Data relating to patients changing family practice
The numbers of patients changing practice without change of address between 1st March 2009 and 28th February 2010 (inclusive) were provided by the UK Department of Health. 8,450 family practices in England had at least one patient who changed practice without changing address. Registered practice list size was also provided by the Department of Health for March 2009 and February 2010 for each practice. All analyses were carried out with the practice as the unit of analysis.
Patient experience and overall satisfaction
Measures of patient experience and overall satisfaction were taken from the 2009/10 General Practice Patient Survey (2,169,718 respondents from 8,362 family practices, response rate 39%) . We then excluded practices with a change in list size of more than 10% during the study year, practices with a mean list size less than 1000, and practices that changed postcode during the study period. This was to exclude closing and merging practices, unconventional practices (such as boarding school practices and specialist clinics) and practices where motivations for disenrollment might have been dictated by special circumstances (e.g. practices moving address or using temporary accommodation). We also excluded 248 practices with fewer than 100 responses on the GP Patient Survey to increase the reliability of patient experience scores. A total of 442,731 voluntary disenrollments without change of address appear in the final data set of 7812 practices.
Case-mix adjusted estimates of practice level scores were calculated for items addressing four components of patient experience: i) communication with doctors and nurses (four questions); ii) access to care (three questions); iii) continuity of care (one question); and iv) overall satisfaction with care (one question). In the case of two communication questions with multiple items, a composite score was calculated as the mean of up to seven sub-items for those patients who answered at least four of the seven. The case-mix adjusted scores were calculated from mixed-effect linear regression models adjusting for patient reported gender, age, ethnicity, deprivation and self-rated health and predict the scores a practice would have received if its case-mix was the same as that of all responders. Full details of the survey and its development have been reported elsewhere [12–14].
Practice and doctor characteristics
For each practice, registered patient numbers (broken down by sex and age group) were provided by the NHS Information Centre. In addition, data for 2009 for each practice were provided to calculate the number of family practitioners (FP) excluding trainees, the number of patients per full time equivalent FP, the mean number of years since qualification of the FPs in each practice, the proportion of male FPs, and the proportion of FPs trained in the UK for their primary medical qualification. A score for socio-economic deprivation for each practice was calculated by applying the 2007 Lower Super Output Area Index of Multiple Deprivation proportionately to the practice population . GP Patient Survey results were also used to estimate the proportion of Black, Asian, Chinese, mixed race and other non-white patients in each practice.
Availability of nearby family practices
The geographic location of each practice was determined by matching the practice postcode to data obtained from UK Borders . Using these data, the UK Ordnance Survey national grid reference of the centroid of each of the practice postcodes was obtained. These grid references were used to calculate the number of other practices within 1km of each practice and the number of practices in the same location (co-located practices are interpreted as separate practices operating within a single health centre or building; we assumed this to be the case where the postcode was common to different practices).
Practice-level scores calculated directly from the GP Patient Survey contain measurement error. Using simple scores in regression models may underestimate true effect sizes, and we therefore used shrunken case-mix adjusted estimates of GP Patient Survey scores. This reduced the effects of measurement error and improves the accuracy of estimated effect sizes [17–20]. The annual rate of patients changing practice without change of address was modelled using Poisson regression with a random effect for practice included to account for any remaining variation in disenrollment rates.
In order to compare the strength of association between disenrollment rates and the measures of patient experience, satisfaction, and other patient and practice characteristics, continuous variables (including proportions) were normalised such that the difference between the 5th and 95th percentile of practices was equal to one  for patient experience measures, and equal to minus one (−1) for all others. This did not affect the continuous nature of the data, but allowed for direct comparison of the magnitude of effect across different variables. The resulting rate ratios should be interpreted as the relative increase in disenrollment rate associated with moving from the 95th to 5th percentile for patient experience measures, and moving from the 5th to the 95th percentile for other continuous variables. Rate ratios of more than one indicate an association with higher disenrollment rates. Where ordered categorical variables have been used (number of doctors in a practice, number of practices within 1km) the reference category was chosen such that it contained either the 5th or 95th percentile. This allowed the largest rate ratios seen for the categorical variables to be compared directly to the rate ratios for the continuous variables and allowed low scoring practices to be compared to high scoring practices (5th vs 95th percentiles).
Models were run in three stages. First, a series of models including a single fixed effect and the random practice effect were run using single measures of patient experience, patient satisfaction, and practice or doctor characteristics. These models (model set 1) were used to assess the crude association between voluntary disenrollment and each of the factors separately. This shows, for example, how much higher the disenrollment rate typically was in a practice on the 5th centile of overall satisfaction compared to a practice on the 95th centile.
A second model included fixed effects for all measures of patient experience, patient satisfaction, practice case-mix, and doctor characteristics, with a random effect to control for practice (model 2). Model 2 was used to assess the same associations as model set 1, but when adjusting for all other variables; for example, to show how much the disenrollment rate typically increased beyond the variations associated with other factors when comparing practices on the 5th and 95th centiles of overall satisfaction.
A final regression model (model 3) was constructed to assess the suitability of using voluntary disenrollment rate as a quality indicator. This model augmented model 2 with log(capitation), log2(capitation) and co-located surgery included in order to explain as much variation as possible. Model 3 was used to assess the variation in disenrollment rates that was not explained by measured characteristics (the random effect).
All analyses were performed in using Stata v11.2 (StataCorp, College Station, Texas, USA) and SAS v9.2 (SAS Institute, Cary, NC, USA).