We have detected that you are using Internet Explorer to visit this website. Internet Explorer is now being phased out by Microsoft. As a result, NHS Digital no longer supports any version of Internet Explorer for our web-based products, as it involves considerable extra effort and expense, which cannot be justified from public funds. Some features on this site will not work. You should use a modern browser such as Edge, Chrome, Firefox, or Safari. If you have difficulty installing or accessing a different browser, contact your IT support team.
An Analysis of Mode Effects Using Data from the Health Survey for England 2006 and the Boost Survey for LondonSurvey, Other reports and statistics
- Publication Date:
- 1 Jul 2008
- Geographic Coverage:
- Geographical Granularity:
- Primary Care Trusts
- Date Range:
- 01 Jan 2006 to 31 Dec 2006
As part of its Survey Programme Review, The NHS IC has commissioned NatCen to examine the extent to which results from the Health Survey for England (HSE) 2006 (carried out using face-to-face interviews) and the Boost Survey for London (carried out largely using a self-completion questionnaire) are comparable and to report on the magnitude and direction of any differences.
Despite the differences in mode, the Boost survey was designed to be as comparable with the Core survey as possible. It was intended from the outset that the two achieved samples should be combined during analyses to maximise the sample size within each Primary Care Trust (PCT) in London.
There were two main strands to the comparisons carried out in this report:
1.Response differences between the two surveys were examined in terms of the overall response levels of households and individuals, and the amount of missing data items: where respondents had refused or skipped individual questions.
2.The second strand of the analysis was to investigate the effects of measurement error attributable to the different modes of questioning.
For the majority of characteristics, there were no differences between the Core and Boost samples. The exceptions were ethnicity and educational qualifications.
The difference in the ethnic profile of the two samples, however, was reduced after the non-response weights were applied. The difference in educational qualifications persisted after weighting, and seemed to be attributable to differences in question format between the two surveys.
For many of the key health measures, there were no significant differences between the two surveys. These included estimates for long-term illness, limiting long-term illness, rates of current smoking, whether respondents drank alcohol, and how often they usually drank.
Estimates produced from combined Boost and Core data will not be biased for these variables. However, there were a number of statistically significant differences between the estimates for some key measures including: general health, GHQ12 score, number of cigarettes smoked, number of alcoholic units consumed on the heaviest drinking day, portions of fruit and vegetables consumed, levels of moderate physical activity, and (for women) perceived social support. For these measures, it is difficult to determine the specific causes of the differences as they are likely to be due to a combination of mode and other effects.
Large differences between the two variables imply some degree of bias in one or both of the estimates. For some of these variables, we are on fairly safe ground in making assumptions about which estimate is likely to be the 'better' one, but this is not the case for all the variables.