Sociological Methods & Research

Based on the local influence approach with an appropriate perturbation scheme, measures are developed to assess the influence of an individual cell in contingency tables. The authors demonstrate that the developed measures are particularly suitable for sensitivity analysis in tables with sampling zeros. The independence model and the log-linear model, together with numerical examples, are presented for illustration.

This article compares estimates of “informed” public opinion derived from the regression-based approach of Bartels, Delli Carpini and Keeter, and Althaus with those from the deliberative polling method of Fishkin on the same sample of respondents. Contrary to low-information rationality perspectives, both methods indicate that across a range of prominent policy domains, level of political awareness has a strong impact on the expressed preferences of individuals. And while self-canceling across respondents tends to translate these individual-level influences into only rather modest effects in the aggregate, on a significant minority of issues, substantial shifts in collective opinion remain. The broad similarity of the estimates produced by these two very different methods, in addition to their convergence with previous studies of information effects, lends some simultaneous support to the validity and reliability of both approaches.

The authors present the results from parallel experiments in two surveys about privacy attitudes. Concerned that the order of the questions might affect the answers, they systematically varied the order of some of the key questions in the questionnaire. In both surveys, four of five question order experiments produced significant effects on responses to the items involved, but the question order variables did not affect responses to attitude items that came later in the questionnaire or relations between the attitude items whose order the authors varied and background characteristics of the respondents. They were also able to determine for most respondents whether their household had mailed back its census questionnaire. Question order did not have a consistent impact on the correlations between the survey responses and actual census returns. These results suggest that question order effects may be common with conceptually related items but that their impact is generally local, affecting answers to the items themselves but not answers to later questions, correlations with respondent background characteristics, or relations to subsequent behaviors.

This article considers modeling response bias when the response is a count. The authors adopt a “structural approach” by using a generalized negative binomial mixture of Poisson distribution to model misreported counts, assuming that the distribution of the true response follows a negative binomial distribution. The model may be interpreted as a “stopped-sum” model. A simulated maximum likelihood estimator is proposed, and its finite sample performance is investigated through Monte Carlo simulations. The approach is then applied to analyzing school victimization data drawn from the National Crime Victimization Survey, which allows the authors to identify the individual- and school-related characteristics that could contribute to school crime victimization and to the possible biases on the reported number of repeat victimizations. The authors find that for the reported number of thefts, about 12 percent of respondents overreport the numbers, most of whom actually have not had any item stolen at school.