Suchergebnisse

Abstract
Data synthesis is an effective statistical approach for reducing data disclosure risk. Generating fully synthetic data might minimize such risk, but its modeling and application can be difficult for data from large, complex surveys. This article extended the two-stage imputation to simultaneously impute item missing values and generate fully synthetic data. A new combining rule for making inferences using data generated in this manner was developed. Two semiparametric missing data imputation models were adapted to generate fully synthetic data for skewed continuous variable and sparse binary variable, respectively. The proposed approach was evaluated using simulated data and real longitudinal data from the Health and Retirement Study. The proposed approach was also compared with two existing synthesis approaches: (1) parametric regressions models as implemented in IVEware; and (2) nonparametric Classification and Regression Trees as implemented in synthpop package for R using real data. The results show that high data utility is maintained for a wide variety of descriptive and model-based statistics using the proposed strategy. The proposed strategy also performs better than existing methods for sophisticated analyses such as factor analysis.

Abstract
Mixed-mode designs have become increasingly common in survey data collection. Although different modes often have different measurement properties, the standard practice is to treat mixed-mode data as if they had been collected with a single mode, neglecting the potential impact of mode effects. To account for potential mode effects when making inferences for mixed-mode samples, we propose (i) a Testimator approach, (ii) a Bayesian approach, and (iii) a model averaging approach. In the Testimator approach, we test whether the means and the variances of mixed-mode samples are the same. If the means are the same, we take the average of mode-specific estimates. If the means are different, we take the average when we have no prior information about preferred modes and take the smaller (or larger) estimate when we have prior information about preferred modes (e.g., a smaller estimate is better). In the Bayesian approach, we assume some prior information. We use a data-driven method to determine whether there are mode effects. If there are no mode effects, we draw inferences using a common mean model. If there are mode effects, we draw inferences using the data collected with the mode that produces smaller estimates. In the model averaging approach, we combine estimates of different models (characterized by whether assume same means and variances across modes) using marginal posteriors as weights. We evaluate the approaches in simulation studies and find that they achieve robust inferences compared to the standard approach. We apply the methods to the Arab Barometer study, which employs a randomized mixed-mode design.

Abstract
Mixing multiple modes of survey data collection has become standard practice in survey research. Mixed-mode surveys are faced with a slew of design decisions regarding which types of modes to administer and which sequence to administer them in. Such decisions are largely based on administrative objectives, such as minimizing costs and maximizing response rates. However, just as important to these mixed-mode decisions is their impact on nonresponse bias, measurement error bias, and total bias, which are understudied issues in the mixed-mode literature. In this article, we report on a sequential mixed-mode experiment of young adult drivers randomized to one of two mode sequences: an interviewer-administered (telephone) mode with self-administered (mail) follow-up, or the reverse sequence. Using a mix of direct and indirect bias estimation strategies, we find support for the notion that implementing a second mode of data collection can reduce nonresponse and measurement error bias, but the sequence in which the modes are administered makes a difference: the mail-telephone sequence minimizes bias to a greater extent than the telephone-mail sequence, relative to the starting mode and overall. However, a backfiring effect was found: despite reducing both nonresponse and measurement error bias, switching from mail to telephone increased the total bias in a key estimate of traffic accidents. A discussion of these findings and their implications for survey practice are provided in conclusion.