Ukoumunne OC, Guilliford MC, Chinn S, Sterne JA, Burney PG; Association for Health Services Research. Meeting.
Abstr Book Assoc Health Serv Res Meet. 1999; 15: 72.
Department of Public Health Sciences, GKT School of Medicine, Kings College London, London SE1 3QD, UK.
OBJECTIVE: To carry out a methodological systematic review with the aim of producing guidelines for evaluating interventions implemented at the level of health service organisational unit or geographical area. METHODS: We searched electronic databases, hand-searched journals, screened citations and consulted experts. Relevant material was synthesised into a review using qualitative judgements concerning validity. RESULTS: We identified twelve key considerations. In general it is advisable to: 1. recognise areas or organisational clusters as the units of intervention; 2. justify the allocation of entire clusters to groups; 3. include a sufficiently large number of clusters in each group to allow statistically valid conclusions; 4. randomise clusters to groups whenever possible; 5. in non-randomised studies include a control group; 6. in single group studies, include repeated measurements over time; 7. when estimating the required number of individuals, multiply the results of standard sample size calculations by the design effect; 8. consider restricted allocation of clusters to groups in order to reduce baseline imbalances, thereby increasing precision in randomised studies and reducing bias in non-randomised studies; 9. choose between cohort and cross-sectional sampling of individuals within clusters; 10. use methods of analysis that allow for correlation of individual responses within clusters; 11. adjust for individual and cluster level covariates by using regression methods for clustered data; 12. publish estimates for intraclass correlation coefficients which may be used in designing future studies. CONCLUSIONS: In area or organisation-based evaluation, a cluster randomised design incorporating a large number of units of allocation will usually minimise potential error and bias. When randomisation is not feasible, quasi-experimental designs with restricted allocation should be used, with due regard to potential sources of bias. Design and analysis should explicitly recognise the correlation of individual responses within clusters.
Publication Types:
Keywords:
- Bias (Epidemiology)
- Cluster Analysis
- Control Groups
- Delivery of Health Care
- Guidelines as Topic
- Organizations
- Periodicals
- Research Design
- Sample Size
- methods
- organization & administration
- hsrmtgs
Other ID:
UI: 102193985
From Meeting Abstracts