I föreliggande rapport redovisas resultaten från den andra, och avslutande, fasen av den metodstudie som genomförts inom ramen för projektet Alternativa datainsamlingsmetoder för Undersökning av Levnadsförhållanden (ULF), som påbörjades 2002.

Metodstudiens huvudsakliga syfte var att jämföra två olika metoder för datainsamling: en mixed mode ansats (MM) med en kombination av insamlingsmetoderna besök- och telefonintervju utan datorstöd respektive datorstödd telefonintervju (CATI). Jämförelser avseende huvudsakligen fyra olika kvalitetsaspekter redovisas: (1) mätkvaliteten, (2) bortfallsfelets storlek och inverkan på skattningar, (3) svarsandelen i Barn-ULF och (4) viljan hos uppgiftslämnarna för att delta i undersökningen.

Den allmänna slutsatsen från studien är att den systematiska felkomponenten (mät- och bortfallsfel) i tillförlitlighetskomponenten bedöms vara oförändrad vid en övergång till CATI. I kombination med att övergången skulle frigöra resurser för en urvalsökning, innebär detta att medelkvadratfelet (MSE) i skattningarna skulle minska vid en övergång från den tidigare ansatsen med MM till CATI som primär insamlingsmetod.

Per-record measures of disclosure risk have potential uses in statistical disclosure control programs as a means of identifying sensitive or atypical records in public-use microdata files. A measure intended for sample data based on the Poisson-inverse Gaussian distribution and overdispersed log-linear modeling is presented. An empirical example indicates that the proposed model performs approximately as well as the Poisson-lognormal model of Skinner and Holmes (1998) and may be a tractable alternative as the required computational effort is significantly smaller. It is also demonstrated how to extend the application to take into account population level information. The empirical results indicate that using population level information sharpens the risk measure.

With the release of public-use microdata filles it is important to assess the risk of disclosing individual information. A measure of disclosure risk often considered in the literature is the proportion of unique records in the file that are also unique in the population. Various methods based on superpopulation models have been proposed for estimating this quantity using sample data. An empirical comparison of a selection of models applied to three real-life data sets is presented. The general conclusion is that no one model is uniformly best with respect to the risk measure used and that performance varies greatly between di¤erent types of data.

An important measure of identification risk associated with the release of microdata or large complex tables is the number or proportion of population units that can be uniquely identified by some set of characterizing attributes which partition the population into subpopulations or cells. Various methods for estimating this quantity based on sample data have been proposed in the literature by means of superpopulation models. In the present paper the Poisson- inverse Gaussian (PiG) distribution is proposed as a possible approach within this context. Disclosure risk measures are discussed and derived under the proposed model as are various methods of estimation. An example on real data is given and the results indicate that the PiG model may be a useful alternative to other models.

Denna rapport är en redovisning av en reliabilitetsanalys som gjorts på uppdrag av Grosin & McNamara HB, avseende ett mätinstrument för att bedöma det pedagogiska och sociala klimatet (PESOK) i skolor. Mätinstrumentet är i form av ett webbaserat frågeformulär och har utvecklats av uppdragsgivarna, se McNamara (1999, 2004). De items som bygger upp instrumentet har a priori delats in i grupper av items för att aggregera till s.k. faktormedelvärden. Den metod som huvudsakligen har använts för att skatta reliabilitetskoefficienterna är Cronbach’s alfa och besläktade mått. Principalkomponentanalys och explorativ faktoranalys har kompletterat analysen tillsammans med annan beskrivande statistik; de förra metoderna har avsett att undersöka antalet latenta faktorer bakom grupperna av items.

Den huvudsakliga slutsatsen är att de olika grupperingarna i faktorer uppvisar relativt sett goda egenskaper med avseende på reliabilitet med undantag för enstaka grupper. Vissa metodologiska och teoretiska problem har också noterats. Bland annat att data har en hierarkisk struktur med individer vid skolorna som uppgiftslämnare. Dessa individdata aggregeras sedan till medelvärden för hela skolan som sedan är tänkta att användas som kollektiva egenskaper för de enskilda skolorna och det är dessa aggregerade värdens egenskaper ur relibilitetsperspektiv som har varit föremål för analysen.

Statistics Sweden has identified a need to increase the level of knowl-edge in the SDC field among methodologists and also to develop a unified view and strategy on application of SDC methodology. A project was thus initiated with the goal to spread knowledge but also to collate current practices and ex-periences within the agency. The project resulted in a course on SDC given at Statistics Sweden and this paper provides a brief first report on the work so far, reviewing the background and describing the course and possible future activities. Furthermore, several practical examples were discussed during the course, two of which are briefly described here; the Industrial production index and Structure of wages and earnings in the private sector.

This book comprises articles on the methodology and applications of official statistics written by some of the best scholars and methodologists in the Nordic-Baltic region to celebrate Professor Daniel THorburn on his 65th birthday. The book's eighteen chapters span a wide spectrum of topics in official statistics, including its foundational aspects, sampling designs, survey methodology, index construction, and methods for registers and adminstrative data.

A data-swapping technique based on ranks for generating synthetic data sets suitable for dissemination is presented. The method is mainly intended to be applied to quantative data down to at least an ordinal scale although it should be applicable in other cases as well. Generalizations and variations of the proposed method are briefly discussed. The expected performance is investigated by means of simulation studies and an analytical result is stated. The results from the simulations studies indicate that the proposed method performs reasonably well in the bivariate normal case when the correlation coefficient is used as the measure of association between pairs of variables.

A data-swapping technique based on ranks is described and suggested as a possible approach to statistical disclosure control. The proposed method is intended to be applied to quantitative data and utilises the rank structure of disjoint subsets of an original data set; values of one subset are exchanged for values of other subsets. The procedure retains the validity of a sample on an intra-variate level but the association between pairs of variables is typically weakened. Theoretical and simulation results indicate that the proposed method performs reasonably well in the bivariate normal case.

Translation of own norms and perception of parents' and friends' values into actual behaviour were studied in the context of the theory of reasoned action. A LISREL model based on this theory was constructed, and the fit and generalizability of the model tested across gender and time. The same norm questionnaire assessing adolescents' own values, their own intentions and actual behavior and perception of parents' and friends' was administered to two samples of approximately 15-year-old adolescents on two occasions (in 1969 and in 1995, in the same town). A LISREL model with an acceptabel fit was obtained, and its structural parameters indicated a strong relationship between adolescents' own evaluations, own intentions and actual behavior. The model was also tenable acorss gender and time. No significant differences were found between groups across gender and time with regard to the four parameters of the structural part of the model which indicate the relationships bwtween the latent variables.

It follows from Swedish legislation that Statistics Sweden, when publishing statistical data, must first assess the risk of disclosing confidential information, and then assess the risk that disclosure will result in any harm or damage. In case of considerable risk of the latter, data should be properly protected. We are attempting to handle this by a simple process describing how disclosure control should be applied within Statistics Sweden. Following the general process orientation of statistics production within Statistics Sweden, the process of statistical disclosure control (SDC) in tables must fit within the general production process, and the recommended methods and tools used should as much as possible be shared by all statistical products. We will describe the process for SDC in tables, the contents of a handbook where the process and appropriate methods are described, and the ongoing work with IT-solutions. Presently, work is concentrated on making it possible to run τ-Argus from SAS. Plans for implementation and training will also be briefly described.

Statistics Sweden is currently investigating the possibility of coordinating the production of social statistics in order to find solutions that fulfill internal, national and international requirements and needs. The coordinative work is conducted within the Population and Welfare Department as a long-term development project called ENAR. The overall goal of the project is to create a social statistical system that provides a coherent image of society that is based on coordinated statistical production to the full extent possible. This paper provides a brief account of completed and on-going work.

The literature presents different approaches to measure information loss in data protection. The probabilistic information loss measure is based on comparison of means, variances, covariance, correlation coefficients and quantiles. In a recent paper the Hellinger distance has also been used to measure information loss. This distance has been used to compare univariate probability distributions. In this paper we will discuss the use of the Hellinger distance to measure information loss, and compare its results with PIL measures. A set of experiments using numerical files protected using a few data protection methods will be presented.

Hellinger distance is a distance between two additive measures defined in terms of the Radon-Nikodym derivative of these two measures. This measure proposed in 1909 has been used in a large variety of contexts. In this paper we define an analogous measure for fuzzy measures. We discuss them for distorted probabilities and give two examples.