“Biodiversity” is a particular term that brings complex ideas of the “abundance of life” to a scientific formula and at the same time marks it as a desirable state. Since the early 1990s, the meta-narrative “threatened biodiversity” has functioned as a meaningful moment of international environmental policy. It is well known that the attitude surveys on social awareness of biodiversity, which the “Bundesamt für Naturschutz” (Federal Agency for Nature Conservation) has had conducted every two years since 2009, are contaminated by social desirability. Since previous analyses either explicitly or implicitly assume that the form of bias is uniform, statements regarding associations and group comparisons are at risk. Causal machine learning methods open up a new analytical approach to the underlying statistical problem. They can be used to uncover how causal effects vary and which variables are related to this variation in such a way that they can explain effect differences. In our specific case, we use causal forest models to show that the influence of social desirability on attitude patterns and behavioral intentions about biodiversity varies systematically, with this effect heterogeneity being significantly associated with the age of the interviewees in a non-linear manner. Furthermore, we present this procedure as a methodological innovation for quality assurance in normatively charged survey topics.

Der Bericht „Jugendstudie Baden-Württemberg 2022 “ informiert bereits zum sechsten Mal über die Einstellungen der Jugend Baden-Württembergs zu Themen rund um Politik und Gesellschaft. Erstmals wurde die Studie von einem Team der Universität Stuttgart durchgeführt, sodass im Vergleich zu früheren Erhebungsrunden methodisch einige Neuerungen anzutreffen sind. Die Befragung von Neuntklässlerinnen und Neuntklässlern fand im Frühjahr 2022 statt; sie erfolgte damit unter dem Eindruck des Krieges in der Ukraine und während die Corona-Pandemie nach mittlerweile zwei Jahren noch immer unseren Alltag mitbestimmt.

Indirect questioning techniques are widely discussed and used as methods to avoid or reduce the effects of social desirability in interview situations on sensitive topics. Nevertheless, current evaluation studies suggest that indirect questioning techniques have a bigger compliance problem than evaluation studies based on the “more is-better” principle would suggest. In our study, we investigate the extent to which question compliance problems can be identified for a variant of the Randomized Response Technique, for the Crosswise Model and Triangular Model. By means of an aggregate and an individual level validation, we examine the response patterns of the participants. Contrary to the actual empirical application context of sensitive topics, we use a non-sensitive question that cannot be distorted by social desirability bias. The resulting “same-is-best” rationale differs from most evaluation studies to date, which work according to the “more” or “less-is-better” principle. Our analyses are based on the data of a convenience sample of 1277 students in the form of an online survey experiment. The results suggest that the indirect questioning techniques show substantial weaknesses in terms of compliance and encourage further individual level evaluations.

Response styles represent a challenge for empirical survey research. Response behavior which is not associated with the content of a survey question can not only influence proportional and average values but also distort model-based parameter estimation. In this pilot study we investigate whether machine learning (ML) methods can be used to construct an empirically based approach to a uniform identification of the most common response style patterns. This method could replace more complex control procedures, which can only find some patterns or are associated with considerable additional effort. Our approach could, therefore, be a more universal and practical option. The method presented here is based on the calibration of the ML model using synthesized data that correspond to a formal definition of response styles (RS) and a proportion of empirical data (European Social Survey) which is not affected by RS. The trained model can then be applied to empirically collected data in order to reliably detect and process RS patterns in survey data. With error classification rates of 0.3 to 3.5 %, the results of this study suggest that our ML-based response style detection is a promising alternative to previous methods.

Aus verschiedenen Perspektiven wird in dieser Festschrift zu Ehren von Prof. Dr. Dieter Urban das thematische Feld der Einstellung-Verhaltens-Forschung beleuchtet. Dies umfasst sowohl neuere theoretisch-analytische Zugänge als auch aktuelle methodisch-statistische Entwicklungen sowie methodologische Überlegungen zur Analyse von Einstellungen und sozialem Verhalten in der empirischen Sozialforschung. Der Band unterteilt sich in die Bereiche Theorie, Methodik, Analysestrategien und Längsschnittanalysen in unterschiedlichen inhaltlichen Anwendungsfeldern und deckt damit eine große Bandbreite der sozialwissenschaftlichen Einstellungs-Verhaltens-Forschung ab. Der Inhalt: Perspektiven zur empirisch-analytischen Soziologie • Theorien und Konzeptionen der Einstellungs- und Verhaltensforschung • Einstellungen und Verhalten in der Forschungspraxis: Erhebungsmethoden und Analyseverfahren • Längsschnittliche Analyseverfahren der Einstellungs- und Verhaltensforschung

This paper evaluates information criteria for class identification in multiple indicator growth mixture models. This identification of latent classes is evaluated under the simulation conditions of missing invariance of the measurement and different class distances. The data of this assessment is based on a Monte Carlo simulation study. This paper extends previous results by evaluating a broader range of IC-Indices. All previously proposed IC measures are examined for their ability to identify latent classes and compared with more recent proposals. These include the Hierarchical BIC, which explicitly takes the latent class structure into account. The evaluations of the simulation results show that the Hierarchical BIC has a completely different identification structure than all other IC-Indices and provides comparatively good results with small distances between the latent classes.

Multiple Indicator Growth Mixture Models (MIGMM) combine the design principles of latent measurement models, growth curve models, and latent class analysis. MIGMMs are thus analytical tools for empirical social research, which consider the measurements as latent constructs and simultaneously allow post-hoc identification and description of group differences with respect to temporal change. By identifying unobserved subpopulations, social change processes and their differences between and within the unobserved subpopulations can be investigated. While simple Growth Mixture Models, based on manifest variables, have already been evaluated in numerous Monte Carlo studies, a systematic analysis of the performance of multiple indicator GMMs is still lacking. This simulation study aims to systematically evaluate the performance of MIGMMs under different data situations, focusing in particular on temporal, group-specific and combined invariance violations of the latent measurement models. The following conditions were manipulated: measurement invariance of the latent constructs, class distance, mixing proportion and the number of observations. It is shown that especially the class distance between the latent growth groups and the violations of the measurement invariance between the latent groups have an influence on the performance of MIGMMs regarding class identification, coefficient estimates and standard error calculation. Based on these results, it is recommended to use MIGMMs only for large and distinct differences in the growth pattern and only for latent constructs with established reliability for group differences. The results of this study are intended to guide the research practice of empirical social research and provide recommendations for the use of this new analytic tool for longitudinal studies.

Dieser Beitrag veranschaulicht am Beispiel der sozialen Exklusion von Kindern und Jugendlichen das Analyseverfahren der “Multiple Indicator Growth Mixture Models” (MIGMMs). Diese, für Längsschnittdaten konzipierte, Analysemethode ermöglicht es latente Klassenentwicklungsverläufe, unter Kontrolle von Prädiktoren, explorativ zu identifizieren und darüber hinaus den Einfluss dieser Entwicklungsverläufe auf mögliche Folgen (distal outcomes) zu überprüfen. Die vorliegenden Resultate stehen teilweise im Widerspruch zu den, aus der Literatur in diesem Forschungsbereich stammenden Ergebnissen. Es hat sich gezeigt, dass die Entwicklungsverläufe der zwei identifizierten Klassen über die Zeit hinweg homogen bleiben und die üblicherweise vorgeschlagenen Risikofaktoren (Prädiktoren) keinen nennenswerten Einfluss auf die Klassenzugehörigkeit haben. Die Ergebnisse dieses Beitrags zeigen die Leistungsfähigkeit - aber auch Grenzen und mögliche Probleme - dieses dynamischen Ansatzes für die Einstellungsund Verhaltensforschung, welcher die Ergebnisse herkömmlicher Verfahren um zusätzliche Perspektiven bereichern kann.

The working paper summarizes the results of a vignette experiment (factorial survey) about threat perceptions in everyday situations. A student research project investigated the possible existence of associations between an outgroup’s visible characteristics (e.g., ethnicity and gender) and threat perceptions, together with dependencies on varying social contexts and respondents’ properties. The analyses are based on data from a student online survey. Multi-level models that account for the hierarchical data structure confirm the hypotheses derived by theory, albeit further variables which influence the survey response and introduce measurement error need to be considered in future research.

This paper describes how to use multi-level models for longitudinal studies with panel data. A special focus is set on the specification and estimation of hybrid models combining some advantages of models for fixed-effect and random-effects panel regression. The paper addresses some basic principles of multi-level- and panel analysis but it also discusses some very special problems of longitudinal analysis like centering of variables. It presents an example of a stepwise, applied panel regression analysis within a multi-level-model approach using the data of the SOEP (the national German socio-economic panel study) and the SPSS-MIXED-software. After reading this paper every social research scholar should be able to perform his/her first longitudinal analysis by multi-level modeling.