Addressing non-response in health survey data to refine alcohol consumption estimates in Scotland

In terms of health, social and economic costs, alcohol is likely to be the most harmful substance present in society, and is closely linked to mental health problems such as depression, anxiety, affective conduct disorders and suicides. Interventions to tackle alcohol harm require good-quality evaluation, the accuracy of which relies on robust consumption estimates over time. However, the best indicator of consumption - alcohol sales per capita - is only available at an aggregate level. Alternative measures can be obtained from sources such as national health surveys which are designed to make major contributions to the monitoring of population health. Mental health and the factors that impact on it - among which alcohol consumption is of great importance - are significant components of health surveys. Data obtained are used to formulate and evaluate public health policies aimed at improving and maintaining mental wellbeing, and identify groups (such as deprived communities) at particular risk of ill-health. However, the validity and usefulness of such data rely on the extent to which the samples surveyed are representative of the general population. In UK settings, response to health surveys has declined over the last decade or so, and there is growing concern that data collected are increasingly non-representative of the general population. Currently, the sole attempt to address bias arising from such non-response uses the corrective procedure of inverse probability weighting (IPW) of respondent data based on limited socio-demographic characteristics, and does not capture or adjust for health aspects, such as those related to alcohol, or allow for differences between responders and non-responders in these aspects. The Scottish Health Survey (SHeS) is based on the Scottish population living in private households. Additionally, there are prospective data available on consenting SHeS participants (92%) from record-linkage to routine hospital admission (SMR) and mortality (GROS) data. Such linkage is rare and provides cost-effective good-quality longitudinal data. SHeS-SMR/GROS are available for the surveys conducted in 1995, 1998 and 2003 (total adult sample size of 23,093), with a maximum follow-up of 15 years. Also available in Scotland we have (unlinked) census information and SMR/GROS data for the general population. The aim of our proposal is to explore and address non-response bias in the health surveys, with a specific focus on alcohol consumption. This will be achieved by exploiting the existing information available from SHeS-SMR/GROS, and census, SMR and GROS data in the general population. Comparison of alcohol-related mortality and hospital admission rates among the health survey respondents with those in the general population provides a means of identifying differential outcomes - and hence underlying differences in the composition of the groups in terms of behaviours such as alcohol consumption, as well as demographic and socioeconomic factors. Once the extent of such differences has been identified, augmentation of survey data with information on the differences can be achieved through the application of particular statistical methodology, such as advanced multiple imputation and IPW. We shall look separately at each round of the survey to see how the decline in non-response affects the estimates. Small area deprivation data will be sought to obtain more accurate disaggregated estimates of alcohol consumption than are currently derived from health surveys. Additionally, data may be combined with sales data. The resulting corrected estimates of alcohol consumption in the population will enable superior testing of the effectiveness of interventions such as the introduction of minimum unit pricing and more finely honed services. An advanced weighting procedure will be developed which may be applied to other surveys.

We propose to examine how representative the respondents to the 1995, 1998 and 2003 SHeSs are in terms of alcohol-related hospitalizations (SMR) and deaths (GROS) to inform the improvement of survey-based estimates of alcohol consumption. Linked SHeS-SMR/GROS data (to 2010; maximum follow-up 15 years) will be compared with 1991 and 2001 census records, mortality and hospital admission data for the entire population. We shall: 1. Compare SHeS-SMR/GROS and general population rates of alcohol-related hospitalizations and deaths; 2. Determine the probability of alcohol-related hospitalization/death in the survey non-responders; 3. Calculate the number of non-respondents for each socio-demographic category (age, sex, area deprivation and health region); 4. Generate observations for non-responders with the corresponding alcohol-related hospitalization and death probabilities in each combination group; 5. Use multiply imputation* - the recommended means of substituting values for missing data items - for unknown alcohol consumption estimates under the assumption that the consumption data are 'missing at random' (MAR)* given socio-demographic category and alcohol-related hospitalization/death; 6. Modify the alcohol consumption imputations to represent likely differences in association between alcohol consumption and alcohol-related hospitalizations and deaths between observed and missing consumption data using a pattern mixture model approach which assumes that the consumption data are 'missing not at random' (MNAR)* given the observed data. We shall look separately at each survey round to assess the effects of non-response level. An advanced correction procedure will be developed for application to other surveys with record-linkage capacity. Further work may include integration of national sales data to further refine consumption estimates. * For technical detail, see the Statistical methodology section of 3.3 in the Case For Support document.

National public sector
At the national public health sector level, the Scottish Government, its Alcohol Evidence Group and NHS Health Scotland's will all benefit from the proposed work in relation to ultimate improvements in mental health and wellbeing. With shared investment in Scotland's alcohol strategy and its evaluation by the MESAS work stream of NHS Health Scotland, they have mutual interest in ensuring that the development and implementation of policy on alcohol-related problems is supported by evidence-based research. The evaluation consortium seeks and will benefit from enhanced knowledge, research capacity and efficiency with the improved estimates of alcohol consumption, and prevalence of harmful drinking and dependency we are offering. In this way, our proposed work will influence the development of public health policy and legislation such as minimum alcohol unit pricing. The corrected population alcohol consumption estimates will also facilitate more finely honed services. Additionally, the correction procedure methodology we offer will inform the Scottish Government planning and reporting of future surveys.

Regional and local public sector
Regionally, more local estimates of alcohol consumption and dependency will benefit the MESAS team with evaluation and NHS health boards and local authorities with service planning, thus informing the practice of health professionals.

Third sector
With a strategic role in tackling associated problems, Alcohol & Drug Partnerships respond to the needs of local areas, based on evidence and outcomes - one of which concerns reducing the prevalence of harmful drinking. By providing appropriately refined small area alcohol consumption estimates, we shall facilitate enhance knowledge, research capacity and efficiency of Alcohol & Drug Partnerships and other voluntary sector organisations.

General public
Our proposed work will facilitate the evaluation of Scotland's alcohol strategy - with the ultimate aim of reducing alcohol-related health and social harms - leading to national improvements in cultural vibrancy, quality of life, and general health and well-being. Specifically, conditions which are likely to be reduced include depression, anxiety, affective conduct disorders and suicides. If successful, gains for the general public include decreases in the alcohol-related economic burden and reductions in alcohol-related crime and antisocial behaviour. Such gains will increase public awareness of the economic and societal benefits of alcohol consumption reduction and the contribution of science.

International influence
With direct relevance to alcohol strategy policy, the proposed work and the resultant increased knowledge have wider implications for public health sectors of other countries of the UK, Western Europe and beyond. In settings where there is interest in alcohol-related harm reduction, the implementation of legislation such as minimum alcohol unit pricing could be considered.

Surveys are widely used by government departments and these improvements in methodology will be of interest to a wide international audience helping government departments and other researchers to better address bias in a variety of settings. The adoption of our proposed methodology in the estimation of alcohol consumption from surveys in England - and other countries with such linkage - would represent operational and organisational change at the wider international level.