LVCN is an ESRC funded project based in the Methodology Institute and the Statistics Department at the LSE. Its aims are to develop and encourage the use of particular statistical tools for cross-national survey analysis that will lead to better utilization of the data, more valid conclusions, and more relevant input into social science and public policy making.
Project overview
In surveys many concepts are too complex to be captured or "measured" by a single question, so batteries of several questions are used to measure such underlying (latent) concepts. Statistical latent variable models are used to represent measurement of concepts by multiple questions and analyse the relations among the concepts. But a pressing challenge in cross-national survey research is the comparability or equivalence of such measurement across countries, i.e. the issue of whether a survey question "works the same way" in all of the countries? If it does not, any comparisons between countries may be distorted, since observed differences may be artefacts of measurement. While survey designers use a variety of techniques, such as advanced translation procedures, to check the meaning of items across countries, the issue of non-equivalence of measurement continues to be a concern. There is a need for complementary statistical methods for examining the equivalence of survey questions and of derived latent concepts.
We will develop tools of latent variable modelling for cross-national survey data. In particular, we will examine methods which correctly account for the fact that typical survey questions are categorical in nature, meaning that the respondents answer them by selecting one of a small set of fixed response options. Latent variable models for categorical data are known as latent class and latent trait models. While much is known about other methods which are strictly speaking only appropriate for the analysis of continuous-level survey data, little work has been done on analysis of international surveys using tools specifically designed for survey questions that yield categorical data.
By drawing upon the researchers' expertise in both statistics and substantive social science, the project will involve a constant interplay between statistical modelling and the analysis of real survey data to ensure the practical applicability of the work. The project will:
(a) provide advice and examples, which are accessible to all survey researchers, of how to use statistical latent trait and latent class models to analyse cross-national survey data, to examine equivalence of survey questions and to find answers to substantive research questions;
(b) examine the general properties of the models and methods of comparing them, using both mathematical and simulation techniques, and analyses of real survey data; and
(c) provide practical guidelines to survey users about what may be done in different situations of measurement equivalence, and how such choices affect conclusions about the substantively interesting research questions.
The methodological work of the project will be tested and illustrated in the context of cross-national comparisons in three substantive areas that the applicants specialise in:
(i) confidence in criminal justice systems (using data from the European Social Survey),
(ii) public attitudes towards science and technology (with Eurobarometer surveys), and
(iii) individual-level characteristics of civil society (using the European and World Values Surveys, and Afrobarometer surveys).
To enable a wide range of potential users to learn from the project, we will provide two-day training workshops for PhD students and survey researchers across the UK. We will also provide information and training materials on our project website, and disseminate our findings through high-level academic conferences and publications.