Abstract
Background: Low health literacy is associated with poorer health. Research has shown that predictive models of health literacy can be developed; however, key variables may be missing from systems where predictive models might be applied, such as health service data. This paper describes an approach to developing predictive health literacy models using variables common to both "source" health literacy data and "target" systems such as health services. Methods: A multilevel synthetic estimation was undertaken on a national (England) dataset containing health literacy, socio-demographic data and geographical (Lower Super Output Area: LSOA) indicators. Predictive models, using variables commonly present in health service data, were produced. An algorithm was written to pilot the calculations in a Family Physician Clinical System in one inner-city area. The minimum data required were age, sex and ethnicity; other missing data were imputed using model values. Results: There are 32,845 LSOAs in England, with a population aged 16 to 65 years of 34,329,091. The mean proportion of the national population below the health literacy threshold in LSOAs was 61.87% (SD 12.26). The algorithm was run on the 275,706 adult working-age people in Lambeth, South London. The algorithm could be calculated for 228,610 people (82.92%). When compared with people for whom there were sufficient data to calculate the risk score, people with insufficient data were more likely to be older, male, and living in a deprived area, although the strength of these associations was weak. Conclusions: Logistic regression using key socio-demographic data and area of residence can produce predictive models to calculate individual- and area-level risk of low health literacy, but requires high levels of ethnicity recording. While the models produced will be specific to the settings in which they are developed, it is likely that the method can be applied wherever relevant health literacy data are available. Further work is required to assess the feasibility, accuracy and acceptability of the method. If feasible, accurate and acceptable, this method could identify people requiring additional resources and support in areas such as medical practice.