a paper by Salisbury et al. (2023) The Social Vulnerability Indicator (SVM) is considered to be an improvement over previous Social Determinants of Health (SDOH) measures such as the Social Vulnerability Index (SVI). SVI uses census tract-level data to construct overall neighborhood rankings based on variables included in four themes:
- “Socioeconomic status” including poverty, unemployment, income, percentage without a high school diploma;
- “Family composition and disability”, including the proportions of those over 65 years old, under 17 years old, disabled, and single-parent families;
- “Minority status and language” includes minority percentages and “not so good” English,
- “Housing Type and Transportation” includes proportions of multi-unit structures, mobile homes, crowding, no vehicles, and group housing
I summarize how to use SVI as part of an allocation cost-benefit analysis (DCEA) methodology here.
On the other hand, the Social Vulnerability Measure (SVM) proposes Salisbury et al. (2023) was constructed using Multidimensional Item Response Theory (MIRT), using data from Agency for Healthcare Research and Quality (AHRQ) SDoH Database. The main difference from SVI.
- importance rather than equal weight. SVMs are built using the MIRT method, specifically the two-factor model with full information items. In contrast to standard latent variable regression, MIRT allows multiple latent variable constructions. The two-factor model specifically “imposes constraints on traditional item factor analysis by requiring each item to be loaded into one main dimension (eg, SDoH) and only one subdomain (eg, physical infrastructure).” The coefficients of MIRT are used to weight the variables in the SVM. This approach differs from measures such as the CDC’s SVI and the Area Deprivation Index (ADI), which give equal weight to all variables,
- geographic unit. SVM is based on observations at the zip code level, while SVI traditionally uses census data.However, AHRQ stated that their SDoH database will be updated to
Future block groups of counties, zip codes, and census tracts (and the SVM will be computed for each of these groups).
The variables included in the SVM span 5 domains.
- Demographics (for example, age and race/ethnicity),
- educate,
- economic background (e.g. unemployment rate),
- physical infrastructure (such as housing and transportation),
- health care (eg, health insurance).
Note that race/ethnicity is not included in the SVM, in part so that the SVM can be used to compare SDoH across racial and ethnic groups.
result
When comparing SVM to SVI, SVM performed better in predicting all-cause, age-adjusted mortality (r=0.68 versus r=0.34). For individuals aged 0-18 years (r = 0.62) and 18 years and older (r = 0.60), SVM was also inversely associated with receipt of one or more COVID-19 vaccinations (r = -0.68) and completion of full vaccination (r = -0.70), and positively associated with age-adjusted asthma ED visits.
While SVM does appear to outperform SVI, it is a bit more complicated to create due to unequal weights and the fact that the coefficients actually span multiple latent variables.
You can read the full text here.
