Quantifying and Correcting Bias in Fertility Estimation from Complete Birth Histories

Thesis (Master's)--University of Washington, 2018Introduction Accurate and unbiased estimates of fertility are critical in creating population policy and evaluating maternal mortality ratio. In settings without the routine collection of vital statistics, fertility data often comes from complete birth histories, which can suffer from biases due to the passage of time, including recall bias—women forgetting their births—and selection bias—women who die having systematically different fertility than those who live to be surveyed. Comprehensive estimates of time-related bias using nationally representative surveys in multiple locations do not exist. Methods This paper develops a method to measure and correct for these biases using 299 complete birth history surveys. I calculate cohort fertility rates of the same cohort from different surveys with different recall periods, find the ratio of a reference cohort fertility rate from the survey with the lowest recall to each other estimated cohort fertility rate, and model the log ratio as a function of recall length, reference recall length, age, maternal mortality ratio, female adult age-specific mortality, and region. These ratios are then predicted as if the reference recall was 0 and applied to the raw cohort Age Specific Fertility Rate (ASFR) to get a value that is adjusted for bias. Results The cohort-recall specific estimates for each country of analysis showed considerable patterning across different recall periods. Predicted ratios of reference ASFR, with the lowest possible recall within a country, cohort, and age, to survey specific ASFR ranged between 0.66 for a three year recall of the 1948 birth cohort in Timor-Leste and 1.40 for a 26 year recall of the 1965 birth cohort also in Timor-Leste, with a mean and median of 0.98. The relationship between recall and the ratio varied by age. For younger ages, including 15-19 and 20-24 years, the model found an underestimation of fertility in relation to the reference with an increasing recall period; for older ages, including 25-29, 30-34, 35-39, and 40-44 years, the model found an overestimation of fertility in relation to the reference with an increasing recall period. Discussion This analysis represents the first systematic quantification and correction of fertility bias due to the passage of time. The results of this analysis suggest that selection bias due to differential fertility among those mothers who die and those who live to be surveyed contributes to the bias in fertility estimates. The application of the results could substantially increase the amount of data available to be used back in time if restrictions on recall are lifted. Limitations include the instability of the input data, the non-inclusion of age heaping and birth transference in analysis, and the focus on complete birth history methods to the exclusion of indirect methods. Additional work, including the development of alternate methods, is necessary to validate the results of this analysis. Supplementary Material Supplemental material includes graphs of cohort ASFR results for each country age survey