GP Access survey
The survey was administered by Ipsos MORI, on behalf of the Department of Health. Patient information was obtained using Primary Care Trust (PCT) registration records from the National Health Application and Infrastructure Services (NHAIS) database. The main outcomes of interest were the survey items that asked respondents about their experience and satisfaction with access to their general practice. Satisfaction and experience with access are two theoretically different aspects of care [11]. The questionnaire included two items relating to satisfaction with aspects of access (Q2, satisfaction with how easy it is to get through to someone on the phone at the surgery; and Q9, satisfaction with the hours the surgery was open) and three items relating to experience with aspects of access (Q4, experience with getting an appointment on the same day or on the next 2 days the surgery was open; Q6, experience with making an appointment with the surgery more than 2 full days in advance; and Q8, experience with making an appointment with a particular doctor at surgery even if it meant waiting for longer - all restricted to patients who had tried in the preceding six months). All five items were dichotomous and answered with a 'Yes' or 'No'. The five items were agreed between the British Medical Association and NHS employers. Ipsos MORI undertook a series of face to face cognitive tests to examine if the questions were clear and easily understood. Some limitations were identified and the questionnaire was redesigned after each round of testing. After the first year of the survey, some changes were made to the questionnaire which did not compromise its consistency [12].
Information on patient characteristics was also collected in the questionnaire: age, gender, number of appointments in the last year, whether the patient was a parent or legal guardian, employment status, travel time from home to work, typical working hours, ability to take time away from work to visit GP surgery, limiting long-standing conditions, difficulty performing day-to-day activities because of limiting long-standing conditions, carer responsibilities and ethnicity. Measures of area deprivation and rurality were assigned to patients based on the Lower Super Output Area in which each resided. The full 2008 questionnaire is presented in Additional file 1.
In 2007/08, 8,307 out of 8,403 practices in England were included in the survey (reasons for exclusion included having fewer than 50 eligible patients). Patients were eligible to be selected for participation if they were aged 18 or over, with a valid NHS number and registered with the same practice continuously from the 1 July 2007 to the date of the sample extraction from the NHAIS on 18/19 November 2007. Patients were randomly sampled from each participating practice, with more patients selected in practices likely to have lower rates of response. The sample size for each practice was determined by the number of returned questionnaires likely to deliver a confidence interval of ± 7 percentage points, at the 95% level, for items Q2, Q3 and Q4 [12]. The questionnaires and a cover letter were posted in the week commencing on the 7th of January 2008, with two reminders sent out in February and March, while the closing date for completed surveys was the 2nd of April 2008. Overall 4,922,080 questionnaires were posted, with no more than 930 issued for any practice. Telephone help lines in 10 languages in addition to English were available for individuals who were unable to complete the questionnaire without additional assistance. The overall response rate was 40.6%, with 1,999,523 completed responses collected [13]. More details on the development and organisation of the survey can be obtained from the technical report published by Ipsos MORI [12]. The dataset is not publicly available.
For this study, we obtained information about practice and PCT characteristics from other additional sources: the General Medical Services (GMS) database 2006; Super Output Area Indices of Multiple Deprivation 2004; and the Quality and Outcomes Framework (QOF) results for 2006/7. Practice level variables were: practice list size, full time equivalent GPs, ratio of full time equivalent GPs per 10,000 patients, overall reported achievement on 48 'stable' QOF indicators (i.e. introduced in 2004/05 and with minor or no changes in the first 5 years of the scheme), distance to nearest practice, emergency admissions per 1000 patients, standardised mortality ratios of people under 65, number of new registrations, total opening hours and extended opening hours. Measures of global practice population deprivation and rurality were created by aggregating scores across the patients in each practice sample. We constructed practice population measures of ethnic mix, percentage of people in full-time employment and age profile, using both the practice samples and the 2001 census. Both estimates are prone to error (those from the sample due to self-selection bias; those from the Census due to changes since 2001), however, the two measures correlated well for ethnic mix (White v non-White; r = 0.856) moderately well for age (mean age, r = 0.614), but less well for rates of full-time employment (r = 0.537). In the analysis we used to estimates from the Census. Regional information was limited to three variables: Strategic Health Authority, number of practice staff in the PCT per 100,000 population, and walk-in centre attendance in the PCT per 100,000 population (walk in centres existed in 49 of the 152 PCTs) and had been established specifically to improve access to primary care.
Statistical Analyses
We used multilevel multivariate regression to investigate relationships between each dimension of satisfaction/experience and patient, practice and regional characteristics. The outcome variables were all binary (e.g. able/unable to get an urgent appointment), therefore we utilised logistic regression. We began with univariate analyses, examining each predictor separately, and followed these up with a multivariate analysis to control for relationships between predictors. We included the patient, practice and regional level predictor variables in the same multi-level analysis. The size of the dataset made it not feasible to model the full hierarchical nature of the data (respondents nested within practices nested within regions), therefore we adopted a two-level model that took account of the nesting of respondents within practices, and assigned the regional variables to the individual practices. Although this may have introduced some small error into the p-values for some predictors, p-values have not been used to gauge the importance of each predictor.
The size of the sample was such that very small differences in scores were statistically significant, making significance alone a poor guide to the effect of each predictor. Therefore to assess strength of effect we used an approach based on the odds-ratio coefficient for each predictor variable. First, we rescaled each continuous predictor variable by subtracting the mean and dividing by twice the variable's standard deviation. In analysis, these rescaled variables then yield odds-ratios comparable to those obtained comparing one level of a categorical variable with another [14]. Second, we defined an important predictor to be one with a calculated odds-ratio of 1.18 or above, or 0.86 or below. These values correspond to an increase/decrease in the satisfaction/experience score of 2.5% or more, from a baseline of 80% (the average across all five domains of satisfaction/experience).
We conducted one multivariate analysis on the full sample of patients, and a second using only those patients in full- or part-time employment. Three variables that were only applicable to patients in employment (travel-time from home to work, typical working hours and ability to take time away from work to visit GP surgery) were included in the second analysis only.
We excluded patients with any missing data, since the sample size was large enough to allow us to avoid less robust approaches. We examined the set of independent variables for multi-collinearity prior to analyses and removed those with a variance inflation factor greater than four [15]. One variable, full time equivalent GPs was removed due to multi-collinearity: this was highly correlated with practice list size and the latter was a stronger predictor in the univariate analyses. Total practice opening hours and extended opening were only available for around half of the practices (53%) and were not included in the presented regressions. However, we repeated the analyses including these variables and using only those practices for which we had data; their effect on the access items was not found to be important. All analyses were undertaken using STATA version 10.1 [16].