The effects of sampling and nonresponse on estimates of transition intensities: Some empirical results from the 1981 Swedish fertility survey
preprintposted on 14.01.2021, 14:40 by Ingrid Lyberg
By utilizing register data covering the entire population, we have investigated the sampling and nonresponse effects on estimates based on the 1981 Swedish fertility survey. The life histories gathered in this survey are analyzed by means of semi-Markov models. Age- or duration-specific occurrence/exposure rates are used as estimates of the transition intensities in such models. For the evaluation study we have used semi-Markov models describing events contained in the register, namely births and changes of marital status. Occurrence/exposure rates based on data for the entire target sample or only the respondent in the survey are found to be subject to large stochastic errors, especially at high ages or durations where the number of observations becomes rather small. In spite of a fairly high response rate in the survey (87 percent)·, there exist systematic nonresponse errors, but they are small compared with the stochastic errors which can be estimated. If we disregard single extreme values and use statistical methods to test results based on data for the entire target sample or for the respondents only, we would hardly draw any wrong conclusions about the changes between birth cohorts. Replicated analysis based on data for the entire population, the entire sample, and the respondents show essentially the same pattern.