Solomon 2001.

Methods	STUDY DESIGN: cluster RCT, service level Risk of Bias: HIGH
Participants	PROVIDERS: 17 Internal Medicine services randomly assigned to intervention (9 services) or control (8 services) PARTICIPANTS: a total of 4500 patients admitted during the baseline and study periods, of whom 260 patients received 278 unnecessary prescriptions for the target drugs; 17 clusters (services) CLINICAL PROBLEM: patients receiving ceftazidime or levofloxacin. SETTING: 1 hospital in the USA POWER CALCULATION: no information. The methods say that the statistical model adjusted for clustering, but no results are given (see risk of bias).
Interventions	FORMAT: Interventions: educational meetings with dissemination of policy for necessary use; educational outreach by review and recommend change, either verbal (face to face or telephone) or by email Intervention Functions: education, enablement, persuasion COMPARISON: randomly assigned control services DESIRED CHANGE: reduce excessive POWER CALCULATION: no information. Note from Statistician: The study adjusted for some clustering, but possibly only in the repeated measures, not in the hospitals. Just using the results from Table 2, I do not get the P value that they state in the table using a unit of analysis error approach. This suggests to me that they are adjusting for "things". I therefore think on balance that it is probably OK to use the results.
Outcomes	PRESCRIBING: Choice: % patients with target antibiotics discontinued. Exposure: % patients with all antibiotics discontinued CLINICAL: inpatient mortality, transfer to ICU, length of stay, and re‐admission within 30 days of discharge FINANCIAL: estimated annual cost of the intervention
Notes	FINANCIAL SUPPORT: Funding: Brigham and Women’s Hospital and Arthritis Foundation Investigator Award. Competing Interests: no information ADDITIONAL DATA: email from authors with information about the intervention
Risk of bias
Bias	Authors' judgement	Support for judgement
Random sequence generation (selection bias)	Low risk	"We assigned services to intervention or control status using a blocked randomization design"
Allocation concealment (selection bias)	Unclear risk	Not concealed
Blinding (performance bias and detection bias) All outcomes	High risk	No blinding
Incomplete outcome data (attrition bias) All outcomes	Unclear risk	Figure 2 and text give %, no denominator.
Selective reporting (reporting bias)	Unclear risk	Figure 2 and text give %, no denominator.
Other bias	Unclear risk	The methods say: "To estimate the relative reduction in unnecessary use of target antibiotics in the intervention group, we used a fixedeffects model (PROC GENMOD in SAS statistical software).20 This model used a log‐linear link function, assumed a Poisson distribution, and accounted for overdispersion. Experimental group assignment (intervention or control) was the independent variable of interest, the individual service was considered a class effect, and covariates included level of baseline prescribing and time, modeled as both a linear and categorical effect. The interaction between assignment and time was also assessed. We further considered a linear randomeffects model to account for variation between services (PROC MIXED in SAS statistical software)20; the results of this analysis were similar to those found in the fixed‐effects models with respect to the level of statistical significance, and only the fixedeffects model results are presented." However, no model outputs are given in the results (only point estimates), and the discussion says only: "This significant effect of the intervention remained after adjusting for baseline prescribing, clustering of repeated measures within a given service, and duration of the intervention."
Baseline Outcomes similar?	Low risk	Figures 1 and 2
Free of contamination?	High risk	The services were in the same hospital.
Baseline characteristics similar?	Low risk	Table 1