Reducing number of postcodes that cannot be ascribed would increase validity of methodBMJ 1995; 311 doi: https://doi.org/10.1136/bmj.311.7009.876 (Published 30 September 1995) Cite this as: BMJ 1995;311:876
- Ronan A Lyons,
- Stephen Monaghan,
- Martin Heaven,
- Alan Willson
- Consultant in public health medicine Senior registrar in public health medicine Information scientist in public health medicine Pharmacy adviser West Glamorgan Family Health Services Authority, Swansea SA1 1LT
EDITOR,--Glen Scrivener and David C E F Lloyd conclude that allocating census data to general practice populations is not sufficiently accurate for the purposes of explaining variations in prescribing despite their finding correlation coefficients of up to 0.84 from two of their methods.1 They acknowledge that perfect correlation is impossible in an ecological comparison. Nevertheless, the correlation coefficients might have been higher if the 11710 patients (3.2%) whose postcode could not be matched had been included in the analysis. This proportion of postcodes that could not be ascribed is fairly high. We have developed a link between enumeration districts and general practices in West Glamorgan to study variation in use of hospitals as well as drug prescribing; only 1% of postcodes held on the West Glamorgan Family Health Services Authority's register cannot be matched on the Post Office address file (personal communication, Royal Mail address manager, Post Office). Local inquiry shows that many of the mismatches are due to keying errors when data are input and can be corrected.
Scrivener and Lloyd suggest that time and effort should be spent on collecting data at the level of individual patients. It is axiomatic that the benefits of collecting additional information should outweigh the costs. Unfortunately, unless very large samples are collected all that will happen is that one type of error (that due to the ecological fallacy) will be replaced by that due to the imprecision of estimates based on small sample size: a sample of around 3000 is required to be 95% confident of acquiring an estimate of +/-1% around an expected rate of permanent sickness of about 8%. Apart from the logistic difficulties and cost of collecting such an amount of data in practices, the validity of the data might be questionable if prescribing budgets were to depend on the results.
While the correlation coefficients between census variables and practice populations are not perfect, in the absence of better indicators they are sufficiently high to be used in practice. In our county, estimates of practices' rates of long term limiting illness (based on enumeration district) and the Townsend index of deprivation explain a quarter of the rates of use of hospitals after adjustment for age (unpublished data). If such factors were not included in resource allocation formulas there would be a substantial transfer of resources from the more deprived to the more affluent areas of our county. Whose interest would this serve?