Time series analysis for evaluation of antimicrobial stewardship interventions

  • Andrea Lynn Patton

Student thesis: Doctoral ThesisDoctor of Philosophy


Antimicrobials have been described as the “wonder drug” of the 20th century. Not only are they crucial for the treatment of infection, antimicrobials are used across the whole healthcare system including to support cancer chemotherapy, and organ transplantation. The use, including avoidable misuse and over use, of antimicrobials drives antimicrobial resistance. The World Health Organisation has described antimicrobial resistance as a serious global health problem and strategies are required to minimise overuse and misuse. Antimicrobial stewardship interventions are implemented in the community setting where it is estimated that up to one third of prescriptions are inappropriate and in the hospital setting where, at any given time, approximately one third of all patients receive antibiotics. It is important to understand what antimicrobial stewardship interventions are effective in reducing inappropriate prescribing and influence resistance without worsening clinical outcomes. Robust methods of evaluation are needed. This thesis examines the advantages and disadvantages of methods available to evaluate antimicrobial stewardship interventions.

A systematic review of the quality of design, analysis and reporting in 107 studies identified in a recent Cochrane systematic review of antimicrobial stewardship interventions for hospital inpatients was performed. An uncontrolled before-after design was used to evaluate the prescribing outcome in 44 (41%) studies. Uncontrolled before-after studies do not adjust for underlying trends in the data and are at high risk of bias. There was agreement on the direction of change and statistical significance of the change in only ten (23%) uncontrolled before-after studies when compared to the results of a segmented regression re-analysis of their data. A time series method was used in 60 (56%) studies and the most commonly used analysis methods were segmented regression (70%) and Autoregressive Integrated Moving Average (ARIMA) (12%). There is no published consensus reporting guideline for interrupted time series studies but approximately half of the studies that used segmented regression or ARIMA also reported the results of an uncontrolled before-after analysis to “support” their results. The findings of this systematic review were used to develop recommendations for the design, analysis and reporting of interrupted time series studies.

Analysis at national and regional health board level of a previously published national interrupted time series analysis reporting the success of two antimicrobial stewardship interventions was conducted. The interventions aimed to reduce the use of carbapenems and piperacillin/tazobactam in Scotland and the first intervention was the introduction of a treatment guideline for infection with multidrug resistant Gram negative bacteria (MDRGNB). The second intervention was a quality improvement programme involving a survey of prescribing and laboratory practice, feedback of survey results to individual health boards and discussion of results at national antimicrobial prescribing group meetings. Intervention one had an immediate impact on carbapenem prescribing but the intervention effect was not sustained, and had no impact on piperacillin/tazobactam prescribing. Intervention two had an immediate and sustained impact on both carbapenem and piperacillin/tazobactam prescribing. Analysis at individual board level showed that the intervention impact was not observed in every health board. Some boards responded to the interventions and other boards did not. It was unclear as to what characteristics of boards were associated with an intervention impact. The intervention was evaluated in line with the recommendations developed above.

Multilevel models are an established method to account for any underlying hierarchical structure within aggregated data but have rarely been applied to evaluations of non-randomised healthcare interventions. A multilevel interrupted time series analysis was developed to examine general practice level effects underlying a previously published health board level analysis reporting the impact of an intervention to reduce the use of 4C (co-amoxiclav, ciprofloxacin, cephalosporins and clindamycin) antimicrobials which are associated with a higher risk of Clostridium difficile infection (CDI). Overall the intervention was successful at changing the level and trend of 4C antimicrobial prescribing at board level although each individual practice did not respond to the intervention. Practices that had a higher level of 4C antimicrobial prescribing at the start of the study period tended to have a bigger decrease in use before the intervention and a bigger decrease in level after the intervention compared to practices with lower baseline rates of prescribing. Multilevel interrupted time series models are more useful than multiple single level interrupted time series analyses to understand what practices are more likely to respond to an intervention and allows for inclusion of a larger number of subunits.

Antimicrobial stewardship is a key strategy in tackling the emerging threat of antimicrobial resistance. Evaluation of antimicrobial stewardship interventions are increasingly reported in the literature but methodological flaws in the design, analysis and reporting of these interventions hinder the interpretation and wider implementation of apparently successful interventions. The number of large-scale antimicrobial stewardship programmes is rapidly growing internationally. High quality interrupted time series studies are a key component of robust evaluation, and this thesis has explored how they can best be used in this context.
Date of Award2019
Original languageEnglish
SponsorsThe Health Foundation
SupervisorCharis Marwick (Supervisor) & Bruce Guthrie (Supervisor)


  • antimicrobial stewardship
  • interrupted time series analysis
  • multilevel modelling
  • study design

Cite this