Abstract
Background
As a consequence of declining levels of participation in health surveys, the results purported to be population-representative may be biased. Traditional adjustments for non-participation, such as weighting, can fail to correct for such biases. We aim to validate our developed methodology, which simulates non-participants, and compare results from the inferred sample to the ’gold standard’ sample of participants and true non-participants, and participants alone.
Methods
Participants and non-participants of the Finnish Health 2000 survey, and a contemporaneous population sample are available, with alcohol-related hospitalisations and deaths (“harms”, individually record-linked for all Health 2000 invitees). Synthetic observations on non-participants were simulated through comparison of participants and population sample. Alcohol consumption of true and inferred non-participants were multiply imputed based on harms and education as well as age and sex, assuming data are Missing At Random (MAR). Results are compared via the relative differences (RD) between the inferred sample and 1) gold standard sample, and 2) participants alone.
Results
Average weekly estimates for men are 129g in the inferred sample, and 130g in the gold standard (RD -1.2%, 95%CI -2.0, -0.4%), and 35g for women in both samples (RD -0.8%; -1.9, 0.3%). Estimates for men with secondary levels of education had the greatest RD (-1.9%; -3.3, -0.5%). Comparisons between the participants and the inferred sample revealed few differences.
Conclusions
All RD between the inferred and gold standard samples lie within our ±5% acceptability limits, in support of the use of our methodology for adjusting for non-participation in health surveys. However, under MAR, there are no significant differences between the results generated from the inferred sample and the participants alone. Further work exploring Missing Not At Random scenarios is required to ensure utility for reliable population health monitoring.
Key messages
Survey weights alone cannot adjust for non-representativeness, but we have shown that data linkage can be used to match the characteristics and outcomes of the selected sample. Non-participation in health surveys may be adjusted for using our methodology, with further exploration into alternative missing data scenarios required.
Improving National Health Surveys through Data Linkage: Implications for Epidemiologic Research
