Empirical findings and research methods in health and mortality: Early life conditions and health and mortality - The case of Mexico and China; Application of multiple imputation to missing data due to sub -sampling

My three dissertation essays share common interests with empirical studies and research methods in health and mortality. Using Chinese Longitudinal Healthy Longevity Survey (CLHLS), Chapter 2 shows that early life conditions are associated with mortality risk at very old age and that this association varies by gender. Father\u27s occupation and sibling compositions are significant predictors of mortality in very late life for both females and males. Arm length, a marker of early life nutrition, predicts mortality only for females. I do not find strong evidence that adult characteristics mediate the association between early life conditions and mortality at very old age, although some adult characteristics, such as living arrangements and physical exercise, exhibit significant associations with mortality risk at very old age. Based on the Mexican Health and Aging Study (MHAS), Chapter 3 indicates that early life conditions predict lower body functional limitations among Mexicans aged 50 years and older, after controlling for a variety of adult characteristics. Whether the respondent often went to bed hungry and whether the respondent had serious health problems before age 10, are significant predictors of adulthood lower body functional limitations for both males and females. I also find that adult characteristics, possibly including chronic diseases and symptoms, such as diabetes, arthritis and pain, mediate the associations between early life conditions and lower body functional limitations. Chapter 4 investigates the application of multiple-imputation (MI) to missing data due to sub-sampling such that only a sub-sample of respondents is chosen for some particular measurements. Based on a simulation study, Chapter 4 suggests that when the data are under a multivariate normal distribution, using the entire sample and imputing values for cases that have missing data is a better choice than using only the sub-sample, in terms of getting unbiased coefficients and smaller standard errors, except for the case where the item measured only in the sub-sample is used as the dependent variable in the analysis. In this situation, both multiple-imputation and list-wise deletion may result in similarly biased estimates. Finally, multiple-imputation may not work well under certain circumstances when the multivariate normality assumption is violated