Search

Rehm et al. highlight the ongoing difficulties in accurately estimating alcohol consumption using surveys. Population surveys, in particular, suffer from non-response and sampling bias, which affects their representativeness, but they are one of the few ways of estimating differences in consumption across population subgroups. In this article, we highlight different approaches that have been taken in Scotland to try to overcome these problems, from the pragmatic to the sophisticated.

Background and aims: Analytical approaches to addressing survey non-participation bias typically use only demographic information to improve estimates. We applied a novel methodology which uses health information from data linkage to adjust for non-representativeness. We illustrate the method by presenting adjusted alcohol consumption estimates for Scotland. Design: Data on consenting respondents to the Scottish Health Surveys (SHeSs) 1995-2010 were linked confidentially to routinely collected hospital admission and mortality records. Synthetic observations representing non-respondents were created using general population data. Multiple imputation was performed to compute adjusted alcohol estimates given a range of assumptions about the missing data. Adjusted estimates of mean weekly consumption were additionally calibrated to per-capita alcohol sales data. Setting: Scotland. Participants: 13 936 male and 18 021 female respondents to the SHeSs 1995-2010, aged 20-64years. Measurements: Weekly alcohol consumption, non-, binge- and problem-drinking. Findings: Initial adjustment for non-response resulted in estimates of mean weekly consumption that were elevated by up to 17.8% [26.5units (18.6-34.4)] compared with corrections based solely on socio-demographic data [22.5 (17.7-27.3)]; other drinking behaviour estimates were little changed. Under more extreme assumptions the overall difference was up to 53%, and calibrating to sales estimates resulted in up to 88% difference. Increases were especially pronounced among males in deprived areas. Conclusions: The use of routinely collected health data to reduce bias arising from survey non-response resulted in higher alcohol consumption estimates among working-age males in Scotland, with less impact for females. This new method of bias reduction can be generalized to other surveys to improve estimates of alternative harmful behaviours.

Aim Response rates in health surveys have diminished over the last two decades, making it difficult to obtain reliable information on health and health‐related risk factors in different population groups. This study compared cause‐specific mortality and morbidity among survey respondents and different types of non‐respondents to estimate alcohol‐, drug‐ and smoking‐related mortality and morbidity among non‐respondents.

Human immunodeficiency virus (HIV) ribo-nucleic acid (RNA) viral load is a measure of actively replicating virus and is used as a marker of disease progression. For a thorough understanding of the dynamics of the evolution of the virus in the early life of HIV-1 vertically infected children, it is important to elucidate the pattern of HIV-RNA viral load over age. An aspect of assay systems used in the quantification of RNA viral load is that they measure values above particular cut-off values for detection, below which the assays used are not sufficiently sensitive. In this way, measurements are potentially left-censored. Recent adult studies suggest that to adequately model RNA pattern over age, it is necessary to account for within-subject correlation, due to repeated measures, and censoring. The aim of this study, therefore, was to establish whether it is necessary to use complex methods to allow for repeated measures within individuals and censoring of the HIV-RNA viral load in children enrolled in a cohort study. The approach involved the identification of an appropriate model for the basic pattern of RNA viral load by age and subsequent assessment of various estimation procedures accounting for repeated measures and censoring in different ways. Methods developed by Hughes involving the expectation-maximization (EM) algorithm and the Gibbs sampler were taken as the benchmark for comparison of simpler alternatives. Other approaches considered involve linear mixed-effects and ordinary least squares in which censoring is dealt with informally by taking the cut-off value as absolute or taking the mid-point between cut-off and zero. Fractional polynomials provided a substantially superior approach for modelling the dynamics of viral load over age compared to conventional polynomials or change-point models. Allowing for repeated measures was necessary to improve the power of the likelihood ratio tests required to establish the final model, but methods beyond taking the mid-point for censored values did not further improve the fit. Although Hughes' methodology is the best approach, its implementation is not necessary for the identification of the optimal model.

We thank Fergusson and Boden for their interest in our paper as detailed in their commentary (1). They point to some interesting issues which, for the main part, have also been debated during the preparation of the paper. The first issue raised is that with the large sample size, the likelihood of finding significant group differences is high. We acknowledge that this should be considered, but as the hazard ratios in the present study are relative large and confidence limits in most cases relative narrow we consider the large sample size to be more a strength than a drawback in the present study. Ferguson and Boden also points to the fact that the low baseline rates of morbidity and mortality will generate relative large hazard ratios even with relative small differences in the absolute number of events between respondents and non-respondents. We acknowledge that looking only at relative differences can be somewhat misleading and we have therefore also provided the absolute number of events and the rates for each group in our paper (2). The most careful way of interpreting results is often to look at both relative and absolute differences. However, as pointed out by Ferguson and Boden, the use of respondent-only data can cause biased estimates even when the absolute difference in the number of events among respondents and non-respondents is small. It is suggested that the future sample sizes could be reduced and the saved cost should be used on contacting the non-contacts. This is an interesting reflection. Consideration of strategies to raise response rate among specific groups of non-respondents is indeed warranted, and different strategies to improve response rate among different types on non-response groups are presented in our paper (2). However, the present study used pooled data from two health surveys. Hence, the sample size in each of the surveys is not as large as it might appear in the commentary by Fergusson and Boden. Both surveys are furthermore designed to provide county and regional representative data, respectively, and hence, a minimum sample size is required in each county/region. Lastly, the number of non-contacts is very small in the present study thanks to a notable effort to establish contact to all invited individuals, and we think that it would be very difficult to establish contact to all invited individuals even if the resources are used differently. However, their suggestion will be considered when planning future surveys. Lastly, Ferguson and Boden argue that non-response bias may have less of an impact when examining exposure-outcome associations compared to studies of prevalence estimates. This is often true (3, 4, 5). There are, however, exceptions and thus the impact of non-response bias in studies of associations cannot be assumed to be negligible (6). In summary, Ferguson and Boden raise some interesting issues, which we agree need to be taken into account when both analyzing and interpreting results on non-response in surveys while the impact of survey non-participation should not be overlooked.