Health Inequalities across Europe: small-area insights from the Euroepan Social Survey

Sara Martino
sara.martino@ntnu.no

Dept. of Mathematical Science, NTNU

This is joint work with

T. Eikemo

I. Backhaus

H. Hoven

M. Beneito

A. Riebler

The European Social Survey

The European Social Survey (ESS)

The ESS is designed to measure the values, attitudes, and behavioral patterns of the populations across Europe.

  • A biennial, cross-national survey conducted in approximately 30 European countries since 2002.
  • In each participating country, a minimum of 1,500 respondents are surveyed.
  • Respondents are drawn from a probabilistic sample representing the countries’ population aged 15 and above.
  • The ESS aims at allowing for cross-country compatibility using harmonized questionnaire design.

Participating countries

Focus areas

How about the survey design: Example of Switzerland

Motivation

Education and Health

Education is one of the strongest social determinants of health

Why this matters:

  • Education influences health through income, employment, health literacy, and access to resources.

  • Understanding these pathways helps explain health inequalities and identify intervention targets.

  • Educational improvements may yield long-term health benefits beyond economic returns.

  • Evidence can guide policies aimed at improving population health and reducing disparities.

Why Use Self-Rated Health?

Self-rated health is a simple but powerful measure of overall health and well-being.

  • 😀 Captures multiple dimensions of health, including physical, mental, and functional well-being.

  • 😀 Strong predictor of future morbidity, healthcare use, and mortality.

  • 😀 Easy and inexpensive to collect in large-scale surveys.

  • 😀 Available across countries and over time, enabling comparative research.

  • 😒 Relies on individual perception and may be influenced by cultural context.

Self-rated health as a general health indicator

“How is your health in general?”

Goal of the work

  • Accumulated evidence has shown educational inequalities in self-rated health over time across Europe. \(\longrightarrow\) However, there is a lack of understanding on how trends evolve subnationally.

  • Monitoring health inequalities at fine spatial scales is crucial for policy planning, but data sparsity and complex survey designs pose challenges.

What we consider..

Variables that are related both to the sampling/design process and to survey response propensity are:

Data

Which data do we use?

  • We use data from 2010 to 2024 (7 ESS waves) across 195 NUTS-2 regions in up to 31 countries giving us about 40 000 respondents per wave.

Measures

  • Self-rated health: \(\Rightarrow\) Dichotomisation into 1: very good/good, 0: fair/bad/very bad

  • Gender: binary

  • Age: 3 groups: less than \(35\), \(35\)–\(55\), older than \(55\).

  • Education: low level, medium, high level.

  • Region: NUTS2/NUTS1 level information per respondent.

Statistical model

Why do we need a model and what should it do?

  • Descriptive analysis does not work for small area estimation due to the small sample size \(\Rightarrow\) We need a model!
  • We would like to borrow strength across time and space \(\Rightarrow\) Bayesian hierarchical model with random effects
  • In the end we would like to draw conclusion for the entire population not just our survey population.

Our model in 1 page

We use a Bayesian hierarchical model for individual-level data with three key properties:

  1. It is a logistic model due to the binary outcome variable SRH.
  2. Inclusion of random effects for space and time that allow for smoothing by borrowing strength of neighbouring time points or regions.
  3. Post-stratification to get representative estimates at population level.

Model details

We fitted separate model for males and females.

Let \(Y_i\) denote the observed value for individual \(i\):

\[\begin{align*} Y_i \mid \pi_i &\sim \text{Bernoulli}(\pi_i)\\ \text{logit}(\pi_i) &= \mu + u_{\ r_i} + \alpha_{\ t_i} + \phi_{\ c_i} +\text{fixed effects} + \text{interaction effects}\\ \end{align*} \]

  • \(\bf{u}\) follows a BYM2 model over the region \(r\).
  • \(\bf{\alpha}\) follows a RW2 model over the ESS wave \(t\).
  • \(\bf{\phi}\) is an iid effect for country \(c\).
  • Fixed effects: Age and education
  • Interaction effects to adjust for the survey design as recommend by Gelman et al.

Multilevel Regression and Post-Stratification (MRP)

We are interested in prevalence estimates at the population level.

MRP is a two stages:

  1. Use a model to estimate probabilities \(\pi_i\)

    • These probabilities reflect the survey population

Multilevel Regression and Post-Stratification (MRP)

We are interested in prevalence estimates at the population level.

MRP is a two stages:

  1. Use a model to estimate probabilities \(\pi_i\)

    • These probabilities reflect the survey population

  2. Weight model predictions for different subgroups by the actual frequency of these subgroups. This idea can be expressed as:

\[ \hat{\pi}^{MRP}_s = \frac{\sum_{i\in s} N_j \hat{\pi}_j}{\sum_{i\in s} N_j } \]

  • These probabilities reflect the actual population
  • We need to know the population in each of the cells se are interested in (GENDER\(\times\)AGE\(\times\)EDUCATION\(\times\)NUTS)

Population in Europe

We need population counts stratified by gender, age, education and NUTS.

Can be problematic as NUTS regions have changed over time

  • We use \(1\times 1\) Km population estimates from WorldPop stratified by gender and age groups.

  • Those estimates were aggregated to the regions of our study (NUTS2).

  • To get estimates stratified by education, we applied a multinomial model incorporating survey weights from the ESS leading to a pseudo-likelihood.

https://www.worldpop.org/

Uncertainty quantification

  • We use a Monte Carlo approach to compute uncertainties for posterior marginal estimates adjusting for how many people we have in the strata.

  • We also account for uncertainty in the population estimates that we need for MRP

Results

Posterior probability of reporting good health - by education

Posterior probability of reporting good health - by age

Mean sex difference in reporting good SRH

Discussion

Discussion I

This study intends to make an empirical and methodological contribution to understanding inequalities in self‑rated health trends in Europe by presenting systematic sub-national analyses over time disaggregated by two prominent determinants of health: gender and education.

  • modest improvements in self-rated health over time, with gains concentrated among women
  • substantial within-country heterogeneity emerges, underscoring the importance of small-area analyses.

Of note, we assessed how well we adjust for the survey design in our model by regressing the survey weights on the control variable and their interactions finding that we explain about \(70\%\) of the design.

Discussion II

More in general:

  • The ESS allows to investigate health inequity across Europe along different variables.

  • Multi-country comparisons are encouraged.

  • Challenges include that there are not so many respondents which complicates spatial analysis and the survey design has to be adjusted for.

Thank you!

Posterior probability of reporting good health

Posterior probability of reporting good health

Posterior probability of reporting good health