Sociodemographic variables for general practices: use of census dataBMJ 1995; 310 doi: https://doi.org/10.1136/bmj.310.6991.1373 (Published 27 May 1995) Cite this as: BMJ 1995;310:1373
- F Azeem Majeed, lecturer in public health medicinea,
- Derek G Cook, senior lecturer in epidemiologya,
- Jan Poloniecki, lecturer in medical statisticsa,
- Jo Griffiths, statisticiana,
- Caroline Stones, information services managerb
- a Department of Public Health Sciences, St George's Hospital Medical School, London SW17 0RE
- b Merton, Sutton and Wandsworth Family Health Services Authority, London SW15 2SW
- Correspondence to: Dr Majeed.
- Accepted 17 February 1995
Measures of the social, ethnic, and demographic characteristics of general practice populations are essential both for planning health services and for research. Such measures can be derived by combining census data for electoral wards with postcoded data from a family health services authority's age-sex register.1 Because electoral wards have large populations (typically 5000-15000 people), the patients registered with individual general practices will rarely be representative of a ward, and this will reduce the accuracy of census derived variables. Enumeration districts, however, have smaller populations (typically 200-600 people). We (a) derived estimates of age structure of practice populations based on enumeration districts and electoral wards and compared them with those obtained from the authority's age-sex registry and (b) compared the predictive power of variables derived from enumeration districts and electoral wards in explaining the variation in breast cancer screening rates among these practices.
Patients, methods, and results
The family health services authority provided a database containing the age, general practice, and postcode of all 601330 patients living in Merton, Sutton, and Wandsworth and registered with 131 general practices. An enumeration district was assigned to each postcode by means of a table.2 Of the 11572 postcodes on the age-sex register, 166 could not be assigned to an enumeration district because they were not listed in the table. The 1149 patients living at these postcodes were excluded from the calculation of the census derived variables, as were the 1543 people living in importing and special enumeration districts (importing enumeration districts contain census data on people living in other enumeration districts, and census data for most special enumeration districts are suppressed). When postcodes lay in only one enumeration district (81.7% (9460) of postcodes) the value of the relevant census variable for the enumeration district was taken as the value for the postcode. When postcodes lay in more than one enumeration district (16.8% (1946) of postcodes) a weighted average of census data for the enumeration districts was taken as the value for the postcode, with the weights being the number of households in each part postcode unit.2 For each census variable we assigned a value to each patient based on their postcode. For each general practice we then calculated the average of these assigned values for the registered patients. Apart from the need to use the table, the method is similar to that described previously for electoral wards.1
The correlations between the observed and expected age structure of the 131 practices were higher in each age group when enumeration district data were used to calculate the expected age variables (table); in all but one case the improvements were significant. The interquartile ranges of the differences between the observed and expected age structure of the practices were also smaller when enumeration district data were used to calculate the expected values (table).
To compare the predictive power of variables derived from enumeration districts and electoral wards we performed regression analysis of breast cancer screening rates on the eight Jarman variables and the census derived estimates of list inflation. Variables derived from enumeration districts gave a significantly better fit than those from electoral wards (P=0.002 by the method of Royston and Thompson4), explaining 58% compared with 53% of the variation in screening rates between practices.
Census derived variables for general practices have many potential uses—for example, they could be used to set targets for cervical cancer screening according to the populations that practices serve. We have shown that postcoded data in family health services authority age-sex registers can be used with census data to produce such variables for general practices. Despite the inaccuracies associated with small area census data,5 variables derived from enumeration districts were more accurate than those derived from electoral wards in measuring the age structure of practice populations and explaining the variation in breast cancer screening rates. Census derived variables could also be produced for schools and hospitals, allowing league tables comparing their performance to be adjusted for the social make up of their catchment populations.
We thank Dr David Martin, lecturer in geography at the University of Southampton, and Dr Patrick Royston, reader in medical statistics, Royal Postgraduate Medical School, for their help. We also thank the University of Manchester Computer Centre for providing access to census data and to the postcode enumeration district table, and the South West London breast screening service for permission to use its data.