Tuesday 16th July
Wednesday 17th July
Thursday 18th July
Friday 19th July
Download the conference book
Multilevel analysis in comparative research |
|
| Convenor | Professor Elmar Schlueter (University of Cologne) |
| Coordinator 1 | Professor Bart Meuleman (University of Leuven) |
We invite researchers to submit paper proposals for the session "Multilevel analysis in comparative research" at the 5th European Survey Research Association Conference, to be held from 15th to 19th of July 2013 in Ljubljana, Slovenia.
Point of departure for this session is the rapid growth in the analysis of survey-based multilevel data structures over the last decade. In large part, this growth has been driven by the availability of novel methodological tools, such as multilevel structural equation models and/or multilevel models for cross-classified data, to name just two examples. These innovations allow comparative researchers not only to explore previously inaccessible research questions, but also to revisit classis research topics with stronger tools. Against this background, the aim of this session is threefold:
- First, this session wants to stimulate the debate on conceptual as well as statistical issues that might arise when applying multilevel techniques in comparative research (e.g. the small-N problem, the black-box problem or Galton's problem).
- Second, this session aims to bring together studies that provide empirical comparisons of multilevel techniques with alternative approaches, e.g. hierarchical geostatistical models, fixed effect models or two-step hierarchical estimation.
- Third, we also invite papers that demonstrate how innovative applications of multilevel modeling techniques further social science understanding of substantial research problems.
The list below presents exemplary topics for contributions:
- Papers dealing with the consequences of and solutions to the small-N problem; techniques for combining mediating and moderating relationships; studies which add a spatial perspective to multilevel research
- Multilevel regression models that address substantive research problems considering hierarchical or non-hierarchical data structures
- Multilevel structural equation models
Paper proposals for this session should be sent to:
Elmar Schlueter (elmar.schlueter@wiso.uni-koeln.de)
Bart Meuleman (bart.meuleman@soc.kuleuven.be)
In the area of comparative value research some studies deal with societal factors like socioeconomic development influencing values (e.g. Inglehart). Other analyses focus on individual factors like social stratification (e.g. Kohn & Schooler). However, it lacks research taking into account societal- and individual-factors within a multilevel framework. This paper applies multilevel confirmatory factor analysis with covariates (Muthen & Muthen 2010) to explain differences in the priority of power- and achievement-values at both levels. Some advantages using this kind of analysis for comparative research are demonstrated, e.g. testing cross-country- and cross-level invariances.
One general problem of multilevel models with international survey data is that the number of cases on the upper level is rather small, often below 30. Consequently, outliers have a high impact on the results of multilevel regression models. Therefore, the analysis of country-comparative data calls for a thorough examination of the data prior to the application of complex regression models. Two-stage (slopes as outcomes) models are a valuable alternative to random effects models, particularly, if the distributional assumptions about the country-specific error terms are at stake.
We provide the programme-package multilevel tools (mlt) for Stata which helps to tackle these issues. The package contains four ados:
- "mltrsq" gives the Bosker/Snijders and Bryk/Raudenbush R-squared values..
- "mltcooksd" calculates the influence statistics Cook's D and DFBETAs for higher-level units. Works with linear, logistic and poisson regressions.
- "mltl2scatter" is an easy way to produce scatter plots at higher levels.
- "mlt2stage" is an easy way to estimate two stage (or slopes as outcomes) multilevel models. The coefficients can be plotted against third variables using "mltl2scatter".
We use examples from real research to stress the importance of considering influential cases and demonstrate how two-stage models can be used to check random effect models for their robustness.
Explanations for radical right-wing populist preferences received a lot of attention in previous research. Anti-immigrant sentiment is the most important predictor for such preferences. Scholars almost exclusively focused on the role of anti-immigrant sentiment as a person-level determinant. A contextual effect of anti-immigrant sentiment is mostly overlooked. In this research, we extend previous knowledge by disentangling the person-level and contextual effect of anti-immigrant sentiments on radical right-wing populist preferences. We argue that above and beyond individual-level anti-immigrant sentiments, the average anti-immigrant sentiment per region exerts a normative influence on personal radical right-wing populist preferences. We examine our theoretical model using voting preferences for the Swiss People's Party (SVP) - one of the most successful radical right-wing populist parties in Western Europe - as our test case. Using the Swiss Election Study 2011 (SELECTS) and based on multilevel structural equation modeling, we find clear evidence for a contextual effect of anti-immigrant sentiment on individual-level SVP preferences.
This study analyzes the direct and interactive effects of the legal regulation of homosexuality on sexual prejudice towards homosexuals from a cross-national, comparative perspective. Using recent survey data from 25 Western and Eastern European Countries and a novel measure of state policies towards homosexuality, the results from multilevel regression analyses reveal higher sexual prejudice in countries where homosexuals are granted relatively fewer rights. Additionally, and equally important, this study finds that the legal regulation of homosexuality systematically shapes the strength of the effect of classic individual-level predictors of sexual prejudice. These results point to the benefits to be achieved from linking contextual- and individual-level predictors of sexual prejudice.
108 words.
Belgium has been found repeatedly to rank among the European immigration countries with the highest ethnic educational inequalities (Marks, 2005; Jacobs, Rea and Teney, 2009). Recent literature has started to explore the impact of tracking, standardization and centralization (Alba and Sperling, 2011; Van de Werfhorst and Mijs, 2010; Montt, 2011), school autonomy (Verschelde et al., 2011) and other features of school systems that differ nationally or even regionally.
Dronkers, Van der Velden and Dunne (2011) have shown that the provisional conclusion that comprehensive educational systems lead to higher educational equality than stratified systems is premature, since it is often based on PISA data, which have problems of misrepresentation of ethnic minorities (Levels, Dronkers and Kraaykamp, 2008), are cross-sectional, and lack many relevant teacher and school characteristics.
The combination of pupil, peer network, classroom, teacher and school data in the panel study of CILS4EU in the Netherlands, Sweden, England and Germany, and LeuvenCILS in Flanders, offer possibilities to overcome these limitations and also to test cross-level moderation and mediation processes. For example, do ethnic composition effects affect ethnic inequalities equally in comprehensive and stratified systems? Do stratified systems lead to higher ethnic inequalities through low teacher expectations for ethnic minority students in lower tracks? The panel data also allow to model educational trajectories rather than achievement levels at a given point. Using multilevel models, we can test differences in causal relationships between countries with educational systems that range from comprehensive to strongly tracked.