2016-Missing Data in Survey Sampling


Missing Data in Survey Sampling 
Chair: Michael Hidiroglou (Statistics Canada)
Organizer: Wes Yung (Statistics Canada) 
[PDF]

SONG CAI, Carleton University
Empirical Likelihood Inference Based on Estimation Equations for Complex Survey with Data Missing Completely At Random  [PDF]
 
We propose an empirical likelihood (EL) method for constructing confidence intervals for population parameters defined by estimation equations (EEs) with data from complex surveys that are missing completely at random. Instead of imputing the original data, we impute the EEs using fractional imputation with fixed number of draws. We then construct an EL ratio based on the imputed EEs and show that this ratio has a chi-bar-square limiting distribution in general. We also show that, when no missing data are present, the proposed EL ratio has a simple chi-square limiting distribution under PPS sampling with replacement. Moreover, we propose a proper bootstrap procedure to approximate the limiting distribution of the EL ratio for constructing confidence intervals for parameters of interest. 
 
DAVID HAZIZA, Université de Montréal
Properties of the Instrument Vector Calibration Estimator in the Presence of Unit Nonresponse [PDF]
 
In recent years, instrument vector calibration has received a lot of attention in the literature in the context of unit nonresponse in surveys. In this paper, we lay out the conditions under which the instrument vector calibration estimator is consistent. Even when the conditions are met, the resulting estimator may suffer from variance amplification. When the conditions are not met, the estimator may suffer from both bias and variance amplification. Results from a simulation study will be presented. 
 
XICHEN SHE, University of Waterloo
Fully Efficient Joint Fractional Imputation for Incomplete Bivariate Ordinal Responses  [PDF]
 
In medical studies and many fields of social sciences, pairwise ordinal responses are one of the common data formats collected for analysis. It is often that one or both of the responses are subject to missingness. In this paper, we propose a fully efficient joint fractional imputation procedure, which creates a single complete dataset with fractional weights. We show that this synthetic dataset not only leads to valid inference for simple parameters such as marginal proportions but also preserves the correlation structure, which further leads to valid association measure estimation. Theoretical results on inference procedures based on the fractionally imputed dataset are presented, and finite sample performances of the methods with comparisons to existing methods are investigated through simulation studies.