A recent paper [Thibaudeau, Slud, and Gottschalck (2017). Modeling log-linear conditional probabilities for estimation in surveys. The Annals of Applied Statistics, 11, 680–697] proposed a ‘hybrid’method of survey ...A recent paper [Thibaudeau, Slud, and Gottschalck (2017). Modeling log-linear conditional probabilities for estimation in surveys. The Annals of Applied Statistics, 11, 680–697] proposed a ‘hybrid’method of survey estimation combining coarsely cross-classified design-based survey-weightedtotals in a population with loglinear or generalised-linear model-based conditional probabilitiesfor cells in a finer cross-classification. The models were compared in weighted and unweightedforms on data from the US Survey of Income and Program Participation (SIPP), a large nationallongitudinal survey. The hybrid method was elaborated in a book-chapter [Thibaudeau, Slud,& Cheng (2019). Small-area estimation of cross-classified gross flows using longitudinal survey data. In P. Lynn (Ed.), Methodology of longitudinal surveys II. Wiley] about estimating grossflows in (two-period) longitudinal surveys, by considering fixed versus mixed effect versionsof the conditional-probability models and allowing for 3 or more outcomes in the later-periodcategories used to define gross flows within generalised logistic regression models. The methodology provided for point and interval small-area estimation, specifically area-level two-periodlabour-status gross-flow estimation, illustrated on a US Current Population Survey (CPS) datasetof survey respondents in two successive months in 16 states. In the current paper, that data analysis is expanded in two ways: (i) by analysing the CPS dataset in greater detail, incorporatingmultiple random effects (slopes as well as intercepts), using predictive as well as likelihood metrics for model quality, and (ii) by showing how Bayesian computation (MCMC) provides insightsconcerning fixed- versus mixed-effect model predictions. The findings from fixed-effect analyseswith state effects, from corresponding models with state random effects, and fom Bayes analysisof posteriors for the fixed state-effects with other model coefficients fixed, all confirm each otherand support a model with normal random state effects, independent across states.展开更多
文摘A recent paper [Thibaudeau, Slud, and Gottschalck (2017). Modeling log-linear conditional probabilities for estimation in surveys. The Annals of Applied Statistics, 11, 680–697] proposed a ‘hybrid’method of survey estimation combining coarsely cross-classified design-based survey-weightedtotals in a population with loglinear or generalised-linear model-based conditional probabilitiesfor cells in a finer cross-classification. The models were compared in weighted and unweightedforms on data from the US Survey of Income and Program Participation (SIPP), a large nationallongitudinal survey. The hybrid method was elaborated in a book-chapter [Thibaudeau, Slud,& Cheng (2019). Small-area estimation of cross-classified gross flows using longitudinal survey data. In P. Lynn (Ed.), Methodology of longitudinal surveys II. Wiley] about estimating grossflows in (two-period) longitudinal surveys, by considering fixed versus mixed effect versionsof the conditional-probability models and allowing for 3 or more outcomes in the later-periodcategories used to define gross flows within generalised logistic regression models. The methodology provided for point and interval small-area estimation, specifically area-level two-periodlabour-status gross-flow estimation, illustrated on a US Current Population Survey (CPS) datasetof survey respondents in two successive months in 16 states. In the current paper, that data analysis is expanded in two ways: (i) by analysing the CPS dataset in greater detail, incorporatingmultiple random effects (slopes as well as intercepts), using predictive as well as likelihood metrics for model quality, and (ii) by showing how Bayesian computation (MCMC) provides insightsconcerning fixed- versus mixed-effect model predictions. The findings from fixed-effect analyseswith state effects, from corresponding models with state random effects, and fom Bayes analysisof posteriors for the fixed state-effects with other model coefficients fixed, all confirm each otherand support a model with normal random state effects, independent across states.
