![Acknowledgements
We acknowledge financial support from:
• The National Institute for Health Research (NIHR) Collaborations for
Leadership in Applied Health Research and Care for West Midlands
(CLAHRC WM).
• The Medical Research Council Midland Hub for Trials Methodology
Research [grant number G0800808].](https://image.slidesharecdn.com/khemminglocalrss2015-160830134909/85/An-introduction-to-the-stepped-wedge-cluster-randomised-trial-41-320.jpg)
![References
• Hemming K, Girling A. The efficiency of stepped wedge vs. cluster
randomized trials: stepped wedge studies do not always require a
smaller sample size. J Clin Epidemiol. 2013;66(12):1427-8.
• Hussey MA, Hughes JP. Design and analysis of stepped wedge
cluster randomized trials. Contemp Clin Trials. 2007;28(2):182-91.
• Hemming K, Girling A. A menu driven facility for sample size for
power and detectable difference calculations in stepped wedge
randomised trials. STATA Journal. 2014;[In Press]](https://image.slidesharecdn.com/khemminglocalrss2015-160830134909/85/An-introduction-to-the-stepped-wedge-cluster-randomised-trial-42-320.jpg)
An introduction to the stepped wedge cluster randomised trial
- 1. The stepped wedge trial (SW-CRT): Recommendations for research methods and reporting 1: University of Birmingham, UK 2: University of Warwick, UK 3: University of Ottawa, Canada 30/08/2016 Karla Hemming1 Alan Girling1, James Martin1, Celia Brown2 Richard Lilford2, Peter Chilton1 Monica Taljaard3
- 2. What is a SW-CRT • Modification of cross-over design: - All clusters start in control - Clusters (or groups of clusters) cross to intervention at randomly assigned times until all have received intervention - Outcome typically observed at each time point 1 2 3 4 5 6 TimeSTEPS (Cluster or Group of Clusters) 1 2 3 4 5 Exposed to intervention Unexposed to intervention
- 3. Cross-sectional designs only • Assume all participants at each step or time point are different • Other types of designs include: – Cohort design – where individuals have repeated measures – Open cohort – where individuals have repeated measures and new individuals can join the study over its duration
- 4. Example: the EPOCH trial • Intervention: – Service delivery intervention to improve care of patients undergoing emergency laparotomy • Setting: – Includes 90 hospitals – Rolled out to 15 geographically close hospitals at a time • Outcome: – 90 day mortality • Sample size: – TSS is 27,500 – 90% power to detect a change from 25% to 22%
- 5. Example: The EPOCH study
- 6. Systematic review of SW-CRTs Rapid update 0 5 10 15 20 25 1985 1990 1995 2000 2005 2010 2015 CumulativenumberofSW-CRTspublished Year Cumulativecompleted Cumulativeprotocols(excludes subsequently completed and published studies)
- 7. Quality of reporting of cluster effects Protocols published in journals (N=14) Results papers published in journals (N=18) SS calculation reported 14 (100%) 12 (67%) ICC stated in methods 12 (86%) 5 (28%) Uncertainty of ICC considered in methods 3 (21%) 1 (6%) Results fully accounted for clustering N/A 6 (33%) ICC given in results N/A 1 (6%)
- 8. Some unresolved (or debated) issues: • Design: –How to determine sample size (number clusters, cluster size)? –Which is the most efficient design? • Analysis: How to analyse these studies: –Temporal confounding
- 9. Design
- 10. Conventional representation of designs 10
- 11. This representation leads us to believe… • The SW-CRT is of longer duration than the PCRT and CRT-BA • The cluster sizes in the SW-CRT are larger than those in the PCRT • The SW-CRT allows staggered roll-out and the PCRT does not
- 12. This representation leads us to believe… • The SW-CRT is of longer duration than the PCRT and CRT-BA – We claim: false • The cluster sizes in the SW-CRT are larger than those in the PCRT – We claim: false • The SW-CRT allows staggered roll-out and the PCRT does not – We claim: false
- 14. Motivating example… • PCRT in primary care • GP practices are clusters • Patients presenting with a new diagnosis of diabetes are the individuals • These patients wont all present at a fixed point in time • Rather they will become eligible for the study over a prolonged period of time Time
- 15. Using this representation: • The SW-CRT is of the same duration as the PCRT and CRT-BA • The cluster sizes in the SW-CRT are the same as those in the PCRT • The SW-CRT allows staggered roll-out and so does the PCRT
- 16. The staggered PCRT • One often cited reason for the SW-CRT is the phased implementation • This is possible under parallel design • Balanced on time, so no time effects 0 1 2 3 4 5 Time (step) Cluster 1 2 3 4 5 6 Intervention Control Staggered Cluster Study Cluster unexposed to intervention Cluster in transition period Cluster exposed to intervention
- 17. Efficiency How to determine which design to use?
- 18. Efficiency – depends on ICC ICC=0.01 ICC=0.1 PCRT PCRT-BA SW-CRT PCRT PCRT-BA SW-CRT Number of clusters 20 20 20 20 20 20 Cluster size 50 50 50 50 50 50 Total sample size 1000 1000 1000 1000 1000 1000 Number of steps 0 1 4 0 1 4 Number of clusters per step 10 5 10 5 Power 0.97 0.87 0.88 0.50 0.77 0.82 Study to detect a moderate effect size of 0.3 (SD 1) at 5% significance
- 19. Sometimes the CRT is infeasible • Example: – NI is 788: 0.2 80% power and 5% significance – ICC is 0.10 – 30 clusters • Can’t run this trial using a parallel design: – minimum number of clusters is (p*NI) – i.e. 788*0.1=79 • Under SW-CRT with 4 steps: – Need 75 observations in 30 clusters – TSS of 2250
- 21. Minimise total sample size or clusters? • Design constraints – NI is 6,426; 10% to 8%; ICC 0.05 – Cluster trial – Over each year possible to recruit M=800 per cluster • SW-CRT : – 4 steps, 26 clusters, 1 year, TSS=20,800 • CRT: – M=800, 330 clusters, 1 year, TSS=264,000!!!! – M=200, 354 clusters, 3 months, TSS=70,000 – M=20, 630 clusters, Random Sample, TSS=12,500
- 22. Sample size and power How do we work it out?
- 23. Simple notation Notation Sample size for RCT NI ICC p Number clusters k Number steps t Cluster size per step m Total cluster size M
- 24. Sample size calculations • Accommodate: – Clustering – Time effects • Seminal paper by Hussey and Hughes – Power for fixed design • Algebraically complicated, BUT: – Stata Function 0 1 2 3 4 5 Time (step) Cluster 1 2 3 4 5 6 Intervention Control Conventional Stepped Wedge Study
- 25. Stata power function Extensions allow: • Two levels – i.e. wards within hospitals • Transition periods – i.e. training periods • Varying cluster size – (work in progress) Hemming K, Girling A. A menu driven facility for sample size for power and detectable difference calculations in stepped wedge randomised trials. STATA Journal. 2014
- 26. Determining number of clusters: • Design effect (Woerterman, 2013): • Sample size needed: N=TSSRCT *DESW * (t+1) Hemming K, Girling A. The efficiency of stepped wedge vs. cluster randomized trials: stepped wedge studies do not always require a smaller sample size. J Clin Epidemiol. 2013;66(12):1427-8.
- 27. Determining number of clusters: • Design effect (Woerterman, 2013): • Sample size needed: N=TSSRCT *DESW * (t+1) Hemming K, Girling A. The efficiency of stepped wedge vs. cluster randomized trials: stepped wedge studies do not always require a smaller sample size. J Clin Epidemiol. 2013;66(12):1427-8. Need this!
- 28. What cluster size do I need? Setting straight the sample size determination for stepped wedge and cluster randomised trials: design effects and illustrative examples Karla Hemming and Monica Taljaard Submitted to J Clin Epi
- 29. Analysis
- 30. Analysis • Summarise key characteristics by exposure / unexposed status – Identify selection biases • Analysis either GEE or mixed models – Clustering – Time effects • Imbalance of calendar time between exposed / unexposed: – The majority of the control observations will be before the majority of the intervention observations – Time is a confounder! • Unadjusted effect meaningless Hemming K., Haines T.P., Chilton P.J., Girling A.J., Lilford R.J. The stepped wedge cluster randomised trial: rationale, design, analysis and reporting. The BMJ, in press
- 31. Example 1: Maternity sweeping 31 • Objective: evaluate a training scheme to improve the rate of membrane sweeping in post term pregnancies – Primary outcome: • Proportion of women having a membrane sweep – Cluster design: • 10 teams (clusters) • Pragmatic design – rolled out when possible • Transition period to allow training
- 32. Example 1: Maternity sweeping (transition period) 32
- 33. Example 1: Underlying trend 0 .2.4.6.8 20005/03/12 23/04/12 18/06/12 13/08/12 week commencing P-value for trend <0.05
- 34. Example 1: results Unexposed to intervention n=1417 Exposed to intervention n=1356 Relative Risk P- value Number of women offered and accepting membrane sweeping Number (%) 629 (44.4%) 634 (46.8%) Cluster adjusted 1.06 (0.97, 1.16) 0.21 Time and cluster adjusted Fixed effects time 0.88 (0.69, 1.12) 0.30 Linear time effect 0.90 (0.73, 1.11) 0.34
- 35. Example 1: results Unexposed to intervention n=1417 Exposed to intervention n=1356 Relative Risk P- value Number of women offered and accepting membrane sweeping Number (%) 629 (44.4%) 634 (46.8%) Cluster adjusted 1.06 (0.97, 1.16) 0.21 Time and cluster adjusted Fixed effects time 0.88 (0.69, 1.12) 0.30 Linear time effect 0.90 (0.73, 1.11) 0.34 Going up!
- 36. Example 1: results Unexposed to intervention n=1417 Exposed to intervention n=1356 Relative Risk P- value Number of women offered and accepting membrane sweeping Number (%) 629 (44.4%) 634 (46.8%) Cluster adjusted 1.06 (0.97, 1.16) 0.21 Time and cluster adjusted Fixed effects time 0.88 (0.69, 1.12) 0.30 Linear time effect 0.90 (0.73, 1.11) 0.34 Going up! Going down!
- 37. Explanations • Rising tide – General move towards improving care – perhaps due to very initiative that prompted study investigators to do this study • Contamination – Unexposed clusters became exposed before their randomisation date • Lack of precision – Intervention wasn’t ruled out as being effective
- 38. Recommendations
- 39. Recommendations • SW-CRT a pragmatic study design which reconciles the need for robust evaluations with political or logistical constraints. – But, can have a staggered parallel CRT • The SW-CRT design is recommended when: – Higher the ICC (process outcomes) – Limited number of clusters – Routinely collected outcome data • Design and analysis – Appropriate consideration of time effects in power and analysis
- 40. Next steps … • Published a set of recommendations for reporting in BMJ.. out soon… • Updating the systematic review of quality of reporting – Look at quality of reporting of SS calculations – Looking at ethical issues around recruitment and concealment of allocation • Consort Extension for SW-CRTs • Alan and James – numerical work on varying cluster size
- 41. Acknowledgements We acknowledge financial support from: • The National Institute for Health Research (NIHR) Collaborations for Leadership in Applied Health Research and Care for West Midlands (CLAHRC WM). • The Medical Research Council Midland Hub for Trials Methodology Research [grant number G0800808].
- 42. References • Hemming K, Girling A. The efficiency of stepped wedge vs. cluster randomized trials: stepped wedge studies do not always require a smaller sample size. J Clin Epidemiol. 2013;66(12):1427-8. • Hussey MA, Hughes JP. Design and analysis of stepped wedge cluster randomized trials. Contemp Clin Trials. 2007;28(2):182-91. • Hemming K, Girling A. A menu driven facility for sample size for power and detectable difference calculations in stepped wedge randomised trials. STATA Journal. 2014;[In Press]
Editor's Notes
- #3: At the moment some of the pictures are black and white – other colour and different colours. This is because I’ve taken them from your pictures are various different stages. I wondered if you could format them so they are all the same format and colour? The ones I think it will be simplest to change are on slides 2 and 16 – and make these the same format as the pictures which are in the green colour which we seem to have adopted now.
- #15: On slide 14 I want to set up an animation. I want a line to represent time (maybe saying time on it). Then I want little people to appear along this line – representing how patients might present in a clinical trial – ie they don’t all present at a fixed point in time…
- #17: At the moment some of the pictures are black and white – other colour and different colours. This is because I’ve taken them from your pictures are various different stages. I wondered if you could format them so they are all the same format and colour? The ones I think it will be simplest to change are on slides 2 and 16 – and make these the same format as the pictures which are in the green colour which we seem to have adopted now.