Longitudinal Nonresponse Weight AdjustmentPrevious Evaluation MethodsModel-based AdjustmentsAdjustment MetricsSubdomain Bias & 1st MetricWhole-Pop Bias vs. Bound2nd Metric: Preserving CellsComputed Results on SIPP 96SummaryLongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryEvaluation and Selection of Models for AttritionNonresponse AdjustmentEric Slud and Leroy BaileyStatistical Research Div., Census BureauFCSM Research Conference, 2007Eric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryOutline1Longitudinal Nonresponse Weight AdjustmentPrevious Evaluation MethodsModel-based Adjustments2Adjustment MetricsSubdomain Bias & 1st MetricWhole-Pop Bias vs. Bound2nd Metric: Preserving Cells3Computed Results on SIPP 96Eric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryDefinitions motivated by SIPPSurvey of Income & Program ParticipationIndividuals i observed in successive ‘Waves’Sample s here denotes Wave-1 respondersWeights wiare inverse inclusion prob’sre-weighted for Wave-1 response & raked.Survey items yie.g. SocSec income indicatorare only Wave 1 data values.ri= indicator of later-wave (say Wave-4) responsetreated as random beyond probability sampleˆpimodel-based estimator of pi= Pr(ri= 1| i ∈ s)Eric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryDefinitions motivated by SIPPSurvey of Income & Program ParticipationIndividuals i observed in successive ‘Waves’Sample s here denotes Wave-1 respondersWeights wiare inverse inclusion prob’sre-weighted for Wave-1 response & raked.Survey items yie.g. SocSec income indicatorare only Wave 1 data values.ri= indicator of later-wave (say Wave-4) responsetreated as random beyond probability sampleˆpimodel-based estimator of pi= Pr(ri= 1| i ∈ s)Eric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryPrevious Evaluation MethodsModel-based AdjustmentsTracking of time-changes in weightsDufour et al. 2001 study adj. weights wiri/ˆpisimulated using Canadian Labor Force survey.Present Methodology – Internal EvaluationTarget ty=Pi∈Uyiwave 1 responder totalEstimates initialPi∈swiyi(Horvitz-Thompson)at later wave, adjusted:Pi∈swiyiri/ˆpiAdj. BiasPi∈swiyinri/ˆpi− 1oSIPP Implemented in Bailey (2004), withSE calculation in Slud & Bailey (2006).Eric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryPrevious Evaluation MethodsModel-based AdjustmentsTracking of time-changes in weightsDufour et al. 2001 study adj. weights wiri/ˆpisimulated using Canadian Labor Force survey.Present Methodology – Internal EvaluationTarget ty=Pi∈Uyiwave 1 responder totalEstimates initialPi∈swiyi(Horvitz-Thompson)at later wave, adjusted:Pi∈swiyiri/ˆpiAdj. BiasPi∈swiyinri/ˆpi− 1oSIPP Implemented in Bailey (2004), withSE calculation in Slud & Bailey (2006).Eric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryPrevious Evaluation MethodsModel-based AdjustmentsTracking of time-changes in weightsDufour et al. 2001 study adj. weights wiri/ˆpisimulated using Canadian Labor Force survey.Present Methodology – Internal EvaluationTarget ty=Pi∈Uyiwave 1 responder totalEstimates initialPi∈swiyi(Horvitz-Thompson)at later wave, adjusted:Pi∈swiyiri/ˆpiAdj. BiasPi∈swiyinri/ˆpi− 1oSIPP Implemented in Bailey (2004), withSE calculation in Slud & Bailey (2006).Eric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryPrevious Evaluation MethodsModel-based AdjustmentsModel-based AdjustmentsCell-based Method: for i in cell C ⊂ U,ˆpi=Xj∈Crjwj.Xj∈CwjLogistic Regression Method:ˆpi= (1 + e−x0iˆβ)−1fitted from weighted ML equation, Wave 1 covariates xiand later-Wave indicator rias response.Score-statistic model comparisons: Kim and Kim (2007).Model comparisons via Subdomain Bias EstimatesEric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryPrevious Evaluation MethodsModel-based AdjustmentsModel-based AdjustmentsCell-based Method: for i in cell C ⊂ U,ˆpi=Xj∈Crjwj.Xj∈CwjLogistic Regression Method:ˆpi= (1 + e−x0iˆβ)−1fitted from weighted ML equation, Wave 1 covariates xiand later-Wave indicator rias response.Score-statistic model comparisons: Kim and Kim (2007).Model comparisons via Subdomain Bias EstimatesEric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryPrevious Evaluation MethodsModel-based AdjustmentsModel-based AdjustmentsCell-based Method: for i in cell C ⊂ U,ˆpi=Xj∈Crjwj.Xj∈CwjLogistic Regression Method:ˆpi= (1 + e−x0iˆβ)−1fitted from weighted ML equation, Wave 1 covariates xiand later-Wave indicator rias response.Score-statistic model comparisons: Kim and Kim (2007).Model comparisons via Subdomain Bias EstimatesEric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummaryPrevious Evaluation MethodsModel-based AdjustmentsModel-based AdjustmentsCell-based Method: for i in cell C ⊂ U,ˆpi=Xj∈Crjwj.Xj∈CwjLogistic Regression Method:ˆpi= (1 + e−x0iˆβ)−1fitted from weighted ML equation, Wave 1 covariates xiand later-Wave indicator rias response.Score-statistic model comparisons: Kim and Kim (2007).Model comparisons via Subdomain Bias EstimatesEric Slud and Leroy Bailey Evaluation of Nonresponse AdjustmentsUSCENSUSBUREAULongitudinal Nonresponse Weight AdjustmentAdjustment MetricsComputed Results on SIPP 96SummarySubdomain Bias & 1st MetricWhole-Pop Bias vs. Bound2nd Metric: Preserving CellsSubdomain