To minimise the impact of non-response on the survey estimates, auxiliary data for both respondents and groups of non-respondents was used. In particular, data services company Experian matched the sample file against their database of variables derived from sources such as the Electoral Registers. In addition, data was available about the sampled areas, and some survey data was available on partial respondents.

Analysis of the factors that influence the different types of non-response was undertaken and a four stage weighting strategy proposed:

1. Weighting to account for complete non-contact with the sampled addresses. As found in other studies, area-level variables are particularly useful for this but the Experian "Household Type" variable which concentrates on the gender and relationships between household members was also useful in explaining the variation. Response rates here show relatively little variation compared with the other types of non-response.
2. Weighting to account for refusal by the whole household to co-operate with the survey. This involves population density and region but mainly uses two Experian address-level variables: "Life-stage" - another household type variable which concentrates on the age and relationships between household members; and "Mosaic Household Type" which attempts to measure lifestyle. Response rates differed substantially between the various sub-groups.
3. Weighting to account for failure to complete the diaries within responding households. This involves a wide range of Experian variables and household variables from the survey as well as region. The single most important predictor of diary response was the survey's household type variable. Somewhat surprisingly, individual characteristics were not good predictors of this type of response. As with household response, response rates differed substantially in the sub-groups identified.
4. Weighting to population distributions by age, gender and region using a calibration method. This aims to align both the household and individual distributions of the sample on these characteristics with the UK population. Additionally, the sample size varied by month, and there were different sampling rates for weekdays and weekends, therefore, to avoid seasonal bias within the data, the calibration also aligned the sample to flat distributions across the 12 months as well as the seven days of the week.

This page last revised: Thursday 2 September 2003