Surveys are commonly tailored to produce estimates of aggregate statistics with a desired level of precision. This may lead to very small sample sizes for subpopulations of interest, defined geographically or by content, which are not incorporated into the survey design. We refer to subpopulations where the sample size is too small to provide direct estimates with adequate precision as small areas or small domains. Despite the small sample sizes, reliable small area estimates are needed for economic and political decision making. Hence, model-based estimation techniques are used which increase the effective sample size by borrowing strength from other areas to provide accurate information for small areas. The paragraph above introduced small area estimation as a field of survey statistics where two conflicting philosophies of statistical inference meet: the design-based and the model-based approach. While the first approach is well suited for the precise estimation of aggregate statistics, the latter approach furnishes reliable small area estimates. In most applications, estimates for both large and small domains based on the same sample are needed. This poses a challenge to the survey planner, as the sampling design has to reflect different and potentially conflicting requirements simultaneously. In order to enable efficient design-based estimates for large domains, the sampling design should incorporate information related to the variables of interest. This may be achieved using stratification or sampling with unequal probabilities. Many model-based small area techniques require an ignorable sampling design such that after conditioning on the covariates the variable of interest does not contain further information about the sample membership. If this condition is not fulfilled, biased model-based estimates may result, as the model which holds for the sample is different from the one valid for the population. Hence, an optimisation of the sampling design without investigating the implications for model-based approaches will not be sufficient. Analogously, disregarding the design altogether and focussing only on the model is prone to failure as well. Instead, a profound knowledge of the interplay between the sample design and statistical modelling is a prerequisite for implementing an effective small area estimation strategy. In this work, we concentrate on two approaches to address this conflict. Our first approach takes the sampling design as given and can be used after the sample has been collected. It amounts to incorporate the survey design into the small area model to avoid biases stemming from informative sampling. Thus, once a model is validated for the sample, we know that it holds for the population as well. We derive such a procedure under a lognormal mixed model, which is a popular choice when the support of the dependent variable is limited to positive values. Besides, we propose a three pillar strategy to select the additional variable accounting for the design, based on a graphical examination of the relationship, a comparison of the predictive accuracy of the choices and a check regarding the normality assumptions.rnrnOur second approach to deal with the conflict is based on the notion that the design should allow applying a wide variety of analyses using the sample data. Thus, if the use of model-based estimation strategies can be anticipated before the sample is drawn, this should be reflected in the design. The same applies for the estimation of national statistics using design-based approaches. Therefore, we propose to construct the design such that the sampling mechanism is non-informative but allows for precise design-based estimates at an aggregate level.
Traditionell werden Zufallsstichprobenerhebungen so geplant, dass nationale Statistiken zuverlässig mit einer adäquaten Präzision geschätzt werden können. Hierbei kommen vorrangig designbasierte, Modell-unterstützte (engl. model assisted) Schätzmethoden zur Anwendung, die überwiegend auf asymptotischen Eigenschaften beruhen. Für kleinere Stichprobenumfänge, wie man sie für Small Areas (Domains bzw. Subpopulationen) antrifft, eignen sich diese Schätzmethoden eher nicht, weswegen für diese Anwendung spezielle modellbasierte Small Area-Schätzverfahren entwickelt wurden. Letztere können zwar Verzerrungen aufweisen, besitzen jedoch häufig einen kleineren mittleren quadratischen Fehler der Schätzung als dies für designbasierte Schätzer der Fall ist. Den Modell-unterstützten und modellbasierten Methoden ist gemeinsam, dass sie auf statistischen Modellen beruhen; allerdings in unterschiedlichem Ausmass. Modell-unterstützte Verfahren sind in der Regel so konstruiert, dass der Beitrag des Modells bei sehr grossen Stichprobenumfängen gering ist (bei einer Grenzwertbetrachtung sogar wegfällt). Bei modellbasierten Methoden nimmt das Modell immer eine tragende Rolle ein, unabhängig vom Stichprobenumfang. Diese Überlegungen veranschaulichen, dass das unterstellte Modell, präziser formuliert, die Güte der Modellierung für die Qualität der Small Area-Statistik von massgeblicher Bedeutung ist. Wenn es nicht gelingt, die empirischen Daten durch ein passendes Modell zu beschreiben und mit den entsprechenden Methoden zu schätzen, dann können massive Verzerrungen und / oder ineffiziente Schätzungen resultieren.
Die vorliegende Arbeit beschäftigt sich mit der zentralen Frage der Robustheit von Small Area-Schätzverfahren. Als robust werden statistische Methoden dann bezeichnet, wenn sie eine beschränkte Einflussfunktion und einen möglichst hohen Bruchpunkt haben. Vereinfacht gesprochen zeichnen sich robuste Verfahren dadurch aus, dass sie nur unwesentlich durch Ausreisser und andere Anomalien in den Daten beeinflusst werden. Die Untersuchung zur Robustheit konzentriert sich auf die folgenden Modelle bzw. Schätzmethoden:
i) modellbasierte Schätzer für das Fay-Herriot-Modell (Fay und Herrot, 1979, J. Amer. Statist. Assoc.) und das elementare Unit-Level-Modell (vgl. Battese et al., 1988, J. Amer. Statist. Assoc.).
ii) direkte, Modell-unterstützte Schätzer unter der Annahme eines linearen Regressionsmodells.
Das Unit-Level-Modell zur Mittelwertschätzung beruht auf einem linearen gemischten Gauss'schen Modell (engl. mixed linear model, MLM) mit blockdiagonaler Kovarianzmatrix. Im Gegensatz zu bspw. einem multiplen linearen Regressionsmodell, besitzen MLM-Modelle keine nennenswerten Invarianzeigenschaften, so dass eine Kontamination der abhängigen Variablen unvermeidbar zu verzerrten Parameterschätzungen führt. Für die Maximum-Likelihood-Methode kann die resultierende Verzerrung nahezu beliebig groß werden. Aus diesem Grund haben Richardson und Welsh (1995, Biometrics) die robusten Schätzmethoden RML 1 und RML 2 entwickelt, die bei kontaminierten Daten nur eine geringe Verzerrung aufweisen und wesentlich effizienter sind als die Maximum-Likelihood-Methode. Eine Abwandlung von Methode RML 2 wurde Sinha und Rao (2009, Canad. J. Statist.) für die robuste Schätzung von Unit-Level-Modellen vorgeschlagen. Allerdings erweisen sich die gebräuchlichen numerischen Verfahren zur Berechnung der RML-2-Methode (dies gilt auch für den Vorschlag von Sinha und Rao) als notorisch unzuverlässig. In dieser Arbeit werden zuerst die Konvergenzprobleme der bestehenden Verfahren erörtert und anschließend ein numerisches Verfahren vorgeschlagen, das sich durch wesentlich bessere numerische Eigenschaften auszeichnet. Schließlich wird das vorgeschlagene Schätzverfahren im Rahmen einer Simulationsstudie untersucht und anhand eines empirischen Beispiels zur Schätzung von oberirdischer Biomasse in norwegischen Kommunen illustriert.
Das Modell von Fay-Herriot kann als Spezialfall eines MLM mit blockdiagonaler Kovarianzmatrix aufgefasst werden, obwohl die Varianzen des Zufallseffekts für die Small Areas nicht geschätzt werden müssen, sondern als bereits bekannte Größen betrachtet werden. Diese Eigenschaft kann man sich nun zunutze machen, um die von Sinha und Rao (2009) vorgeschlagene Robustifizierung des Unit-Level-Modells direkt auf das Fay-Herriot Model zu übertragen. In der vorliegenden Arbeit wird jedoch ein alternativer Vorschlag erarbeitet, der von der folgenden Beobachtung ausgeht: Fay und Herriot (1979) haben ihr Modell als Verallgemeinerung des James-Stein-Schätzers motiviert, wobei sie sich einen empirischen Bayes-Ansatz zunutze machen. Wir greifen diese Motivation des Problems auf und formulieren ein analoges robustes Bayes'sches Verfahren. Wählt man nun in der robusten Bayes'schen Problemformulierung die ungünstigste Verteilung (engl. least favorable distribution) von Huber (1964, Ann. Math. Statist.) als A-priori-Verteilung für die Lokationswerte der Small Areas, dann resultiert als Bayes-Schätzer [=Schätzer mit dem kleinsten Bayes-Risk] die Limited-Translation-Rule (LTR) von Efron und Morris (1971, J. Amer. Statist. Assoc.). Im Kontext der frequentistischen Statistik kann die Limited-Translation-Rule nicht verwendet werden, weil sie (als Bayes-Schätzer) auf unbekannten Parametern beruht. Die unbekannten Parameter können jedoch nach dem empirischen Bayes-Ansatz an der Randverteilung der abhängigen Variablen geschätzt werden. Hierbei gilt es zu beachten (und dies wurde in der Literatur vernachlässigt), dass die Randverteilung unter der ungünstigsten A-priori-Verteilung nicht einer Normalverteilung entspricht, sondern durch die ungünstigste Verteilung nach Huber (1964) beschrieben wird. Es ist nun nicht weiter erstaunlich, dass es sich bei den Maximum-Likelihood-Schätzern von Regressionskoeffizienten und Modellvarianz unter der Randverteilung um M-Schätzer mit der Huber'schen psi-Funktion handelt.
Unsere theoriegeleitete Herleitung von robusten Schätzern zum Fay-Herriot-Modell zeigt auf, dass bei kontaminierten Daten die geschätzte LTR (mit Parameterschätzungen nach der M-Schätzmethodik) optimal ist und, dass die LTR ein integraler Bestandteil der Schätzmethodik ist (und nicht als ``Zusatz'' o.Ä. zu betrachten ist, wie dies andernorts getan wird). Die vorgeschlagenen M-Schätzer sind robust bei Vorliegen von atypischen Small Areas (Ausreissern), wie dies auch die Simulations- und Fallstudien zeigen. Um auch Robustheit bei Vorkommen von einflussreichen Beobachtungen in den unabhängigen Variablen zu erzielen, wurden verallgemeinerte M-Schätzer (engl. generalized M-estimator) für das Fay-Herriot-Modell entwickelt.
Data fusions are becoming increasingly relevant in official statistics. The aim of a data fusion is to combine two or more data sources using statistical methods in order to be able to analyse different characteristics that were not jointly observed in one data source. Record linkage of official data sources using unique identifiers is often not possible due to methodological and legal restrictions. Appropriate data fusion methods are therefore of central importance in order to use the diverse data sources of official statistics more effectively and to be able to jointly analyse different characteristics. However, the literature lacks comprehensive evaluations of which fusion approaches provide promising results for which data constellations. Therefore, the central aim of this thesis is to evaluate a concrete plethora of possible fusion algorithms, which includes classical imputation approaches as well as statistical and machine learning methods, in selected data constellations.
To specify and identify these data contexts, data and imputation-related scenario types of a data fusion are introduced: Explicit scenarios, implicit scenarios and imputation scenarios. From these three scenario types, fusion scenarios that are particularly relevant for official statistics are selected as the basis for the simulations and evaluations. The explicit scenarios are the fulfilment or violation of the Conditional Independence Assumption (CIA) and varying sample sizes of the data to be matched. Both aspects are likely to have a direct, that is, explicit, effect on the performance of different fusion methods. The summed sample size of the data sources to be fused and the scale level of the variable to be imputed are considered as implicit scenarios. Both aspects suggest or exclude the applicability of certain fusion methods due to the nature of the data. The univariate or simultaneous, multivariate imputation solution and the imputation of artificially generated or previously observed values in the case of metric characteristics serve as imputation scenarios.
With regard to the concrete plethora of possible fusion algorithms, three classical imputation approaches are considered: Distance Hot Deck (DHD), the Regression Model (RM) and Predictive Mean Matching (PMM). With Decision Trees (DT) and Random Forest (RF), two prominent tree-based methods from the field of statistical learning are discussed in the context of data fusion. However, such prediction methods aim to predict individual values as accurately as possible, which can clash with the primary objective of data fusion, namely the reproduction of joint distributions. In addition, DT and RF only comprise univariate imputation solutions and, in the case of metric variables, artificially generated values are imputed instead of real observed values. Therefore, Predictive Value Matching (PVM) is introduced as a new, statistical learning-based nearest neighbour method, which could overcome the distributional disadvantages of DT and RF, offers a univariate and multivariate imputation solution and, in addition, imputes real and previously observed values for metric characteristics. All prediction methods can form the basis of the new PVM approach. In this thesis, PVM based on Decision Trees (PVM-DT) and Random Forest (PVM-RF) is considered.
The underlying fusion methods are investigated in comprehensive simulations and evaluations. The evaluation of the various data fusion techniques focusses on the selected fusion scenarios. The basis for this is formed by two concrete and current use cases of data fusion in official statistics, the fusion of EU-SILC and the Household Budget Survey on the one hand and of the Tax Statistics and the Microcensus on the other. Both use cases show significant differences with regard to different fusion scenarios and thus serve the purpose of covering a variety of data constellations. Simulation designs are developed from both use cases, whereby the explicit scenarios in particular are incorporated into the simulations.
The results show that PVM-RF in particular is a promising and universal fusion approach under compliance with the CIA. This is because PVM-RF provides satisfactory results for both categorical and metric variables to be imputed and also offers a univariate and multivariate imputation solution, regardless of the scale level. PMM also represents an adequate fusion method, but only in relation to metric characteristics. The results also imply that the application of statistical learning methods is both an opportunity and a risk. In the case of CIA violation, potential correlation-related exaggeration effects of DT and RF, and in some cases also of RM, can be useful. In contrast, the other methods induce poor results if the CIA is violated. However, if the CIA is fulfilled, there is a risk that the prediction methods RM, DT and RF will overestimate correlations. The size ratios of the studies to be fused in turn have a rather minor influence on the performance of fusion methods. This is an important indication that the larger dataset does not necessarily have to serve as a donor study, as was previously the case.
The results of the simulations and evaluations provide concrete implications as to which data fusion methods should be used and considered under the selected data and imputation constellations. Science in general and official statistics in particular benefit from these implications. This is because they provide important indications for future data fusion projects in order to assess which specific data fusion method could provide adequate results along the data constellations analysed in this thesis. Furthermore, with PVM this thesis offers a promising methodological innovation for future data fusions and for imputation problems in general.
This dissertation deals with consistent estimates in household surveys. Household surveys are often drawn via cluster sampling, with households sampled at the first stage and persons selected at the second stage. The collected data provide information for estimation at both the person and the household level. However, consistent estimates are desirable in the sense that the estimated household-level totals should coincide with the estimated totals obtained at the person-level. Current practice in statistical offices is to use integrated weighting. In this approach consistent estimates are guaranteed by equal weights for all persons within a household and the household itself. However, due to the forced equality of weights, the individual patterns of persons are lost and the heterogeneity within households is not taken into account. In order to avoid the negative consequences of integrated weighting, we propose alternative weighting methods in the first part of this dissertation that ensure both consistent estimates and individual person weights within a household. The underlying idea is to limit the consistency conditions to variables that emerge in both the personal and household data sets. These common variables are included in the person- and household-level estimator as additional auxiliary variables. This achieves consistency more directly and only for the relevant variables, rather than indirectly by forcing equal weights on all persons within a household. Further decisive advantages of the proposed alternative weighting methods are that original individual rather than the constructed aggregated auxiliaries are utilized and that the variable selection process is more flexible because different auxiliary variables can be incorporated in the person-level estimator than in the household-level estimator.
In the second part of this dissertation, the variances of a person-level GREG estimator and an integrated estimator are compared in order to quantify the effects of the consistency requirements in the integrated weighting approach. One of the challenges is that the estimators to be compared are of different dimensions. The proposed solution is to decompose the variance of the integrated estimator into the variance of a reduced GREG estimator, whose underlying model is of the same dimensions as the person-level GREG estimator, and add a constructed term that captures the effects disregarded by the reduced model. Subsequently, further fields of application for the derived decomposition are proposed such as the variable selection process in the field of econometrics or survey statistics.
In politics and economics, and thus in the official statistics, the precise estimation of indicators for small regions or parts of populations, the so-called Small Areas or domains, is discussed intensively. The design-based estimation methods currently used are mainly based on asymptotic properties and are thus reliable for large sample sizes. With small sample sizes, however, this design based considerations often do not apply, which is why special model-based estimation methods have been developed for this case - the Small Area methods. While these may be biased, they often have a smaller mean squared error (MSE) as the unbiased design based estimators. In this work both classic design-based estimation methods and model-based estimation methods are presented and compared. The focus lies on the suitability of the various methods for their use in official statistics. First theory and algorithms suitable for the required statistical models are presented, which are the basis for the subsequent model-based estimators. Sampling designs are then presented apt for Small Area applications. Based on these fundamentals, both design-based estimators and as well model-based estimation methods are developed. Particular consideration is given in this case to the area-level empirical best predictor for binomial variables. Numerical and Monte Carlo estimation methods are proposed and compared for this analytically unsolvable estimator. Furthermore, MSE estimation methods are proposed and compared. A very popular and flexible resampling method that is widely used in the field of Small Area Statistics, is the parametric bootstrap. One major drawback of this method is its high computational intensity. To mitigate this disadvantage, a variance reduction method for parametric bootstrap is proposed. On the basis of theoretical considerations the enormous potential of this proposal is proved. A Monte Carlo simulation study shows the immense variance reduction that can be achieved with this method in realistic scenarios. This can be up to 90%. This actually enables the use of parametric bootstrap in applications in official statistics. Finally, the presented estimation methods in a large Monte Carlo simulation study in a specific application for the Swiss structural survey are examined. Here problems are discussed, which are of high relevance for official statistics. These are in particular: (a) How small can the areas be without leading to inappropriate or to high precision estimates? (b) Are the accuracy specifications for the Small Area estimators reliable enough to use it for publication? (c) Do very small areas infer in the modeling of the variables of interest? Could they cause thus a deterioration of the estimates of larger and therefore more important areas? (d) How can covariates, which are in different levels of aggregation be used in an appropriate way to improve the estimates. The data basis is the Swiss census of 2001. The main results are that in the author- view, the use of small area estimators for the production of estimates for areas with very small sample sizes is advisable in spite of the modeling effort. The MSE estimates provide a useful measure of precision, but do not reach in all Small Areas the level of reliability of the variance estimates for design-based estimators.
Statistical matching offers a way to broaden the scope of analysis without increasing respondent burden and costs. These would result from conducting a new survey or adding variables to an existing one. Statistical matching aims at combining two datasets A and B referring to the same target population in order to analyse variables, say Y and Z, together, that initially were not jointly observed. The matching is performed based on matching variables X that correspond to common variables present in both datasets A and B. Furthermore, Y is only observed in B and Z is only observed in A. To overcome the fact that no joint information on X, Y and Z is available, statistical matching procedures have to rely on suitable assumptions. Therefore, to yield a theoretical foundation for statistical matching, most procedures rely on the conditional independence assumption (CIA), i.e. given X, Y is independent of Z.
The goal of this thesis is to encompass both the statistical matching process and the analysis of the matched dataset. More specifically, the aim is to estimate a linear regression model for Z given Y and possibly other covariates in data A. Since the validity of the assumptions underlying the matching process determine the validity of the obtained matched file, the accuracy of statistical inference is determined by the suitability of the assumptions. By putting the focus on these assumptions, this work proposes a systematic categorisation of approaches to statistical matching by relying on graphical representations in form of directed acyclic graphs. These graphs are particularly useful in representing dependencies and independencies which are at the heart of the statistical matching problem. The proposed categorisation distinguishes between (a) joint modelling of the matching and the analysis (integrated approach), and (b) matching subsequently followed by statistical analysis of the matched dataset (classical approach). Whereas the classical approach relies on the CIA, implementations of the integrated approach are only valid if they converge, i.e. if the specified models are identifiable and, in the case of MCMC implementations, if the algorithm converges to a proper distribution.
In this thesis an implementation of the integrated approach is proposed, where the imputation step and the estimation step are jointly modelled through a fully Bayesian MCMC estimation. It is based on a linear regression model for Z given Y and accounts for both a linear regression model and a random effects model for Y. Furthermore, it yields its validity when the instrumental variable assumption (IVA) holds. The IVA corresponds to: (a) Z is independent of a subset X’ of X given Y and X*, where X* = X\X’ and (b) Y is correlated with X’ given X*. The proof, that the joint Bayesian modelling of both the model for Z and the model for Y through an MCMC simulation converges to a proper distribution is provided in this thesis. In a first model-based simulation study, the proposed integrated Bayesian procedure is assessed with regard to the data situation, convergence issues, and underlying assumptions. Special interest lies in the investigation of the interplay of the Y and the Z model within the imputation process. It turns out that failure scenarios can be distinguished by comparing the CIA and the IVA in the completely observed dataset.
Finally, both approaches to statistical matching, i.e. the classical approach and the integrated approach, are subject to an extensive comparison in (1) a model-based simulation study and (2) a simulation study based on the AMELIA dataset, which is an openly available very large synthetic dataset and, by construction, similar to the EU-SILC survey. As an additional integrated approach, a Bayesian additive regression trees (BART) model is considered for modelling Y. These integrated procedures are compared to the classical approach represented by predictive mean matching in the form of multiple imputations by chained equation. Suitably chosen, the first simulation framework offers the possibility to clarify aspects related to the underlying assumptions by comparing the IVA and the CIA and by evaluating the impact of the matching variables. Thus, within this simulation study two related aspects are of special interest: the assumptions underlying each method and the incorporation of additional matching variables. The simulation on the AMELIA dataset offers a close-to-reality framework with the advantage of knowing the whole setting, i.e. the whole data X, Y and Z. Special interest lies in investigating assumptions through adding and excluding auxiliary variables in order to enhance conditional independence and assess the sensitivity of the methods to this issue. Furthermore, the benefit of having an overlap of units in data A and B for which information on X, Y, Z is available is investigated. It turns out that the integrated approach yields better results than the classical approach when the CIA clearly does not hold. Moreover, even when the classical approach obtains unbiased results for the regression coefficient of Y in the model for Z, it is the method relying on BART that over all coefficients performs best.
Concluding, this work constitutes a major contribution to the clarification of assumptions essential to any statistical matching procedure. By introducing graphical models to identify existing approaches to statistical matching combined with the subsequent analysis of the matched dataset, it offers an extensive overview, categorisation and extension of theory and application. Furthermore, in a setting where none of the assumptions are testable (since X, Y and Z are not observed together), the integrated approach is a valuable asset by offering an alternative to the CIA.