Using the Welsh Index of Multiple Deprivation in research: estimating the effect of excluding domains on a routine health data study.

backgroundThe Welsh Index of Multiple Deprivation (WIMD) is an area-based deprivation measure comprising eight domains, produced by the Welsh Government to rank Lower Layer Super Output Areas (LSOAs) in Wales. Researchers use the WIMD to account for deprivation, however, as one domain contains health indicators, there is a risk of endogeneity bias when using the WIMD in research on health outcomes. This study evaluated the effect on study results of removing the health domain from the overall WIMD or using only the income domain as deprivation measures.

methodsWIMD 2019 scores were linked to 2,760,731 individuals in the SAIL Databank. Original WIMD scores including decile and quintile rankings for each LSOA 2011 were obtained from Welsh Government. The first alternative method removed the health domain from the original WIMD scores. In the second alternative method, WIMD scores were based on only the income domain. Spearman's correlation and Cohen's kappa were used to assess the agreement of ranks, deciles, and quintiles between each method. To quantify the change in association between WIMD quintile and diabetes mellitus prevalence for each alternative method, binary logistic regression obtained age-adjusted odds ratios and 95% confidence intervals.

resultsRemoving the health domain from the original WIMD scores resulted in 17.28% of LSOAs changing decile (8.64% to a more deprived group and 8.64% to a less deprived group) and 9.00% changing quintile (4.50% more deprived, 4.50% less deprived). The income-domain-only method caused 50.49% of LSOAs to change decile (26.87% more deprived, 23.62% less deprived) as compared with the original WIMD, and 29.65% changed quintile (15.14% more deprived, 14.51% less deprived). There was a significant association between each of the three methods and diabetes prevalence, with odds ratios increasing with more deprived quintiles, but the 95% confidence intervals for each method showed little or no overlap with each other.

conclusionTo avoid biased estimates, researchers using WIMD in studies on health, education, housing, physical environment, income, employment, community safety, and access to services should consider how these domains are related to their outcomes. We describe a methodology for researchers to quantify any bias in their own studies.