STATISTICS IN MEDICINE Statist. Med. 2001; 20:367–376 Sample size calculations for intervention trials in primary care randomizing by primary care group: an empirical illustration from one proposed intervention trial Sandra Eldridge 1; *; † , Colin Cryer 2 , Gene Feder 3 and Martin Underwood 3 1 Department of Environmental and Preventive Medicine; Wolfson Institute of Preventive Medicine; St Bartholomew’s and the Royal London School of Medicine and Dentistry; Queen Mary and Westeld College; Charterhouse Square; London EC1M 6BQ; U.K. 2 Health and Healthcare; Kings College London; Oak Lodge; David Salomon’s Estate; Broomhill Road; Tunbridge Wells; Kent TN3 OXT; U.K. 3 Department of General Practice and Primary Care; St Bartholomew’s and the Royal London School of Medicine and Dentistry; Queen Mary and Westeld College; Mile End Road; London E1 4NS; U.K. SUMMARY Because of the central role of the general practice in the delivery of British primary care, intervention trials in primary care often use the practice as the unit of randomization. The creation of primary care groups (PCGs) in April 1999 changed the organization of primary care and the commissioning of sec- ondary care services. PCGs will directly aect the organization and delivery of primary, secondary and social care services. The PCG therefore becomes an appropriate target for organizational and educa- tional interventions. Trials testing these interventions should involve randomization by PCG. This paper discusses the sample size required for a trial in primary care assessing the eect of a falls prevention programme among older people. In this trial PCGs will be randomized. The sample size calculations involve estimating intra-PCG correlation in primary outcome: fractured femur rate for those 65 years and over. No data on fractured femur rate were available at PCG level. PCGs are, however, similar in size and often coterminous with local authorities. Therefore, intra-PCG correlation in fractured femur rate was estimated from the intra-local authority correlation calculated from routine data. Three alterna- tive trial designs are considered. In the rst design, PCGs are selected for inclusion in the trial from the total population of England (eight regions). In the second design, PCGs are selected from two regions only. The third design is similar to the second except that PCGs are stratied by region and baseline value of fracture rate. Intracluster correlation is estimated for each of these designs using two methods: an approximation which assumes cluster sizes are equal and an alternative method which takes account of the fact that cluster sizes vary. Estimates of sample size required vary between 26 and 7 PCGs in each intervention group, depending on the trial design and the method used to calculate sample size. Not unexpectedly, stratication by baseline value of the outcome variable decreases the sample size re- quired. In our analyses, geographic restriction of the population to be sampled reduces between-cluster variability in the primary outcome. This leads to an increase in precision. When allowance for variable cluster size is made, the increase in precision is not as great as would be expected with equal cluster sizes. This paper highlights the usefulness of routine data in work of this kind, and establishes one of * Correspondence to: Sandra Eldridge, Department of General Practice and Primary Care, St Bartholomew’s and the Royal London School of Medicine and Dentistry, Queen Mary College, London University, Mile End Road, London E1 4NS, U.K. † E-mail: s.eldridge@mds.qmw.ac.uk Copyright ? 2001 John Wiley & Sons, Ltd.