Register based human population size estimation: a geometric perspective toward assessing sensitivity to unverifiable independence assumptions

Open Access
- Author:
- Johnston, Ann
- Graduate Program:
- Statistics
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- July 01, 2020
- Committee Members:
- Aleksandra B Slavkovic, Thesis Advisor/Co-Advisor
Ephraim Mont Hanks, Committee Member
Murali Haran, Program Head/Chair - Keywords:
- Log-linear models
Capture-recapture
Sensitivity analysis
Data integration
Official statistics - Abstract:
- Human population size estimation based on capture-recapture methodology is a problem of substantial interest within the field of official statistics. With data sourced from T administrative registers, a log-linear models framework can be used to model the counts of a 2^T–tabulation of the population of interest, where the sum of the estimates for these counts yield an estimate of population size. Within this framework, the structure of the available data affords no ability to estimate a T–way interaction effect. Thus, the standard approach assumes no T–way interaction, even though this may be highly unrealistic in contexts of administrative records on human populations. Given capture-recapture data from two administrative registers, recent work has described a strategy for assessing the sensitivity/robustness of the classic population size estimate to deviations in the odds-ratio away from one (the value under independence). In this work, we generalize this strategy to the setting of data from T administrative registers. Moreover, we introduce a univariate sensitivity quantifier that is novel, geometrically motivated, and easy to interpret. This quantifier, which is supported on the interval (0, 1), measures the sensitivity of the classic capture-recapture population size estimate to deviations in the true value of the order–T odds ratio away from one (the value under no T–way interaction). Taking advantage of our proposed sensitivity quantifier, we outline an alternative sensitivity assessment strategy; and, we demonstrate its use through several illustrative examples. The approach we propose is simple in its interpretation and computationally straightforward, making it attractive for use by practitioners.