Biomedical survey participants: characteristics, missing data, and the association between maternal education and metabolic syndrome
- Chukwuma Iwundu(Author),
- Yannis Pappas(Supervisor)
Student Thesis:
Student thesis
Doctoral thesis
About the thesis
Missing data arising from non-response to surveys poses a significant threat to the validity of estimates in longitudinal cohort studies, potentially introducing substantial bias in exposure-outcome associations. This challenge is particularly salient in birth cohort data that examine the long-term effects of maternal education on offspring health outcomes.
This thesis addresses this critical methodological gap using the 1958 British Birth Cohort (National Child Development Study - NCDS). The research meticulously investigates non-response patterns in biomedical surveys, evaluates strategies for mitigating missing data bias, and assesses the implications of adjustment methods in an epidemiological context.
This thesis encompasses two interrelated studies. Study 1 identifies participant
characteristics predictive of non-response in the midlife NCDS biomedical survey and performs a scoping review of bias-minimising adjustment methods. Study 2 applies these insights to mitigate missing data issues, while conducting a thorough epidemiological analysis of the risk markers and mediators in the association between childhood maternal educational level (CMLE) and midlife metabolic syndrome. This analysis employs directed acyclic graphs (DAGs) for causal identification and structural equation modelling (SEM) for pathway estimation.
Both studies utilised the NCDS subsample eligible for the 2002–2004 biomedical follow-up. Regression models, drawing on harmonised data from seven NCDS sweeps (0–6), were used to delineate responder/non-responder differences and evaluate missing data adjustments in the CMLE–Metabolic syndrome relationship.
Study 1 identified robust predictors of non-response to the biomedical survey. Elevated odds ratios (ORs) were observed for characteristics such as male gender (OR 1.31, 95% CI 1.10–1.55, p<0.01), lower paternal social class (OR 2.40, 95% CI 1.18–4.88, p<0.01), and suboptimal self-reported general health (OR 1.87, 95% CI 1.28–2.32, p<0.05). The scoping review established Multiple Imputation (MI) as the least biased method, outperforming Single Imputation (SI), Complete Case Analysis (CCA), and Maximum Likelihood (ML) estimation.
Building on this, Study 2 demonstrated that MI, Inverse Probability Weighting (IPW), and CCA yielded comparable estimates for the CMLE–Metabolic syndrome association, with MI producing smaller standard errors due to sample size augmentation. Regression analyses revealed significant inverse associations between low CMLE and midlife Metabolic syndrome risk markers, including waist circumference (β=−1.62), systolic blood pressure (β=−1.50), HDL-cholesterol (β=−1.62), and HbA1c (β=−0.05) (all p<0.001). Individuals with low CMLE displayed a higher prevalence of clustered (≥3) risk markers.
The SEM mediation analysis was a key output. Body Mass Index (BMI) accounted for the largest proportion of the total effect (77%), followed by smoking (7.6%), exercise (2.48%), and alcohol frequency (0.46%). This finding strongly suggests that BMI mediates the considerable majority of the protective relationship between CMLE and metabolic syndrome, with only a minor proportion attributable to direct effects or other unmeasured mechanisms.
In conclusion, this thesis highlights the critical importance of understanding non-response determinants and implementing robust missing data adjustments in longitudinal research. By integrating methodological and epidemiological approaches, it demonstrates that maternal education exerts a lasting impact on offspring metabolic health. The findings advocate for the use of Multiple Imputation as the preferred method for bias reduction and contribute to advancing methodological rigour in cohort-based health research.
This thesis addresses this critical methodological gap using the 1958 British Birth Cohort (National Child Development Study - NCDS). The research meticulously investigates non-response patterns in biomedical surveys, evaluates strategies for mitigating missing data bias, and assesses the implications of adjustment methods in an epidemiological context.
This thesis encompasses two interrelated studies. Study 1 identifies participant
characteristics predictive of non-response in the midlife NCDS biomedical survey and performs a scoping review of bias-minimising adjustment methods. Study 2 applies these insights to mitigate missing data issues, while conducting a thorough epidemiological analysis of the risk markers and mediators in the association between childhood maternal educational level (CMLE) and midlife metabolic syndrome. This analysis employs directed acyclic graphs (DAGs) for causal identification and structural equation modelling (SEM) for pathway estimation.
Both studies utilised the NCDS subsample eligible for the 2002–2004 biomedical follow-up. Regression models, drawing on harmonised data from seven NCDS sweeps (0–6), were used to delineate responder/non-responder differences and evaluate missing data adjustments in the CMLE–Metabolic syndrome relationship.
Study 1 identified robust predictors of non-response to the biomedical survey. Elevated odds ratios (ORs) were observed for characteristics such as male gender (OR 1.31, 95% CI 1.10–1.55, p<0.01), lower paternal social class (OR 2.40, 95% CI 1.18–4.88, p<0.01), and suboptimal self-reported general health (OR 1.87, 95% CI 1.28–2.32, p<0.05). The scoping review established Multiple Imputation (MI) as the least biased method, outperforming Single Imputation (SI), Complete Case Analysis (CCA), and Maximum Likelihood (ML) estimation.
Building on this, Study 2 demonstrated that MI, Inverse Probability Weighting (IPW), and CCA yielded comparable estimates for the CMLE–Metabolic syndrome association, with MI producing smaller standard errors due to sample size augmentation. Regression analyses revealed significant inverse associations between low CMLE and midlife Metabolic syndrome risk markers, including waist circumference (β=−1.62), systolic blood pressure (β=−1.50), HDL-cholesterol (β=−1.62), and HbA1c (β=−0.05) (all p<0.001). Individuals with low CMLE displayed a higher prevalence of clustered (≥3) risk markers.
The SEM mediation analysis was a key output. Body Mass Index (BMI) accounted for the largest proportion of the total effect (77%), followed by smoking (7.6%), exercise (2.48%), and alcohol frequency (0.46%). This finding strongly suggests that BMI mediates the considerable majority of the protective relationship between CMLE and metabolic syndrome, with only a minor proportion attributable to direct effects or other unmeasured mechanisms.
In conclusion, this thesis highlights the critical importance of understanding non-response determinants and implementing robust missing data adjustments in longitudinal research. By integrating methodological and epidemiological approaches, it demonstrates that maternal education exerts a lasting impact on offspring metabolic health. The findings advocate for the use of Multiple Imputation as the preferred method for bias reduction and contribute to advancing methodological rigour in cohort-based health research.