the use of target antibiotics in all five studies and, with one exception (Kim 2008), the 95% CI for effect size did not include decrease in use of target antibiotics. Kim 2008 was the only one of these five interventions including enablement by audit and feedback.

Microbial outcomes (antibiotic resistance and CDI)

There were 1 CBA and 5 RCTs with microbial outcome data, and these were too heterogeneous for data synthesis (Table 8).

We performed meta‐regression on 26 ITS studies including reliable data about prescribing outcomes at 6 months and microbial outcomes at 12 months after the intervention (Table 9). Six unplanned interventions (in response to outbreaks) were associated with markedly greater effect on microbial outcomes (Figure 13). When studies were ranked in descending order of effect size for microbial outcome at 12 months, the top five studies were all unplanned interventions (Kim 2008; May 2000; McNulty 1997; Tangdén 2011; Valiquette 2007), with the remaining unplanned intervention ranking 9th (Lautenbach 2003).

Figure 13

---

---

Meta‐regression by effect modifiers for 34 microbial outcomes 12 months' postintervention from 26 ITS studies. The bars show the results for unadjusted versus adjusted analyses, the comparison for unplanned interventions is with planned interventions in both the unadjusted and adjusted analysis.

CDI: Clostridium difficile infection
GPC: infection with antibiotic‐resistant gram‐positive cocci
GNB: infection with antibiotic‐resistant gram‐negative bacteria

Other infection control: 'Yes' means there were changes to infection control processes during the study period.

In the 20 studies of planned intervention, there were six studies with unclear information about other infection control interventions or changes during the study period (Chan 2011; Grohs 2014; Jump 2012; Liebowitz 2008; Meyer 2009; Petrikkos 2007). We performed meta‐regression on the remaining 14 studies from Table 9 (Figure 14). In contrast with the meta‐regression of all 27 studies (Figure 13), the effects of setting, other infection control interventions, and microbial outcome type were all reversed so that each of these variables was associated with increase in effect size in the 14 studies with planned interventions and details of other infection control interventions (Figure 14).

Figure 14

---

---

Meta‐regression by effect modifiers for 20 microbial outcomes 12 months' postintervention from 14 ITS studies of planned interventions that provided details about other infection control changes or interventions.

CDI: Clostridium difficile infection
GPC: infection with antibiotic‐resistant gram‐positive cocci
GNB: infection with antibiotic‐resistant gram‐negative bacteria

Other infection control: 'Yes' means there were changes to infection control processes during the study period.

The antibiotic targets for the 20 studies of planned interventions were single antibiotic classes in nine studies (Cook 2011b; Grohs 2014; Knudsen 2014; Lafaurie 2012; Lee 2007; Meyer 2009; Petrikkos 2007; Willemsen 2010; Yoon 2014), high‐risk antibiotics in nine studies (Aldeyab 2012; Aldeyab 2014; Ananda‐Rajah 2010; Buising 2008a; Chan 2011; Dancer 2013; Fowler 2007; Liebowitz 2008; Talpaert 2011), and all antibiotics in the remaining two studies (Cook 2011a; Jump 2012). High‐risk antibiotics were a combination of drugs from more than one class of antibiotic, which were all considered to be high risk for the microbial outcome. The prescribing outcome data reported in these nine studies varied from just one of the high‐risk antibiotics, in Dancer 2013, through individual results for all of the high‐risk antibiotics, in Buising 2008a, Chan 2011, Fowler 2007, and Talpaert 2011, to combined results for all of the high‐risk antibiotics (Aldeyab 2012; Aldeyab 2014; Ananda‐Rajah 2010; Liebowitz 2008).

One study can be used to demonstrate the technical challenges of estimation of intervention effect on microbial outcomes (Dancer 2013). The intervention was addition of complete restriction of ceftriaxone and ciprofloxacin to a pre‐existing multifaceted intervention introduced seven months before restriction and remaining in place throughout the restrictive period (Dancer 2013). We could not analyse the effect of the initial multifaceted intervention because there were no pre‐intervention data about prescribing or microbial outcomes. However, the available data showed CDI was lower by ‐0.143 cases per 1000 occupied bed days per month in the nine months prior to the addition of the restrictive intervention. At the start of the restrictive intervention, CDI rates were already low (1.5 cases per 1000 occupied bed days). After the introduction of restriction, CDI rates continued to decline for five months, and then stabilised at around 0.5 cases per 1000 occupied bed days. These data suggest that the restrictive intervention had no additional effect on the rate of CDI. However, the segmented regression analysis estimated that there was a relative increase of 35.8% in CDI rate 12 months after the restrictive intervention with very wide confidence intervals (from 81.0% decrease to 152.7% increase).

Our review did include one multicentre controlled ITS study comparing CDI rates in six hospitals with antimicrobial stewardship programmes versus four control hospitals (Ostrowsky 2014). We did not include this study in evidence synthesis because neither the interventions nor the prescribing outcomes were standardised across the six hospitals with stewardship programmes. Baseline rates of CDI were only 0.8 cases per 1000 occupied bed days in the intervention and control hospitals before the intervention, and the authors did not report a decrease in aggregate CDI rates either between intervention and non‐intervention groups or within the intervention groups over time (Ostrowsky 2014).

We have not attempted to synthesise microbial outcome data because of the small number of studies, the heterogeneity of intervention targets and prescribing outcomes, and the wide confidence intervals for estimated relative effect. We have focused on the 20 ITS studies of planned interventions and separated the results by microbial outcome type. Interventions were associated with consistent reduction in CDI (median ‐48.6%, interquartile range ‐80.7% to ‐19.2%) but inconsistent effect on resistant gram‐negative bacteria (median ‐12.9%, interquartile range ‐35.3% to 25.2%) and resistant gram‐positive bacteria (median ‐19.3%, interquartile range ‐50.1% to 23.1%). There were too few studies with too much variance in microbial outcomes to reliably assess the relationship between change in antibiotic use and each of the microbial outcomes.