Explaining the Global Life-Expectancy Gap
Why do countries with similar economic resources have different health outcomes? An evidence-led extension of Global Health Atlas, combining data audit, statistical reasoning, SQL and interactive D3.
Global Health Atlas began as a five-chapter D3 application. This extension audits its assumptions, preserves its country exploration and rebuilds the analytical story around a specific question: what remains unexplained after accounting for economic resources?
The project combines interactive country exploration with a reproducible analytical workflow: data validation, economic-reference modelling, statistical analysis and predictive evaluation.
Which countries consistently sit above or below an economic reference for life expectancy? Investigate distributions, GDP relationships, persistent residuals, adjusted associations and later-year model errors. Each stage answers a different question.
The objective is transparent country comparison and analytical explanation. Residuals are not measures of healthcare efficiency or policy effectiveness.
The dataset contains 2,864 rows, 21 columns and 179 country labels across 2000–2015. All 2,864 expected country-year combinations are present.
All numerical rows and regional labels match the publisher download one-to-one. Twenty-two country labels differ. Publisher metadata identifies the dataset as CC0 and attributes fields to WHO, World Bank and Our World in Data. GDP is current USD; schooling is mean formal-education years for adults aged 25+.
No duplicate country-years or explicit missing cells were found, and the documented range/flag checks passed. The original JavaScript parser nevertheless converted missing numerical values to zero. It now preserves missingness and true zeros. This correction does not change the complete dataset.
The original mortality pie summed overlapping and non-commensurable indicators. It was replaced with distributions and country comparisons. Country-average correlations remain available, but are distinguished from cross-sectional and repeated country-year analysis.
Completeness is not raw-data quality: the publisher filled missing observations with nearby-year and regional averages. Without imputation flags, we cannot exclude future or cross-group information already entering the values.
Median country life expectancy rose from 69.7 to 73.0 years between 2000 and 2015.
The country-level 90th–10th percentile gap changed from 27.6 to 21.4 years.
Compare distributions, not a population-weighted world average.
Limit: Same 179 country labels; supplied processed values and national aggregates.
Reusable SQLite queries calculate paired endpoint improvements, persistent gaps, coverage and year summaries. The schema enforces unique country-years and plausible core ranges. SQL outputs are generated and checked against the Python analytical table.
In 2015, Pearson r was 0.63 for GDP and 0.82 for log GDP.
Spearman rank correlation was 0.85; each comparison uses 179 countries.
A log scale better reveals differences at lower GDP levels; neither coefficient establishes causation.
Limit: GDP is nominal current USD, not PPP; upstream imputation and ecological associations do not identify causal mechanisms.
| Specification | Adjusted R² |
|---|---|
| raw gdp | 0.359 |
| log gdp | 0.643 |
| quadratic log gdp | 0.656 |
The largest positive mean baseline residual belongs to Vietnam (+10.2 years); the largest negative to Swaziland (-20.3 years).
Mean over all 16 years relative to the same full-sample log-GDP + year-indicator baseline; median, persistence and volatility are also exported.
These are deviations from an economic reference, not rankings of healthcare efficiency.
Limit: In-sample, model-dependent ecological residuals; omitted conditions and source processing remain.
The schooling coefficient was 0.40 life-expectancy years per additional schooling year (country-clustered 95% CI 0.09 to 0.72).
Conditional on log GDP, BMI, three immunisation fields and year indicators, using 2,864 country-year rows.
An adjusted ecological association, not the effect of an education intervention.
Limit: Polio/DTP VIFs exceed 11; overlapping predictors make individual vaccine coefficients unstable. Shared shocks, source imputation and omitted conditions remain.| Field | Coefficient | 95% CI |
|---|---|---|
| log_gdp | 3.249 | 2.587 to 3.912 |
| Schooling | 0.404 | 0.089 to 0.720 |
| BMI | 0.284 | -0.139 to 0.707 |
| Polio | 0.128 | 0.021 to 0.236 |
| Diphtheria | 0.080 | -0.036 to 0.195 |
| Hepatitis_B | -0.047 | -0.074 to -0.019 |
Models include year indicators but no country fixed effects. They mix between- and within-country associations. Polio/DTP3 VIFs exceed 11; individual vaccine coefficients should not be interpreted as intervention effects. Influence and heteroscedasticity diagnostics are exported.
Gradient boosting achieved 1.99-year MAE on 2013–2015 (537 rows).
RMSE 2.71; R² 0.886. Family selected on 2011–2012; refitted on 2000–2012.
Contemporaneous covariates predict later-year outcomes for already observed countries.
Limit: This is not a forecast or causal model. Publisher three-year/regional imputation may include future information; upstream leakage cannot be excluded.Ten contemporaneous features are used. Country IDs, region, year, mortality outcomes, HIV and economy flags are excluded from the primary predictor set. Training preprocessing is fitted inside each pipeline, but cannot undo the publisher’s earlier imputation.
| Model | MAE / years | RMSE / years | R² |
|---|---|---|---|
| Mean baseline | 7.36 | 8.51 | -0.128 |
| Linear regression | 3.33 | 4.22 | 0.723 |
| Ridge regression | 3.32 | 4.22 | 0.723 |
| Random forest | 2.04 | 2.85 | 0.874 |
| Gradient boosting | 1.99 | 2.71 | 0.886 |
A separate geographic stress test holds out 36 countries and gives 2.31-year MAE (Gradient boosting). Selection uses only the 143 training countries. This is an across-country historical test, not a future-year test; regional source imputation may cross the split.
Largest country-level temporal test MAE: Cambodia, 11.65 years. The fixed-model country-block bootstrap 95% interval for overall MAE is 1.77–2.25 years; it excludes source and selection uncertainty.
The six-chapter D3 experience separates world distributions, economic resources, baseline gaps, contextual associations, prediction and findings. Select a year and country to inspect the GDP scatter, actual/reference readout, maps and country trajectory.
All 179 country labels resolve uniquely to local world-atlas geometry using explicit aliases and name normalisation. Missing/unmatched observations have a neutral state; exact values and a native country selector/table provide alternatives to hover. Charts redraw for viewport changes; no autoplay or long entrance animations are used.
The next analytical improvement is reconstructing source observations with imputation flags and training-only preprocessing. Then test baseline sensitivity and repeated temporal/geographic folds. Within-country modelling should follow a clear question and a defensible panel design.
Global Health Atlas combines interactive D3 country exploration with reproducible Python/SQL analysis, economic-reference modelling and predictive evaluation.
The GitHub repository and public deployment contain Global Health Atlas. The source download also includes the extended Python/SQL analysis, evaluation reports and interactive country comparisons.