Posters

Rodrigo Arruda, Luiza Campos,
Tatiane de Menezes
and Cristiano Ferraz

A Didactic Review of Hot-Deck Imputation Methods and its Applications

Rodrigo Arruda, Economics Department, Federal University of Pernambuco, Luiza Campos, Economics Department, Federal University of Pernambuco, Department of Economics, Federal University of Pernambuco,Tatiane de Menezes, Department of Economics, Federal University of Pernambuco,and Cristiano Ferraz,Department of Statistics and NEPS - Research Group on Criminality, Violence and Security Policy

Studies on criminality and security policy analyze data sets from census and sample surveys subjected to non-response errors. We are currently interested in studying how violence may affect basic education in Brazil, analyzing violence at school related data, collected from the SAEB/Prova Brasil. The SAEB/Prova Brasil is a national large-scale survey applied to students, professors and directors of private and public schools every two years since 1990. In this poster we summarize the first steps done to treat the survey missing data, by didactically introducing the subject of imputation, focusing on a review of hot-deck methods. The key-concepts are introduced and the operational imputation steps are illustrated with an artificial data set based on the SAEB/Prova Brasil collected information.

Sixia Chen

Jackknife Empirical Likelihood Method for Inference with Regression Imputation

Sixia Chen,Senior survey methodologist, Westat, 1600 Research blvd, Rockville, MD, USA

We propose novel jackknife empirical likelihood (JEL) methods for constructing confidence intervals of mean functionals with nonparametric and semi-parametric regression imputation, respectively, under ignorable and non-ignorable missingness assumption. The JEL is constructed based on the adjusted jackknife pseudo-values (Rao and Shao, 1992). It is shown that the proposed JEL ratios evaluated at the true value converge to the standard chi-square distribution under both missing mechanisms for simple random sampling. Thus the JEL can be applied to construct a Wilk-type confidence interval directly without any secondary estimation. We then extend the proposed method to accommodate Poisson sampling design in survey sampling. The proposed methods are compared with some existing methods in simulation studies. Both theory and simulation results show the benefits of our proposed methods. We also demonstrate the proposed method in an application to Italy Household Income Panel Survey data.

Hemílio
Fernandes Campos Coelho,
Cristiano Ferraz and
Camila Ribeiro da Silva

Proportion estimators in dual frame surveys with auxiliary
information

Hemílio Fernandes Campos Coelho, Universidade Federal da Paraíba, Departamento de Estatística, Brazil, Cristiano Ferraz, Universidade Federal de Pernambuco, Departamento de Estatística, Brazil,and Camila Ribeiro da Silva, Universidade Federal da Paraíba, Departamento de Estatística, Brazil

Let A and B denote the sets of elements identified by each one of two frames covering the same target population, U=A∪B. In dual frame surveys, probability samples are independently drawn from A and from B, with A∩ B≠∅. We consider the problem of incorporating auxiliary information, available from at least one of the frames, into a generalized regression type estimator for population proportion. The work proposed by Lehtonen and Veijanen (1998) on logistic regression estimators for single frame surveys is reviewed and generalized for a dual frame version. Preliminary results based on Monte Carlo simulation experiments are presented.

Luis Fernando Contreras-Cruz

Varying Coefficient Models in Finite Population Sampling

Luis Fernando Contreras-Cruz, Colegio de Postgraduados, México

A model-assisted semiparametric method of estimating finite-population totals is investigated to improve the precision of survey estimators by incorporating multivariate auxiliary information. The proposed superpopulation model is a varying-coefficient model. Under standard design conditions, the proposed estimators are asymptotically design-unbiased, consistent and asymptotically normal. Both simulated and real data examples are given to illustrate the model and the proposed estimation methodology, which have provided strong evidence that corroborates with the asymptotic theory.

Andreea L. Erciulescu

Small Area Estimation with Incomplete Auxiliary Information

Andreea L. Erciulescu and Wayne A. Fuller, Department of Statistics, Iowa State University, Ames, IA

Surveys are often designed to achieve specifi
c information about totals and means, but direct estimates for small areas or subpopulations may not be reliable because of small sample sizes. Procedures based on models have been used to construct estimates for small areas, by exploiting auxiliary information. In this paper, we
fit nested models with a binary response and random area effects. These models form a subclass of generalized linear mixed models. Because the response variable is binary, the estimation and prediction are not as straight forward as in linear mixed models. We consider three cases of auxiliary information, when the covariates have known mean, when the covariates have estimated
fixed mean, and when the covariates have estimated random mean.

Andreea Erciulescu and Emily Berg

Evaluating Impacts of Nonsampling Errors on Estimates for the Conservation Effects Assessment Project

Andreea Erciulescu and Emily Berg, Department of Statistics, Iowa State University, Ames IA

The Conservation Effects Assessment Project (CEAP) consists of a collection of surveys intended to evaluate envirnomental outcomes associated with conservation practices. Sources of nonsampling error in CEAP include nonresponse and undercoverage of the sampling frame. Nonresponse error occurs because some sampled farm operators refuse to complete the CEAP survey. Coverage error arises due to incomplete information at the design stage.  The population of interest for CEAP consists of locations with cultivated cropland in the survey year, but only historical information on landuse is available when a sample of locations is selected.  Information on soil charachteristics from the National Resources Inventory (NRI) and the Soil Survey is used to investigate possible biases due to nonresponse and undercoverage.

Wade Herndon

When employing model based approaches to the analysis of survey data it is not always immediately clear what role the sampling weights should play in the analysis. When tting a statistical model to survey data, the goal is to t a model that holds at the population level, and the problem arises that the distribution of the data given that they are included in the sample may be dierent from the distribution of the data in the population. Because of this, there is a desire in practice to account for the sampling design while at the same time staying within a classical model based estimation framework. Semiparametric methods for  tting models to survey data that account for the selection process are discussed. This poster will include further motivation of the problem and details of the proposed semiparametric estimator, with an emphasis on an application to data from a Canadian labor force survey.

Jie (Kate) Hu

Application of Z-estimation theory to calibrated estimators for semiparametric models with two-phase stratified sampling

Jie (Kate) Hu, Department of Biostatistics, University of Washington, Gary Chan, Norman Breslow

In epidemiology studies, we are usually interested in parameters specified in a (semi)parametric model describing the association between an exposure and an outcome. When the outcome and the exposure are rare, two-phase stratified sampling design is sometimes implemented to collect the data in order to reduce the cost. In this study design, epidemiologists often obtain auxiliary variables for all phase I observations. To further improve the efficiency, this additional information can also be incorporated into the estimated parameter by calibration.  In this work, we develop a new method to derive the calibrated estimator obtained from this sampling design for epidemiology studies and show how the Z (or M)-estimation theory is applied to the asymptotic variance estimation

Jae-Kwang Kim and Jongho Im

The use of follow-ups for propensity score adjustment with non-ignorable non-response

Jae-Kwang Kim and Jongho Im, Department of Statistics, Iowa State University, Ames, IA

Propensity score weighting method is a popular tool for handling unit non-response in survey sampling. In case of non-ignorable non-response, estimation of the propensity scores is complicated and often requires additional surrogate variables to estimate model parameters (Chang and Kott, 2008). Another way is to use paradata which is defined as data about survey processes such as call records, interviewee aptitudes and the number of attempts to complete interview. Follow-ups can be used to adjust non-ignorable non-response in views of paradata.
When there are several follow-ups and the final follow-up sample is also subject to missingness, Drew and Fuller (1980) developed multinomial likelihood approach for categorical response variables and Alho (1990) extended the approach of Drew and Fuller to the case of general continuous response variable by adopting a logistic regression model as the conditional response probability.
We propose calibration weighting method to estimate the model parameters in the conditional response model already assumed in Alho (1990) and to incorporate auxiliary information using generalized method of moments. We provide some simulation studies to compare performance between our estimator and estimators of Alho (1990) and Chang and Kott (2008). The proposed method is applied to real data example in a Korean household survey of employment.

Nuanpan Nangsue

Estimation of Cluster-level Regression Model under Nonresponse within Clusters

Nuanpan Nangsue, Social Sciences, University of Southampton, UK

Surveys are sometimes conducted using two-stage sampling, where the
first-stage units (clusters) are of analytic interest. In this paper, we suppose that the aim is to
fit a regression model at the cluster level. Nonresponse in two-stage surveys can operate at either stage and, in this paper, we focus on the problem
when nonresponse occurs at the second stage. We are specifi
cally concerned with how to use observed data to make inference about regression coe
fficients in a linear regression model of cluster-level variables when some of the response variable data is missing. A naive approach estimates the regression coe
fficients with-out considering the nonresponse. Section 2 defi
nes our notations and chosen framework. Section 3 defi
nes our model of interest. In Section 4, we propose a new method for estimating regression coe
fficients which incorporate information on nonresponse at the cluster level by extending Heckman (1976) estimators to our clustered model. In Section 5, we present a simulation study to compare the new method with the naive approach and in Section 6 we show a real application using the Workplace Employment Relations Survey(WERS) 2004 data.

Raphael Nishimura

Item Nonresponse in Auxiliary Variables Used in Weighting Adjustments for Survey Sample Data

Raphael Nishimura, Michigan Program in Survey Methodology, Institute for Social Research, Ann Arbor, MI

The present study aims to compare the empirical properties of the Horvitz-Thompson estimator with the Generalized Regression (GREG) estimator in the presence of item nonresponse in one of two variables used in the weighting adjustment through simulations that select simple random samples from different simulated finite populations. These simulations vary the levels of correlations between the auxiliary variable with missing data and the out-come variable, the missing rate levels and the missing mechanisms. The impacts of these conditions are evaluated on each estimation approach and it is verified which estimator performs best in each situation.

Jiwei Zhao

Variance Estimation after Multiple Imputation

Jiwei Zhao, Yale University

In complex survey data, we usually encounter appreciable amount of missing values. The missing data mechanism can be missing completely at random (MCAR), missing at random (MAR), or nonignorable. Multiple Imputation (MI, Rubin 1987) is a well-known and well-established procedure to handle missing values and it is an important technique in the literature. In this paper, we consider the regression model with response variable subject to missing values and we concentrate on the variance estimation after MI. At first, we briefly review the results when the missing data mechanism is MAR (Wang and Robins, 1998). Under MAR assumption, the estimates after MI are always less efficient than the ones before MI. In the following, we focus on the situation when the missing data mechanism is nonignorable. We first propose an estimation procedure before MI under some assumptions on the missing data mechanism. We then conduct MI based on the proposed estimates. However, different from MAR, the variance after MI is not generally necessarily larger than the one before MI. Hence, there is no definite answer to which one, before or after MI, is more efficient under nonignorable assumption. It is possible that MI could increase the estimation efficiency when the missing data mechanism is nonignorable. This is a different phenomenon comparing MAR and nonignorable missing data mechanisms. Finally, intensive simulation studies are conducted to illustrate the finite sample behaviors.