Assessing non-response bias in Finnish register-based health survey data
Reliable estimates of the population prevalence of health-related behaviours, such as those for alcohol consumption, are critical to public health for multiple reasons. These include formulating and evaluating policies aimed at improving and maintaining population health and wellbeing. Surveys, among other sources, provide prevalence estimates of such health-related measures. However, the reliability of inference from survey data is dependent on how representative they are of the general population. This can be threatened by compromised survey participation levels, which have been declining in recent years in Europe as well as elsewhere. This decline in unit response is widely recognised as a considerable and growing problem.
We aim to explore the validity of dedicated methodology we have developed, which uses record-linked survey data with reference to data on the general population to infer on non-participants. We compare this to the approach based on record-linked data on non-participants, as well as participants which are available from the Finnish Health 2000 survey data. The focus is on refining the measurement of alcohol consumption but the methodology has wider applicability.
We have recently developed and applied methods that utilise record linkage of survey participants to hospitalisation and mortality files, comparing alcohol-related harms as well as sociodemographics to those in the general population to allow inference on non-participants. We then used this inference as the basis of generated partial observations for non-participants with the corresponding death rates in demographic sub-groups and then used multiple imputation to reliably fill-in their “missing” alcohol measurements. Rather than simply incorporate mortality into survey weights, we instead choose the multiple imputation-based approach since it has the flexibility to accommodate differences between participants and non-participants within groups defined by sociodemographics and mortality. This has been performed in settings in which information on individual non-participants is unknown.
In the absence of direct information on the non-participants, we have had to make indirect inference on the non-participants by comparing the participants and the general population. With the register-based systems in Finland, non-response bias can be assessed in direct comparison of participants and non-participants; in this setting the sociodemographic characteristics of the non-participants are already known at the individual level. Of course, the ability to also make the indirect comparisons with the general population, as we have thus far, holds in Finland too. There is therefore the opportunity to compare the results of both approaches: having access to the linked hospitalisation and death records as well as socio-demographic information for non-participants (providing the “gold standard”), in addition to the general population data, allows us to assess the validity of our existing approach.
We base this on the Health 2000 survey with follow-up through to the end of 2013. For both participants and non-participants we have: age-group, sex, socioeconomic measures, deaths due to alcohol (date of death; ICD codes), hospitalisations due to alcohol (date of event; ICD codes). Additionally, for the participants we have: alcohol consumption, measures of alcohol-related dysfunction (eg alcohol dependence); these data have all been acquired. Reference will be made to the following information on the contemporaneous Finnish general population which we are in the process of acquiring: population count in 2000 aggregated by gender, age-group, socioeconomic measures; counts of subsequent alcohol-related deaths and hospitalisations through to the end of 2013 aggregated by gender, age-group and socioeconomic measures.
Pekka Martikainen, University of Helsinki
Tommi Härkänen, Finnish National Institute for Health and Welfare (THL)
Hanna Tolonen, Finnish National Institute for Health and Welfare (THL)