Databases
Six Dutch GPRNs participated in this research; the Continuous Morbidity Registration Nijmegen (CMR-N), the Academic Network of General Practitioners of the VUmc (ANH-VUmc), the Netherlands Information Network of General Practice (LINH)a, the Registration Network of General Practitioners Associated with Leiden University (RNUH-LEO), the Study of Medical Information and Lifestyle in Eindhoven (SMILE) and the Transition project (Trans). More detailed information of these GPRNs can be found elsewhere [5]. These Dutch GPRNs were selected, because they collect information on all health problems of individual patients. GPRNs which exclusively collect information on chronic, permanent or recurring diseases were left out of this study.
Using the data
We performed an observational study without any intervention. In the Netherlands, no approval is necessary from an ethical committee for analysing data from general practice registration networks. The data are not openly available, permission to use the data is granted by ANH VUmc, RNUH_LEO, SMILE, Transition project, LINH steering committee and the chair of CMR-N.
Selection of diseases
For our analyses we selected twelve health problems: urinary tract infection, gastro-intestinal infection, neck and back problems, eczema, asthma, chronic obstructive pulmonary disease (COPD), osteoarthritis, diabetes mellitus, coronary heart disease (CHD), stroke, depression, and anxiety disorders. The selection of these health problems was based on three criteria: (1) The expected incidence of the specific disorder in the Dutch general practice population was at least 3 per 1000 per year; (2) The total set of diseases represented several ICD chapters (e.g. circulatory system, respiratory system) to obtain a broad spectrum of diseases; (3) The occurrence of incidence and prevalence of included diseases should vary between different patient subgroups (e.g. age, gender).
Incidence and prevalence rates
In this study, we used data of 2007. To determine incidence rates, all patients diagnosed with a new episode of a certain disease between the 1st of January 2007 and the 31st of December 2007 were counted per 1,000 patient years. Prevalence rates were calculated by counting the number of patients with a new or existing episode of a specific disease in 2007 per 1,000 patient years (period prevalence). Incidence rates were calculated for all twelve diseases; prevalence rates were only calculated for the ten chronic or recurring diseases. Five GPRNs record diagnoses according to the International Classification of Primary Care (ICPC), one used the so-called E-list codes [11]-[13]. When necessary, we combined different classification codes to determine morbidity [1],[14]. For example, to measure depression we used ICPC codes P03 and P76.
Socio-demographic characteristics
This study starts with analysing the variation in incidence and prevalence figures between GPRNs and general practices adjusted for patient characteristics: age (in years), gender (male versus female), socio-economic status (high-medium-low), urbanization level (`rural', `urban' and `large cities') and ethnic origin [6]. The latter three measures were determined by proxy using 4-digit postal codes of the patients' home address (the population size is about 4,000 per postal code area) [15],[16].
Practice characteristics
Within a general practice, patients are generally registered with one specific GP, but most patients are not exclusively treated by that GP. The care in general practice has become more multi-practitioner and multi-disciplinary organized [3]. In most networks, the information of an individual patient cannot be related to an individual doctor and therefore, inter-doctor variation cannot be assessed validly. Instead, we analysed GP characteristics on practice level.
The practice characteristics used in the analysis are type of practice (one GP = solo, two GPs = duo and three GPs or more = group practice), percentage female GPs, mean years of working experience, employment of a practice nurse (yes/no), EHR software package used in the general practice, province, distance to the nearest out-of-hours service location and distance to the nearest hospital.
To be sure all practice characteristics are based on the same type of data we consulted the 'Register of General Practitioners' (HAREG) of NIVEL [17]. This database holds information on all practising GPs and practices in the Netherlands about e.g. gender, age, and working experience. We received the information about the employment of a practice nurse and type of electronic patient record directly from the GPRN. The distances have been calculated with the so-called driving time model of Automotive Navigation Data (AND) in combination with the localisations of the out-of-hours service locations and hospitals, using 4-digit postal codes [18].
We are interested in the influence of practice characteristics on the variation in morbidity estimates between GPRNs and practices. Therefore, we only used the practices with all population and practice characteristics available. As a consequence, 9 out of 81 practices of LINH, 2 out of 9 practices of ANH VUmc and 1 out of 9 practices of SMILE and 1 out of 5 practices of Trans were excluded from analyses.
Analyses
Descriptive analyses were applied to give an overview of the distribution of the population and practice characteristics. To explore the variation in morbidity estimates between GPRNs and general practices we used multilevel logistic regression analysis with three levels (patient, practice and network). We used random intercepts on network and practice level to determine the unexplained variation between GPRNs and practices. We analysed the variations in morbidity estimates by calculating the corresponding median odds ratio (MOR) and 95% confidence intervals (95%CI); we also calculated the odds ratios (ORs) of the significant practice characteristics. MOR quantifies the variation between clusters by comparing two `identical' persons from two randomly chosen, but different clusters. MOR expresses the heterogeneity on an odds ratio scale between clusters and represents the median increased risk. Consequently MOR can never be smaller than one. MOR has been calculated on practice and network level. In this study, MOR refers to the (statistical) increased risk of being diagnosed with a certain disease between two randomly chosen practices or GPRNs. For example, if MOR is 2.0 the risk of being diagnosed with a specific disease is twice as high for a person in one network compared to an `identical' person in another network [19],[20].
First, we analysed for each disease the variations in morbidity estimates between general practices and GPRNs without taking any practice characteristic into account. Second, we analysed the influence of six practice characteristics on the variations in morbidity estimates for all diseases in separate models. This results in a total of 154 models (incidence of12 diseases and prevalence of 10 diseases, analysing the variation in one model without any practice characteristics and 6 models with just one practice characteristic (22 × 7 = 154)). Before we performed multilevel analyses, we checked the correlation between characteristics. A high correlation (r >0.70) was found between the urbanization level of the patient's home address and the distance to the nearest hospital of the general practice. We therefore left urbanization level out of the analyses when measuring the effect of distance to the nearest hospital.
The analyses of type of EHR software package and province could not be performed in a three level analysis, as most GPRNs are located in one province and use only one or two types of EHR software package. The influence of these characteristics was only analysed using LINH data in a two level analysis (patient and practice), since this is the only GPRN located in all provinces and including seven different EHR software packages [21]. All analyses were performed with SAS version 9.2.