[Paper Review] Predicting Long-Term Self-Rated Health in Small Areas Using Ordinal Regression and Microsimulation

Seán Caulfield Curley, Karl Mason+1 more|arXiv (Cornell University)|Jan 20, 2026

demographic modeling and climate adaptation0 citations

TL;DR

The paper combines an open-source microsimulation for Ireland with ordinal regression to forecast future self-rated health (SRH) at fine geographic scales, aligning microdata with census distributions and exploring future scenarios under aging and migration.

ABSTRACT

This paper presents an approach for predicting the self-rated health of individuals in a future population utilising the individuals' socio-economic characteristics. An open-source microsimulation is used to project Ireland's population into the future where each individual is defined by a number of demographic and socio-economic characteristics. The model is disaggregated spatially at the Electoral Division level, allowing for analysis of results at that, or any broader geographical scales. Ordinal regression is utilised to predict an individual's self-rated health based on their socio-economic characteristics and this method is shown to match well to Ireland's 2022 distribution of health statuses. Due to differences in the health status distributions of the health microdata and the national data, an alignment technique is proposed to bring predictions closer to real values. It is illustrated for one potential future population that the effects of an ageing population may outweigh other improvements in socio-economic outcomes to disimprove Ireland's mean self-rated health slightly. Health modelling at this kind of granular scale could offer local authorities a chance to predict and combat health issues which may arise in their local populations in the future.

Motivation & Objective

• Predict future distributions of self-rated health (SRH) for Ireland at fine geographic scales (Electoral Divisions).
• Integrate a dynamic open-source microsimulation (SEMIPro) with ordinal regression to predict SRH from socio-economic characteristics.
• Align microdata SRH distributions with Census distributions to improve prediction accuracy.
• Explore how demographic and socio-economic changes (e.g., aging, education, migration) affect SRH trajectories.
• Demonstrate usefulness for local health planning and scenario analysis.

Proposed method

• Use SEMIPro microsimulation to create synthetic individuals with attributes: age, sex, marital status, citizenship, moved to Ireland recently, education level, and primary economic status.
• Apply an ordinal regression model on SRH with predictors age group, sex, marital status, economic status, education, and region; use a cumulative model with latent Y* and thresholds to estimate probabilities for SRH categories.
• Align predicted SRH distributions per cohort with national Census distributions via element-wise ratios and renormalization.
• Predict future SRH distributions using per-cohort additive log-ratio (ALR) transformation and Gaussian Process (GP) regression to model log-ratios across years and cohorts; perform Monte Carlo sampling (1000 samples) for uncertainty.
• Validate predictions by comparing ED-level mean SRH in 2022 to observed 2022 SRH, achieving high fit (R^2 ~ 0.9, MSE ~ 0.0037).
• Use case studies and scenario analyses (e.g., M1 migration scenario and 2057 projection) to illustrate geographic disparities and policy relevance.

Figure 1: A graphical summary of how the approach operates. Note that synthetic individuals are created for this example and are not based on individuals from either dataset.

Experimental results

Research questions

• RQ1 Can ordinal regression accurately predict individual SRH from socio-economic attributes in a microsimulation context?
• RQ2 How well can microsimulation-generated populations reproduce observed SRH distributions at small-area geography (EDs)?
• RQ3 What is the impact of aging and education changes on country-wide SRH distributions by 2057 under defined migration scenarios?
• RQ4 How does aligning microdata SRH with census SRH distributions affect predictive accuracy?
• RQ5 What geographic patterns emerge in SRH trajectories, and how can they inform local health planning?

Key findings

• Predicted SRH distributions at the Electoral Division level align closely with observed 2022 distributions, with mean R^2 around 0.9 and mean MSE around 0.0037.
• The strongest negative predictor of SRH deterioration is “Unable to work due to sickness or disability,” while positive predictors include being a student, working, or Looking after home/family.
• Age is a strong negative predictor of SRH, and higher education generally correlates with better SRH across regions.
• Future projections (up to 2057 under the M1 migration scenario) show rising proportions of Very Good and Good SRH initially, with bidirectional uncertainty captured via 90% confidence intervals; patterns stabilize around mid-2030s.
• Actual_outcome_note_to_usersOnly_from_source: The paper reports that Ireland’s SRH distribution has been relatively stable across recent censuses, but projections indicate notable changes by 2057.

Figure 2: A flowchart summarising the workings of the SEMIPro Irish microsimulation model (Caulfield Curley et al., 2025a ) .