An economic analysis of retirement decisions in Taiwan

Abstract

Over the last 20 years there has been a growth in the relative importance of labour economics as an area of economics, particularly for labour force participation, retirement, and labour force transition. However in Taiwan, due to a lack of suitable data, most of the work in this area has been cross-sectional and time-series data analyses. This thesis uses micro panel data to fill this gap. The data is from the Survey of Health and Living Status of the Middle Aged and Elderly in Taiwan, a rich source of information on employment history from 1989 to 2003. The main econometric methods use the binary response models and continuous-time hazard models to analyse labour force participation, retirement, and labour force transition, paying particular to gender differences. The main empirical results show that older workers, female workers, Mainlander workers, and workers with poor health have a lower probability of labour force participation and a higher hazard rate of retirement. In contrast, Hakka workers, workers with better educational attainment, married male workers, and rural workers have a higher probability of participation in work and a lower hazard rate of retirement. In particular, there is an interesting and conditional result for the Pension variable that for workers with less than 35 years employment duration, the survival curve for workers eligible for a pension lies above that of workers ineligible for a pension; and after 35 years, die results are expected to change, particularly for women. Furthermore, workers with higher predicted earnings have a lower hazard rate of retirement, and workers with higher predicted pension income have a higher hazard rate of retirement. Finally, in the case of labour force transitions, the duration models incorporate time-varying covariate factors and show that being in poor health increases the hazard rate of retirement, other things being equal. In addition, as the models consider unobserved heterogeneity factors and find that most estimated coefficients on the regressors are lightly larger in magnitude than the corresponding coefficients in the reference model. Further, unobserved heterogeneity factors are also found to be less serious once time-varying covariates are included in the hazard model