Article Text
Abstract
Objective To compare the effects of accounting for different missing data mechanisms in an investigation of the role of lifecourse socio-economic position (SEP) on later-life verbal cognitive ability.
Design Two UK prospective cohort studies.
Participants A nationally representative sample born in 1946 (NSHD) (original N=5362), and a sample of British civil servants (Whitehall II) (original N=10 308).
Methods Linear regression models were used to test associations between SEP at different life stages and verbal ability. Results from complete case analysis (assuming missing completely at random) were compared with those using multiple imputation (assuming missing at random) and a Heckman selection model (assuming missing not at random) for each cohort.
Main outcome measure Verbal cognitive ability in adulthood; the National Adult Reading Test at age 53 years (NSHD), and the Mill Hill Test at ages 55–79 years (Whitehall II).
Results NSHD: Educational qualifications and head of household occupational SEP at age 53 were significantly related to verbal ability using all missing data methods, after adjusting for sex and cognitive function at age 8. The effect of childhood SEP was not significant at the 5% level when using Heckman selection (regression coefficient 0.51 (95% CI −0.25 to 1.27)) but was significant for complete case analysis (regression coefficient 0.83 (95% CI 0.11 to 1.54)). Compared with complete case analysis, the coefficients for SEP were generally higher when multiple imputation was used, but the overall conclusions remained the same. The coefficients using Heckman selection differed from those for the complete case and multiple imputation analyses, with lower coefficients for all SEP variables. Whitehall II: Educational qualifications and current occupational SEP were significantly associated with verbal ability for all missing data methods, after adjusting for age, sex, marital status, employment status (working/retired/long-term sick) and number of times the cognitive tests had been taken. The effect of childhood SEP was not significant. The coefficients were generally higher for multiple imputation than complete case analysis, whereas the Heckman selection coefficients were lower for educational qualifications and adult SEP (regression coefficient (95% CI) −4.46 (−5.12 to −3.78) for Heckman selection vs −5.15 (−5.75 to −4.55) for complete case).
Conclusion Educational qualifications and adult SEP were significant predictors of verbal ability in middle to older age, but results for childhood SEP were inconclusive. Greater differences exist between the results from different missing data methods in the older Whitehall II sample, which may be due to greater selective dropout, which is better accounted for by Heckman selection.