Each panel shows the relationship between inequality and one outcome variable. We include a linear regression line and report the R-squared.
outcome_vars = {
'life_expectancy': ('Life expectancy (years)', False),
'infant_mortality': ('Infant mortality (per 1000)', True),
'homicide_rate': ('Homicides (per 100k)', True),
}
if not spirit_data.empty and 'gini_disp' in spirit_data.columns:
available = {k: v for k, v in outcome_vars.items() if k in spirit_data.columns}
n_plots = len(available)
if n_plots > 0:
fig, axes = plt.subplots(1, n_plots, figsize=(6 * n_plots, 5))
if n_plots == 1:
axes = [axes]
for ax, (var, (label, higher_is_worse)) in zip(axes, available.items()):
plot_df = spirit_data.dropna(subset=['gini_disp', var])
if plot_df.empty:
continue
colors = [country_color(c) if c in COUNTRY_NAMES else '#999'
for c in plot_df['country_code']]
ax.scatter(plot_df['gini_disp'], plot_df[var],
c=colors, s=80, zorder=5, edgecolors='white', linewidth=0.5)
# Regression line
x = plot_df['gini_disp'].values
y = plot_df[var].values
mask = np.isfinite(x) & np.isfinite(y)
if mask.sum() > 3:
slope, intercept, r, p, se = stats.linregress(x[mask], y[mask])
x_line = np.linspace(x[mask].min(), x[mask].max(), 100)
ax.plot(x_line, intercept + slope * x_line,
color='gray', linestyle='--', alpha=0.7)
ax.text(0.05, 0.95, f'R\u00b2 = {r**2:.2f}\np = {p:.3f}',
transform=ax.transAxes, fontsize=9, va='top',
bbox=dict(boxstyle='round', facecolor='white', alpha=0.8))
annotate_countries(ax, plot_df['gini_disp'].values,
plot_df[var].values,
plot_df['country_code'].values)
ax.set_xlabel('Gini (disposable income)')
ax.set_ylabel(label)
direction = 'worse \u2192' if higher_is_worse else '\u2190 better'
ax.set_title(f'Inequality vs. {label.split("(")[0].strip()}')
fig.suptitle('The Spirit Level: Inequality and Social Outcomes (~2019)',
fontsize=14, y=1.02)
plt.tight_layout()
save_figure(fig, '05_spirit_level_small_multiples')
plt.show()
else:
print('No outcome variables available in the data.')
else:
print('Spirit Level data or Gini column not available.')
The US is an extreme outlier on both axes: high inequality AND poor outcomes. A critical question is whether the correlation survives removing the US. If a single data point drives the relationship, causal claims are weak.
if not spirit_data.empty and 'gini_disp' in spirit_data.columns:
available = {k: v for k, v in outcome_vars.items() if k in spirit_data.columns}
n_plots = len(available)
if n_plots > 0:
fig, axes = plt.subplots(1, n_plots, figsize=(6 * n_plots, 5))
if n_plots == 1:
axes = [axes]
for ax, (var, (label, _)) in zip(axes, available.items()):
for include_us, style in [(True, {'alpha': 0.3, 'ls': ':'}),
(False, {'alpha': 0.9, 'ls': '--'})]:
if include_us:
plot_df = spirit_data.dropna(subset=['gini_disp', var])
lbl = 'With US'
else:
plot_df = spirit_data[spirit_data['country_code'] != 'US'].dropna(
subset=['gini_disp', var])
lbl = 'Without US'
x = plot_df['gini_disp'].values
y = plot_df[var].values
mask = np.isfinite(x) & np.isfinite(y)
if mask.sum() > 3:
slope, intercept, r, p, se = stats.linregress(x[mask], y[mask])
x_line = np.linspace(x[mask].min(), x[mask].max(), 100)
ax.plot(x_line, intercept + slope * x_line,
label=f'{lbl} (R\u00b2={r**2:.2f}, p={p:.3f})',
**style)
# Plot points (all countries)
full_df = spirit_data.dropna(subset=['gini_disp', var])
colors = ['#c75b5b' if c == 'US' else
(country_color(c) if c in COUNTRY_NAMES else '#999')
for c in full_df['country_code']]
sizes = [150 if c == 'US' else 80 for c in full_df['country_code']]
ax.scatter(full_df['gini_disp'], full_df[var],
c=colors, s=sizes, zorder=5, edgecolors='white', linewidth=0.5)
ax.set_xlabel('Gini (disposable income)')
ax.set_ylabel(label)
ax.set_title(f'{label.split("(")[0].strip()}')
ax.legend(fontsize=8)
fig.suptitle('Sensitivity Test: Does the Relationship Survive Without the US?',
fontsize=14, y=1.02)
plt.tight_layout()
save_figure(fig, '05_sensitivity_remove_us')
plt.show()
Deaton’s critique: the correlation between inequality and health may simply reflect that poorer countries are both more unequal and less healthy. If we control for GDP per capita, does the inequality effect survive?
if (not spirit_data.empty and 'gini_disp' in spirit_data.columns
and 'gdp_per_capita_ppp' in spirit_data.columns):
for var, (label, _) in outcome_vars.items():
if var not in spirit_data.columns:
continue
df = spirit_data.dropna(subset=['gini_disp', var, 'gdp_per_capita_ppp'])
if len(df) < 5:
continue
# Simple bivariate
r_bivar = stats.pearsonr(df['gini_disp'], df[var])
# Partial correlation: residualize both on GDP per capita
x = df[['gini_disp', 'gdp_per_capita_ppp']].values
y = df[var].values
# Residualize Gini on GDP
slope_g, intercept_g, _, _, _ = stats.linregress(
df['gdp_per_capita_ppp'], df['gini_disp'])
gini_resid = df['gini_disp'] - (intercept_g + slope_g * df['gdp_per_capita_ppp'])
# Residualize outcome on GDP
slope_o, intercept_o, _, _, _ = stats.linregress(
df['gdp_per_capita_ppp'], df[var])
outcome_resid = df[var] - (intercept_o + slope_o * df['gdp_per_capita_ppp'])
r_partial = stats.pearsonr(gini_resid, outcome_resid)
print(f"\n{label}:")
print(f" Bivariate: r = {r_bivar[0]:.3f}, p = {r_bivar[1]:.3f}")
print(f" Partial (controlling for GDP/capita): r = {r_partial[0]:.3f}, "
f"p = {r_partial[1]:.3f}")
else:
print('Data not available for partial correlation analysis.')
Life expectancy (years):
Bivariate: r = -0.002, p = 0.993
Partial (controlling for GDP/capita): r = 0.018, p = 0.935
Infant mortality (per 1000):
Bivariate: r = 0.543, p = 0.009
Partial (controlling for GDP/capita): r = 0.555, p = 0.007
Homicides (per 100k):
Bivariate: r = 0.569, p = 0.006
Partial (controlling for GDP/capita): r = 0.579, p = 0.005Life expectancy is the most commonly cited outcome in the Spirit Level literature. This plot adds GDP per capita as a bubble size to visualize the three-way relationship.
if (not spirit_data.empty and 'gini_disp' in spirit_data.columns
and 'life_expectancy' in spirit_data.columns
and 'gdp_per_capita_ppp' in spirit_data.columns):
df = spirit_data.dropna(subset=['gini_disp', 'life_expectancy', 'gdp_per_capita_ppp'])
fig, ax = plt.subplots(figsize=(11, 7))
# Bubble size proportional to GDP per capita
gdp_norm = df['gdp_per_capita_ppp'] / df['gdp_per_capita_ppp'].max()
sizes = 50 + 250 * gdp_norm
colors = [country_color(c) if c in COUNTRY_NAMES else '#999'
for c in df['country_code']]
ax.scatter(df['gini_disp'], df['life_expectancy'],
c=colors, s=sizes, zorder=5, edgecolors='white',
linewidth=1, alpha=0.8)
annotate_countries(ax, df['gini_disp'].values,
df['life_expectancy'].values,
df['country_code'].values)
# Regression line
x = df['gini_disp'].values
y = df['life_expectancy'].values
mask = np.isfinite(x) & np.isfinite(y)
if mask.sum() > 3:
slope, intercept, r, p, se = stats.linregress(x[mask], y[mask])
x_line = np.linspace(x[mask].min(), x[mask].max(), 100)
ax.plot(x_line, intercept + slope * x_line,
color='gray', linestyle='--', alpha=0.7)
ax.text(0.05, 0.05, f'R\u00b2 = {r**2:.2f}, p = {p:.3f}',
transform=ax.transAxes, fontsize=10,
bbox=dict(boxstyle='round', facecolor='white', alpha=0.8))
ax.set_xlabel('Gini coefficient (disposable income)')
ax.set_ylabel('Life expectancy at birth (years)')
ax.set_title('Inequality and Life Expectancy (~2019)\n'
'Bubble size = GDP per capita (PPP)', fontsize=13)
plt.tight_layout()
save_figure(fig, '05_gini_vs_life_expectancy')
plt.show()
else:
print('Data not available for life expectancy plot.')
Greener (2025) found that replicating the Spirit Level with updated data shows weakening correlations for several outcomes. We compare the inequality-health relationship at two time points.
years_to_compare = [2005, 2019]
results = []
for yr in years_to_compare:
data = health_social.get_spirit_level_data(year=yr)
if not data.empty and 'gini_disp' in data.columns:
for var, (label, _) in outcome_vars.items():
if var in data.columns:
df = data.dropna(subset=['gini_disp', var])
if len(df) >= 5:
r, p = stats.pearsonr(df['gini_disp'], df[var])
results.append({
'year': yr, 'outcome': label,
'r': r, 'r_squared': r**2, 'p_value': p,
'n_countries': len(df),
})
if results:
results_df = pd.DataFrame(results)
# Display comparison table
print('Gini-outcome correlations: 2005 vs 2019')
print('=' * 70)
for outcome in results_df['outcome'].unique():
subset = results_df[results_df['outcome'] == outcome]
print(f"\n{outcome}:")
for _, row in subset.iterrows():
sig = '***' if row['p_value'] < 0.01 else ('**' if row['p_value'] < 0.05
else ('*' if row['p_value'] < 0.1 else ''))
print(f" {row['year']}: r = {row['r']:.3f}, R\u00b2 = {row['r_squared']:.3f}, "
f"p = {row['p_value']:.3f}{sig} (n={row['n_countries']})")
# Bar chart comparing R-squared values
outcomes_with_both = [o for o in results_df['outcome'].unique()
if len(results_df[results_df['outcome'] == o]) == 2]
if outcomes_with_both:
fig, ax = plt.subplots(figsize=(10, 5))
x = np.arange(len(outcomes_with_both))
width = 0.35
for i, yr in enumerate(years_to_compare):
r2_vals = [results_df[(results_df['outcome'] == o) &
(results_df['year'] == yr)]['r_squared'].values[0]
for o in outcomes_with_both]
color = '#4878a8' if yr == 2005 else '#c75b5b'
ax.bar(x + i * width, r2_vals, width,
label=f'~{yr}', color=color, alpha=0.8)
short_labels = [o.split('(')[0].strip() for o in outcomes_with_both]
ax.set_xticks(x + width / 2)
ax.set_xticklabels(short_labels, rotation=30, ha='right')
ax.set_ylabel('R\u00b2 (Gini vs. outcome)')
ax.set_title('Has the Spirit Level Relationship Weakened?\n'
'Comparing ~2005 vs ~2019 correlations', fontsize=13)
ax.legend()
ax.set_ylim(0, 0.6)
plt.tight_layout()
save_figure(fig, '05_spirit_level_over_time')
plt.show()
else:
print('Not enough data to compare across time periods.')Gini-outcome correlations: 2005 vs 2019
======================================================================
Life expectancy (years):
2005: r = -0.021, R² = 0.000, p = 0.928 (n=22)
2019: r = -0.002, R² = 0.000, p = 0.993 (n=22)
Infant mortality (per 1000):
2005: r = 0.643, R² = 0.414, p = 0.001*** (n=22)
2019: r = 0.543, R² = 0.295, p = 0.009*** (n=22)
Homicides (per 100k):
2005: r = 0.486, R² = 0.237, p = 0.022** (n=22)
2019: r = 0.569, R² = 0.323, p = 0.006*** (n=22)
A common argument against redistribution is that it reduces economic growth. The cross-sectional evidence does not support a simple equality-prosperity trade-off among rich countries.
if (not spirit_data.empty and 'gini_disp' in spirit_data.columns
and 'gdp_per_capita_ppp' in spirit_data.columns):
df = spirit_data.dropna(subset=['gini_disp', 'gdp_per_capita_ppp'])
fig, ax = plt.subplots(figsize=(10, 7))
colors = [country_color(c) if c in COUNTRY_NAMES else '#999'
for c in df['country_code']]
ax.scatter(df['gini_disp'], df['gdp_per_capita_ppp'] / 1000,
c=colors, s=100, zorder=5, edgecolors='white', linewidth=1)
annotate_countries(ax, df['gini_disp'].values,
(df['gdp_per_capita_ppp'] / 1000).values,
df['country_code'].values)
x = df['gini_disp'].values
y = (df['gdp_per_capita_ppp'] / 1000).values
mask = np.isfinite(x) & np.isfinite(y)
if mask.sum() > 3:
slope, intercept, r, p, se = stats.linregress(x[mask], y[mask])
x_line = np.linspace(x[mask].min(), x[mask].max(), 100)
ax.plot(x_line, intercept + slope * x_line,
color='gray', linestyle='--', alpha=0.7)
ax.text(0.05, 0.95, f'R\u00b2 = {r**2:.2f}, p = {p:.3f}',
transform=ax.transAxes, fontsize=10, va='top',
bbox=dict(boxstyle='round', facecolor='white', alpha=0.8))
ax.set_xlabel('Gini coefficient (disposable income)')
ax.set_ylabel('GDP per capita, PPP ($1000s)')
ax.set_title('Inequality vs. Prosperity (~2019)\n'
'No clear trade-off among rich countries', fontsize=13)
plt.tight_layout()
save_figure(fig, '05_gini_vs_gdp')
plt.show()
else:
print('Data not available for GDP scatter.')
The Spirit Level correlations are real but fragile. More unequal rich countries do tend to have worse health and social outcomes, but the relationships are sensitive to country selection and time period.
The US is a massive outlier. Removing the US weakens many of the correlations considerably, raising questions about generalizability.
Controlling for GDP matters. Some correlations weaken when controlling for national income, consistent with Deaton’s alternative explanation.
The relationship may be weakening. Greener (2025) finds that updated data shows weaker correlations for several outcomes, possibly because health improvements in some unequal countries have outpaced inequality growth.
There is no equality-prosperity trade-off. Cross-sectional evidence does not support the claim that less inequality means less prosperity. Nordic countries achieve both.
Causation remains debated. The four proposed pathways (psychosocial stress, material deprivation, political economy, social capital) are all plausible but difficult to disentangle empirically.