Journal of Official Statistics

We investigate disproportionate stratified sampling as a possibly efficient method of surveying members of a rare domain in circumstances in which there is no acceptable list of members. In this work, we assume that information is available at the sampling stage to stratify the general-population sampling frame into high- and low-density strata. Under a fixed constraint on the variance of the estimator of the domain mean, we make the optimum allocation of sample size to the several strata and show that, in comparison to proportional allocation, the optimum allocation requires (a) a smaller total sample size but (b) a larger number of interviews of members of the rare domain. We illustrate the methods using information about American consumers maintained by market-research companies. Such companies are able to develop lists of households that are thought to have a defined attribute of interest, such as at least one resident in a user-specified age range. The lists are imperfect, with false positives and negatives. We apply an age-targeted list to the National Immunization Survey (NIS), conducted by the Centers for Disease Control and Prevention, which targets the relatively rare population of children age 19–35 months. The age-targeted list comprises the high-density stratum and the rest of the survey’s sampling frame comprises the low-density stratum. Given the optimum allocation, we demonstrate potential cost savings for the NIS in excess of ten percent.

Response burden in business surveys has long been a concern for National Statistical Institutes (NSIs) for three types of reasons: political reasons, because response burden is part of the total administrative burden governments impose on businesses; methodological reasons, because an excessive response burden may reduce data quality and increase data-collection costs; and strategic reasons, because it affects relations between the NSIs and the business community. This article investigates NSI practices concerning business response burden measurement and reduction actions based on a survey of 41 NSIs from 39 countries. Most NSIs monitor at least some burden aspects and have implemented some actions to reduce burden, but large differences exist between NSIs’ methodologies for burden measurement and actions taken to reduce burden. Future research should find ways to deal with methodological differences in burden conceptualization, operationalization, and measurement, and provide insights into the effectiveness and efficiency of burden-reduction actions.

Using as much administrative data as possible is a general trend among most national statistical institutes. Different kinds of administrative sources, from tax authorities or other administrative bodies, are very helpful material in the production of business statistics. However, these sources often have to be completed by information collected through statistical surveys. This article describes the way Insee has implemented such a strategy in order to produce French structural business statistics. The originality of the French procedure is that administrative and survey variables are used jointly for the same enterprises, unlike the majority of multisource systems, in which the two kinds of sources generally complement each other for different categories of units. The idea is to use, as much as possible, the richness of the administrative sources combined with the timeliness of a survey, even if the latter is conducted only on a sample of enterprises. One main issue is the classification of enterprises within the NACE nomenclature, which is a cornerstone variable in producing the breakdown of the results by industry. At a given date, two values of the corresponding code may coexist: the value of the register, not necessarily up to date, and the value resulting from the data collected via the survey, but only from a sample of enterprises. Using all this information together requires the implementation of specific statistical estimators combining some properties of the difference estimators with calibration techniques. This article presents these estimators, as well as their statistical properties, and compares them with those of other methods.

Survey nonresponse may increase the chances of nonresponse error, and different interviewers contribute differentially to nonresponse. This article first addresses the relationship between initial impressions of interviewers in survey introductions and the outcome of these introductions, and then contrasts this relationship with current viewpoints and practices in telephone interviewing. The first study described here exposed judges to excerpts of interviewer speech from actual survey introductions and asked them to rate twelve characteristics of the interviewer. Impressions of positive traits such as friendliness and confidence had no association with the actual outcome of the call, while higher ratings of “scriptedness” predicted lower participation likelihood. At the same time, a second study among individuals responsible for training telephone interviewers found that when training interviewers, sounding natural or unscripted during a survey introduction is not emphasized. This article concludes with recommendations for practice and further research.

In this article we propose a methodology for estimating the GDP of a country’s different regions, providing quarterly profiles for the annual official observed data. Thus the article offers a new instrument for short-term monitoring that allows the analysts to quantify the degree of synchronicity among regional business cycles. Technically, we combine time-series models with benchmarking methods to process short-term quarterly indicators and to estimate quarterly regional GDPs ensuring their temporal and transversal consistency with the National Accounts data. The methodology addresses the issue of nonadditivity, explicitly taking into account the transversal constraints imposed by the chain-linked volume indexes used by the National Accounts, and provides an efficient combination of structural as well as short-term information. The methodology is illustrated by an application to the Spanish economy, providing real-time quarterly GDP estimates, that is, with a minimum compilation delay with respect to the national quarterly GDP. The estimated quarterly data are used to assess the existence of cycles shared among the Spanish regions.

Sample coordination seeks to maximize or to minimize the overlap of two or more samples. The former is known as positive coordination, and the latter as negative coordination. Positive coordination is mainly used for estimation purposes and to reduce data collection costs. Negative coordination is mainly performed to diminish the response burden of the sampled units. Poisson sampling design with permanent random numbers provides an optimum coordination degree of two or more samples. The size of a Poisson sample is, however, random. Conditional Poisson (CP) sampling is a modification of the classical Poisson sampling that produces a fixed-size πps sample. We introduce two methods to coordinate Conditional Poisson samples over time or simultaneously. The first one uses permanent random numbers and the list-sequential implementation of CP sampling. The second method uses a CP sample in the first selection and provides an approximate one in the second selection because the prescribed inclusion probabilities are not respected exactly. The methods are evaluated using the size of the expected sample overlap, and are compared with their competitors using Monte Carlo simulation. The new methods provide a good coordination degree of two samples, close to the performance of Poisson sampling with permanent random numbers.

Privacy is an important feature of the interview interaction mainly due to its potential effect on reporting information, especially sensitive information. Here we examine the effect of third-party presence on reporting both sensitive and relatively neutral outcomes. We investigate whether the effect of third-party presence on reporting sensitive information is moderated by the respondent’s need for social conformity and the respondent’s country of residence. Three types of outcomes are investigated: behavioral, attitudinal, and relatively neutral health events. Using data from 22,070 interviews and nine countries in the cross-national World Mental Health Survey Initiative, we fit multilevel logistic regression to study reporting effects on questions about suicidal behavior and marital ratings, and contrast these with questions about having high blood pressure, asthma, or arthritis. We find that there is an effect of third-party presence on reporting sensitive information and no effect on reporting of neutral information. Further, the effect of the interview privacy setting on reporting sensitive information is moderated by the need for social conformity and the cultural setting.

The article presents central frameworks for guiding the development and improvement of population statistics. A shared understanding between producers and users of statistics is needed with regard to the concepts, data, processes, and outputs produced. In the United Kingdom, population estimates are produced by conducting decennial censuses and by estimating intercensus populations through the addition and subtraction of the demographic components of change derived from registers of vital events and from a combination of administrative data and surveys for internal and international migration. In addition, data cleaning, imputation, and modelling may be required to produce the desired population statistics. The frameworks presented in this paper are useful for aligning the required concepts of population statistics with the various sources of available data. Taken together, they provide a general ‘recipe’ for the continued improvement and expansion of official statistics on population and demographic change.

Link-tracing designs are often used to estimate the size of hidden populations by utilizing the relational links between their members. A major problem in studies of hidden populations is the lack of a convenient sampling frame. The most frequently applied design in studies of hidden populations is respondent-driven sampling in which no sampling frame is used. However, in some studies multiple but incomplete sampling frames are available. In this article, we introduce the B-graph design that can be used in such situations. In this design, all available incomplete sampling frames are joined and turned into one sampling frame, from which a random sample is drawn and selected respondents are asked to mention their contacts. By considering the population as a bipartite graph of a two-mode network (those from the sampling frame and those who are not on the frame), the number of respondents who are directly linked to the sampling frame members can be estimated using Chao’s and Zelterman’s estimators for sparse data. The B-graph sampling design is illustrated using the data of a social network study from Utrecht, the Netherlands.

Scientific- or public-use files are typically produced by applying anonymisation methods to the original data. Anonymised data should have both low disclosure risk and high data utility.

Data utility is often measured by comparing well-known estimates from original data and anonymised data, such as comparing their means, covariances or eigenvalues.

However, it is a fact that not every estimate can be preserved. Therefore the aim is to preserve the most important estimates, that is, instead of calculating generally defined utility measures, evaluation on context/data dependent indicators is proposed.

In this article we define such indicators and utility measures for the Structure of Earnings Survey (SES) microdata and proper guidelines for selecting indicators and models, and for evaluating the resulting estimates are given. For this purpose, hundreds of publications in journals and from national statistical agencies were reviewed to gain insight into how the SES data are used for research and which indicators are relevant for policy making.

Besides the mathematical description of the indicators and a brief description of the most common models applied to SES, four different anonymisation procedures are applied and the resulting indicators and models are compared to those obtained from the unmodified data. The disclosure risk is reported and the data utility is evaluated for each of the anonymised data sets based on the most important indicators and a model which is often used in practice.

Determining sample sizes in multistage samples requires variance components for each stage of selection. The relative sizes of the variance components in a cluster sample are dramatically affected by how much the clusters vary in size, by the type of sample design, and by the form of estimator used. Measures of the homogeneity of survey variables within clusters are related to the variance components and affect the numbers of sample units that should be selected at each stage to achieve the desired precision levels. Measures of homogeneity can be estimated using standard software for random-effects models but the model-based intracluster correlations may need to be transformed to be appropriate for use with the sample design. We illustrate these points and implications for sample size calculation for two-stage sample designs using a realistic population derived from household surveys and the decennial census in the U.S.

The problem of inference about the joint distribution of two categorical variables based on knowledge or observations of their marginal distributions, to be referred to as categorical data fusion in this paper, is relevant in statistical matching, ecological inference, market research, and several other related fields. This article organizes the use of proxy variables, to be distinguished from other auxiliary variables, both in terms of their effects on the uncertainty of fusion and the techniques of fusion. A measure of the gains of efficiency is provided, which incorporates both the identification uncertainty associated with data fusion and the sampling uncertainty that arises when the theoretical bounds of the uncertainty space are unknown and need to be estimated. Several existing techniques for generating fusion distributions (or datasets) are described and some new ones proposed. Analysis of real-life data demonstrates empirically that proxy variables can make data fusion more precise and the constructed fusion distribution more plausible.