Open access Protocol BMJ Open: first published as 10.1136/bmjopen-2018-026187 on 4 April 2019. Downloaded from Validation of non-participation bias methodology based on record-linked Finnish register-based health survey data: a protocol paper Megan A McMinn, 1 Pekka Martikainen,2 Emma Gorman,3 Harri Rissanen,4 Tommi Härkänen,4 Hanna Tolonen,4 Alastair H Leyland,1 Linsay Gray1 To cite: McMinn MA, ABSTRACT Strengths and limitations of this study Martikainen P, Gorman E, et al. Introduction Decreasing participation levels in health Validation of non-participation surveys pose a threat to the validity of estimates ► This study will validate a dedicated methodolo- bias methodology based on intended to be representative of their target population. record-linked Finnish register- gy that aims to adjust for non-participation bias in If participants and non-participants differ systematically, based health survey data: a health surveys through the use of record linkage. the results may be biased. The application of traditional protocol paper. BMJ Open ► We use an individual-level dataset on the entire se- non-response adjustment methods, such as weighting, can 2019;9:e026187. doi:10.1136/ lected sample for a Finnish national health survey, fail to correct for such biases, as estimates are typically bmjopen-2018-026187 from which the characteristics of non-participants based on the sociodemographic information available. can be identified, with linkage to morbidity and mor- ► Prepublication history and Therefore, a dedicated methodology to infer on non- additional material for this tality records, providing the ‘gold standard’ for the participants offers advancement by employing survey data paper are available online. To methodology validation process. linked to administrative health records, with reference view these files, please visit ► Previous applications of this methodology have been to data on the general population. We aim to validate the journal online (http:// dx. doi. able to use data on the total population for compar- such a methodology in a register-based setting, where org/ 10. 1136/ bmjopen- 2018- ison. This study is limited to a population sample individual-level data on participants and non-participants 026187). available for this analysis. are available, taking alcohol consumption estimation as ► The estimated gradient in the risk of alcohol-related Received 21 August 2018 the exemplar focus. Revised 4 March 2019 harms may be stronger using individual measures of Methods and analysis We made use of the selected http://bmjopen.bmj.com/ socioeconomic position than area-level measures of Accepted 8 March 2019 sample of the Health 2000 survey conducted in deprivation; therefore, these reference comparisons Finland and a separate register-based sample of the may not mirror the methodology based on less infor- contemporaneous population, with follow-up until 2012. mative area-based measures. Finland has nationally representative administrative and ► This validation exercise is confined to assessing the health registers available for individual-level record linkage reliability of inferring on non-participants from com- to the Health 2000 survey participants and invited non- parisons of the participants and the reference popu- participants, and the population sample. By comparing lation; other aspects of the methodology, such as the the population sample and the participants, synthetic extent to which alcohol-related hospitalisations and on September 29, 2021 by guest. Protected copyright. observations representing the non-participants may be © Author(s) (or their deaths provide sufficient information to impute un- generated, as per the developed methodology. We can employer(s)) 2019. Re-use known alcohol consumption estimates, are beyond permitted under CC BY. compare the distribution of the synthetic non-participants the scope of this study. Published by BMJ. with the true distribution from the register data. Multiple 1MRC/CSO Social and Public imputation was then used to estimate alcohol consumption Health Sciences Unit, University based on both the actual and synthetic data for non- of Glasgow, Glasgow, UK participants, and the estimates can be compared to 2 such as smoking prevalence, levels of physical Population Research Unit, evaluate the methodology’s performance. activity and alcohol consumption for entire Faculty of Social Science, Ethics and dissemination Ethical approval and access to University of Helsinki, Helsinki, populations, not confined to the subpop- the Health 2000 survey data and data from administrative Finland ulation in contact with health services. 3 and health registers have been given by the Health 2000 Department of Economics, However, the decreasing levels of participa- Lancaster University, Lancaster, Scientific Advisory Board, Statistics Finland and the National Institute for Health and Welfare. The outputs will tion in these surveys threaten their ability UK 1–3 4Department of Public Health include two publications in public health and statistical to provide reliable estimates. The propor- Solutions, National Institute for methodology journals and conference presentations. tions of non-participation are typically not Health and Welfare, Helsinki, uniform across sociodemographic groups, Finland meaning that selected groups, such as men Correspondence to INTRODUCTIon or those from deprived backgrounds, are 4 Dr Megan A McMinn; Health surveys enable the production of esti- often under-represented in health surveys. megan. mcminn@ glasgow. ac. uk mates of various health-related behaviours, Non-participation has also been found to McMinn MA, et al. BMJ Open 2019;9:e026187. doi:10.1136/bmjopen-2018-026187 1 Open access BMJ Open: first published as 10.1136/bmjopen-2018-026187 on 4 April 2019. Downloaded from correlate with higher rates of morbidity and mortality5 6; how extreme the differences in sex-specific mean weekly in particular, substantially lower rates of alcohol-related consumption between participants and non-participants harms (deaths and hospitalisations) have been found were assumed to be, with little impact on estimates for among participants, compared with the general popu- women.12 lation.7 Where it is possible to identify non-participants, This project aims to validate the methodology developed findings of higher harm rates among the non-participants for addressing non-participation bias. More specifically, to relative to the participants have been reported.8 9 A set evaluate whether it is valid to infer on the non-participants of health studies conducted in Finland found that deaths from comparisons of the participants and a total regis- due to alcohol-related diseases, injuries and poisonings ter-based population sample without non-response. Valida- had the largest relative mortality differences between tion requires a setting whereby some true information on participants and non-participants for men and were the individual non-participants of a health survey is known, second largest for women, exceeded only by deaths due and these can be compared with the synthetic observa- to suicides.9 In Denmark, non-participants were found to tions generated by our methodology. Finland provides have significantly increased hazard ratios for alcohol-re- this opportunity as it maintains a nationally representative lated hospitalisations and deaths relative to participants.8 register that forms the sampling frame for surveys and Under such circumstances, there is bias present in the has the ability to interlink sociodemographic information, participant sample and, as a consequence, in the derived morbidity and mortality databases, and survey responses at estimates of alcohol consumption. Attempts to correct for the individual level using personal identification codes.13 such non-participation bias typically make use of weights Therefore, through the use of this register, the sociodemo- based on sociodemographic characteristics10; however, graphic, hospitalisation and death categories of the true this may not fully capture health differences. The success non-participants are known (providing the ‘gold standard’). of the weighting is dependent on the extent to which With the addition of the general population data, we are those participating are representative of their subgroups able to make indirect inference using the synthetic observa- of the population. For instance, individuals in harder-to- tions. We can then compare the results of the synthetic and reach subgroups, such as younger men from disadvan- true non-participants, allowing us to assess the validity of our taged backgrounds, that do participate, are unlikely to existing methodology. be representative of their entire demographic, and so weighting does not resolve the bias.11 We have developed11 and applied6 12 a dedicated meth- METHODS AND ANALYSIS odology that uses additional health information from Health 2000 survey data data linkage and reference to population data to adjust The Health 2000 Survey (thl. fi/ health2000) is a nation- for non-participation bias. This methodology has previ- ally representative health examination survey conducted http://bmjopen.bmj.com/ ously been used to improve estimates of population-level in Finland between 2000 and 2001. A regional two-stage alcohol consumption, although it could be applied stratified cluster sampling strategy was used to iden- to other health-related behaviours of interest, such as tify approximately 8000 persons aged 30 and over, in tobacco smoking. the main survey.14