Investigating the Macro Determinants of Self-Rated Health and Well-Being Using the European Social Survey: Methodological Innovations across Countries and Time

Abstract

At present, there is a debate over the relative importance and contribution of household income to well-being, and the link between economic growth, welfare, and well-being is not fully understood. We sought to examine how changes in contextual and individual income (spanning the Great Recession) are associated with changes in self-reported well-being in the European Social Survey (ESS) 2002–2011. A multivariate multilevel analysis was performed on 237,253 individuals nested within 128 country cohorts covering 30 countries. In this article, we focus specifically on the analysis and some of the methodological challenges and issues faced when making international comparisons across nations and time.

Keywords:

• inequality
• European Social Survey
• multilevel modeling
• subjective well-being
• welfare regimes

Acknowledgments

The authors thank Jon Kvist and Olli Kangas for their invitation to present at the 11th ESPAnet 2013 Conference, and the discussants. They are also grateful to IJS referees and editors for helpful comments on an earlier version of the manuscript.

Notes

The European Social Survey (ESS) is a biennial multicountry survey covering more than 30 countries. The first round was fielded in 2002/3, the fifth in 2010/11. The project is funded jointly by the European Commission, the European Science Foundation, and academic funding bodies in each participating country, and is designed and carried out to exceptionally high standards. The project is directed by a Core Scientific Team at the Centre for Comparative Social Surveys, City University, London. The questionnaire includes two main sections, each consisting of approximately 120 items; a “core” module that remains relatively constant from round to round, plus two or more “rotating” modules, repeated at intervals. The core module aims to monitor change and continuity in a wide range of social variables, including health and well-being; social and public trust; political interest and participation; sociopolitical orientations; governance and efficacy; moral, political, and social values; social exclusion, national, ethnic, and religious allegiances; human values; demographics and socio-economics. The ESS collects a wide range of methodological data, including tests of reliability, call records, data on interview settings, and event data. Upon registration, the ESS data are available free of charge and without restrictions, for not-for-profit and academic research purposes.

Our GDP per capita data cover the years 2002–10 and are available from the World Bank Web site (http://data.worldbank.org/indicator/NY.GDP.PCAP.CD [consulted June 2015]). GDP per capita is gross domestic product divided by midyear population. GDP is the sum of gross value added by all resident producers in the economy plus any product taxes and minus any subsidies not included in the value of the products. It is calculated without making deductions for depreciation of fabricated assets or for depletion and degradation of natural resources. Data are in current U.S. dollars.

The facility to model unbalanced data is based on the rather undemanding missing at random (MAR) assumption (Little and Rubin 2002) so that the missingness itself is uninformative and the nonresponse process depends only on observed variables and not on unobserved ones. In particular it is assumed that the probability of missing does not depend on responses we would have observed had they not been missing.

This is based on the MAR assumption. We will not get biased results providing that there is not something about the outcomes that has resulted in a particular survey wave in a country not being undertaken.

Estimation involves the inversion of matrixes of this order in an iterative fashion until convergence is achieved. Fairbrother (2014) used maximum likelihood with a Laplace approximation in his study of trust and postmaterialism in the World Values Survey (WVS). However, he did not use a Bernoulli model at the individual level but employed the binomial method of Subramanian, Duncan, and Jones (2001), thereby aggregating over individuals. This was appropriate for his study, but not for ours, because no level-1 predictor variables were specified and therefore the modeling is based on the much smaller data set of the proportions aggregated to waves and countries. Fairbrother was therefore able to use the more computationally demanding likelihood methods whereas we have used the computationally efficient quasi- likelihood ones in their least biased PQL-second-order form (Goldstein and Rasbash 1996).

It is perfectly possible for the higher-level variance to increase as lower level predictors are included (Jones 1992). This is particularly the case when dealing with Bernoulli models because the level-1 variance cannot decrease when predictor variables are included (Jones and Subramanian 2013). This does not happen here except for a very small increase with Model 4.

Additional information

Notes on contributors

Christopher Deeming

Christopher Deeming is ESRC Senior Research Fellow in the School of Geographical Sciences at the University of Bristol.

Kelvyn Jones

Kelvyn Jones is a professor of human quantitative geography and codirector of the Centre for Multilevel Modelling (CMM) at the University of Bristol. CMM at Bristol (http://www.bristol.ac.uk/cmm/) develops new statistical methodology and the MLwiN software used here to address unsolved issues in the quantitative modeling of social processes.