Association between regional income inequality and infant mortality rates in Türkiye: an ecological panel analysis of NUTS-2 regions, 2014–2024

Abstract

Background

Income inequality may affect infant survival through socioeconomic and health-care pathways, but regional evidence from Türkiye remains limited. This study examined the association between regional income inequality and infant mortality rates across the 26 NUTS-2 regions of Türkiye during 2014–2024.

Methods

This longitudinal ecological panel study used publicly available data from the Turkish Statistical Institute. Income inequality was measured using the Gini coefficient and P80/P20 ratio. Linear mixed-effects models included a regional random intercept and a first-order autoregressive covariance structure for repeated annual observations. Year was modelled categorically in the primary analyses, and between-region and within-region effects were examined separately. Live-birth-weighted sensitivity analyses were used to assess whether differences in the precision of regional IMR estimates affected the findings.

Results

The study included 286 region-year observations; income inequality models included 285. From 2014 to 2024, the regional mean Gini coefficient increased from 0.364 to 0.390 and the mean P80/P20 ratio from 6.42 to 7.22, while mean IMR declined from 10.81 to 8.86 per 1,000 live births. After adjustment for categorical year effects, no statistically significant association was detected for either the Gini coefficient (B = 0.041; 95% CI: −0.025 to 0.106; p = 0.221) or the P80/P20 ratio (B = 0.054; 95% CI: −0.145 to 0.252; p = 0.594). No statistically significant between-region or within-region associations were detected. Year-adjusted regional means ranged from 6.82 to 15.03, and 83.1% of the remaining variance was between regions. Live-birth-weighted sensitivity analyses likewise detected no statistically significant association for the Gini coefficient (B = 0.013; 95% CI: −0.041 to 0.067; p = 0.640) or the P80/P20 ratio (B = − 0.021; 95% CI: −0.189 to 0.146; p = 0.803).

Conclusions

No statistically significant association was detected between regional income inequality and IMR after accounting for year-specific effects and repeated observations. However, the confidence interval for the primary Gini estimate did not exclude small-to-moderate effects, so the findings should not be interpreted as evidence of no association. Persistent regional heterogeneity suggests that other demographic, social, and health-system factors may contribute to regional differences.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12889-026-29172-5.

Background

Income inequality is a major structural determinant of population health. Unequal income distribution affects not only purchasing power but also access to education, adequate nutrition, secure housing, safe environments, and social protection. These disadvantages may increase exposure to health risks and limit the use of health services. Income inequality is therefore regarded as an important social determinant associated with mortality, morbidity, and access to care [1, 2]. Children are particularly vulnerable. Poverty during childhood can impair growth and development, with effects that may persist into adulthood. Social protection programmes can reduce household vulnerability and support children’s health, nutrition, and education [3, 4].

The infant mortality rate (IMR) is the number of deaths before 1 year of age per 1,000 live births. It reflects more than the quality of clinical care. It also provides information on living conditions, maternal health, educational attainment, and the accessibility of the health system. IMR is therefore a key indicator of population health and health-system performance. Child mortality has declined substantially worldwide, but progress has slowed in recent years. Survival still varies markedly by household wealth, maternal education, and place of residence [5]. Evidence from Türkiye likewise links child mortality with education, economic vulnerability, and regional conditions [6].

Income inequality may affect infant survival through several pathways. Low income can restrict access to adequate nutrition and safe housing and increase health risks before and during pregnancy. Transport costs, user charges, and concerns about income loss may delay care seeking. Maternal education, fertility patterns, and birth spacing may also mediate the association. Conversely, vaccination, antenatal care, facility-based delivery, and postnatal follow-up reduce preventable deaths. Coverage, however, is not equally distributed across social groups. UNICEF reports that unvaccinated children are concentrated in poor, remote, and marginalised communities [7]. A study of 88 countries also found lower coverage of antenatal, delivery, and postnatal care in rural areas [8].

Universal health coverage (UHC) is a major policy instrument for reducing the translation of socioeconomic inequalities into adverse health outcomes. Insurance coverage alone is insufficient. Services must also be accessible, of adequate quality, and affordable. Recent global monitoring shows that progress in service coverage has slowed, while financial hardship and unmet health-care needs persist [9]. In an analysis of 4.1 million births from 60 low- and middle-income countries, improvements in UHC were associated with a lower risk of infant death, although the benefits were not equally distributed across income groups [10]. These findings underline the need to assess service coverage and financial protection together.

Primary health care (PHC) is central to this framework. A multicountry study in Latin America found that higher PHC coverage was associated with lower postneonatal and under-five mortality. Expanding PHC coverage was also projected to mitigate the adverse effects of economic crises on child health [11]. Continuity and broad coverage of PHC are therefore important for protecting child health.

Türkiye restructured the financing, organisation, and delivery of health services under the Health Transformation Program launched in 2003. Health insurance schemes were integrated under the General Health Insurance system [12], and family medicine was expanded nationwide by the end of 2010 [13]. Maternal and child health services subsequently reached high overall coverage. According to the Health Statistics Yearbook 2024, 99.4% of live births occurred in hospitals, while coverage for routine childhood vaccines ranged from 94% to 96% [14]. This institutional context may reduce the extent to which income differences translate into barriers to essential maternal and child health services.

Recent national data also indicate that infant mortality in Türkiye is concentrated in the neonatal period. In 2024, neonatal mortality rates were 4.9 per 1,000 live births among girls and 6.2 among boys, compared with postneonatal rates of 3.2 and 3.4, respectively [14]. The same yearbook reports neonatal disorders and congenital birth anomalies among important causes of premature mortality in its 2023 burden-of-disease estimates [14]. This profile provides relevant context for early-life mortality, particularly the contribution of conditions arising around birth and congenital disorders.

High population coverage, however, does not ensure equity in effective access. Studies from Türkiye show that cost-related unmet need is concentrated among lower-income groups [15, 16]. Lower educational attainment, lack of social security, chronic illness, and poor self-rated health have also been associated with unmet need [17, 18]. Geographic proximity alone may be insufficient. A capacity-adjusted analysis in western Türkiye found that effective PHC coverage declined and rural gaps became more pronounced when facility capacity was considered [19]. Household budget data also suggest that health insurance reduces the risk of catastrophic health expenditure, but financial protection remains incomplete for low-income and vulnerable households [20]. Inequalities in health-care utilisation across income groups also persist [21].

Evidence on the relationship between income inequality and IMR in Türkiye remains limited and inconsistent. Some regional and provincial studies have reported positive associations, whereas several panel analyses found no independent effect of the Gini coefficient [22–25]. Previous studies used different periods, units of analysis, and econometric methods, and many covered short intervals or a single year. Longer regional panels that account for repeated observations are therefore needed.

The broad coverage of maternal and child health services in Türkiye may partly buffer the effect of income inequality on infant mortality. However, persistent unmet need, financial barriers, and regional capacity differences suggest that this protection is not complete [14–21]. Rising income inequality may place additional pressure on this buffering capacity. This study therefore examined the association of regional income inequality, measured by the Gini coefficient and P80/P20 ratio based on household disposable income, with IMR across the 26 Nomenclature of Territorial Units for Statistics level 2 (NUTS-2) regions of Türkiye from 2014 to 2024. It also assessed long-term between-region differences and annual within-region changes separately.

Methods

Study design, setting, and data source

This secondary-data study used a longitudinal ecological panel design to examine the association between regional income inequality and IMR in Türkiye. The reporting of this study followed the relevant recommendations of the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement [26]. The unit of analysis was each region-year observation at the NUTS-2 level. Data from all 26 NUTS-2 regions for 2014–2024 yielded 286 region-year observations. Because all NUTS-2 regions and all available annual observations within the study period were included, no sampling procedure or formal sample size calculation was performed.

Data were obtained from the Turkish Statistical Institute Central Dissemination System (TURKSTAT-MEDAS) [27]. The study used NUTS-2-level data on IMR, the Gini coefficient based on household disposable income, and the P80/P20 income ratio. Regional live-birth counts were also obtained from TURKSTAT-MEDAS and were used in the sensitivity analyses.

The Gini coefficient and P80/P20 ratio are regional income-distribution indicators derived from the TURKSTAT Income and Living Conditions Survey. Higher values of both indicators represent greater income inequality. The P80/P20 ratio is the ratio of the total income received by the highest-income 20% of the population to that received by the lowest-income 20%. The income reference period corresponds to the calendar year preceding the survey year [28].

IMR was defined as the number of deaths before 1 year of age divided by the number of live births in the same year and was expressed per 1,000 live births [29].

The 2023 Gini coefficient and P80/P20 ratio for the TR63 region were not published because fieldwork could not be conducted following the earthquake [27]. No imputation or interpolation was performed. Analyses including income inequality indicators were therefore conducted on a complete-case basis and included 285 region-year observations, whereas analyses involving IMR alone included all 286 observations.

Study variables

The dependent variable was IMR. The main independent variables were the Gini coefficient and P80/P20 ratio. Regression coefficients were interpreted per 0.01-unit increase in the Gini coefficient and per 1-unit increase in the P80/P20 ratio. Year was included as the temporal adjustment variable for 2014–2024, and the NUTS-2 region was defined as the panel unit. The number of live births in each region-year was used as the weighting variable in sensitivity analyses.

Statistical analysis

Statistical analyses were performed using IBM SPSS Statistics, version 31. The region-level cluster bootstrap used to estimate the VPC confidence interval was implemented separately in Python 3.13. The Gini coefficient, P80/P20 ratio, and IMR were summarised by year using means and standard deviations. Temporal patterns were displayed using line graphs based on annual national values for Türkiye for descriptive comparison.

Unadjusted associations between the income inequality indicators and IMR were examined using Spearman rank-order correlations. Because these analyses did not account for repeated observations within regions or temporal effects, they were considered exploratory.

Regional differences in IMR were evaluated using a linear mixed-effects model in which region and year were entered as categorical fixed effects. Year-adjusted estimated marginal means (EMMs) were then calculated. Regional heterogeneity was further assessed in a complementary model with region specified as a random intercept and year as a categorical fixed effect. In both models, repeated annual observations within each region were modelled using a first-order autoregressive [AR(1)] covariance structure. The variance partition coefficient (VPC) was calculated, and the contribution of the regional random intercept was evaluated using a boundary-adjusted likelihood-ratio test. A bias-corrected and accelerated (BCa) 95% confidence interval for the VPC was estimated from 5,000 cluster-bootstrap samples generated by resampling NUTS-2 regions with replacement while retaining each sampled region’s complete annual series.

In the main models examining the association between income inequality and IMR, the NUTS-2 region was included as a random intercept, and successive annual observations within each region were modelled using an AR(1) covariance structure. Degrees of freedom were estimated using the Satterthwaite approximation. Because the Gini coefficient and P80/P20 ratio were highly correlated, they were not entered simultaneously; separate models were fitted for each indicator. The primary models assigned equal weight to each region-year observation. Equal weighting was retained in the primary analysis because the region-year, rather than the individual birth, was the prespecified ecological unit of analysis; the influence of differential precision related to live-birth counts was evaluated separately in the sensitivity analyses. To assess whether differences in the precision of IMR estimates related to regional live-birth counts affected the findings, the categorical-year Gini and P80/P20 models were refitted in sensitivity analyses using the number of live births as a regression weight applied to the residual covariance structure.

Year was initially modelled as a continuous variable. Because temporal changes did not follow a linear pattern, year was entered as a categorical fixed effect in the primary models. To distinguish between-region and within-region associations, the mean income inequality level across the study period was calculated for each region, together with each annual observation’s deviation from its region-specific mean. This decomposition was used to separate long-term differences between regions from annual changes occurring within the same region. Because the continuous- and categorical-year models differed in their fixed-effect structures, these alternative time specifications were additionally refitted using maximum likelihood for model comparison. Akaike information criterion (AIC), Bayesian information criterion (BIC), and likelihood-ratio tests were examined, while restricted maximum likelihood was retained for the primary coefficient estimates.

Model assumptions, including residual normality, homoscedasticity, and potential outlying observations, were assessed using Q–Q plots of residuals, box plots, and scatterplots of residuals against predicted values. Model estimates were reported as unstandardised regression coefficients with 95% confidence intervals. All tests were two-sided, and p < 0.05 was considered statistically significant. No convergence warnings were observed in the fitted models.

Results

A total of 286 region-year observations from the 26 NUTS-2 regions were evaluated for the 2014–2024 period. Because the 2023 Gini coefficient and P80/P20 ratio for TR63 were not published following the earthquake, the 2023 descriptive statistics for these two variables were calculated across 25 regions.

The mean Gini coefficient across NUTS-2 regions increased from 0.364 ± 0.025 in 2014 to 0.390 ± 0.031 in 2024. The highest mean Gini coefficient was observed in 2023 (0.394 ± 0.032). The mean P80/P20 ratio also increased, from 6.42 ± 0.81 in 2014 to 7.22 ± 0.98 in 2024, with the highest value recorded in 2023 (7.30 ± 1.11). Mean IMR declined from 10.81 ± 2.88 per 1,000 live births in 2014 to 8.50 ± 2.51 in 2020. It subsequently rose to 9.75 ± 3.53 in 2023 before decreasing to 8.86 ± 2.48 in 2024 (Table 1).

Table 1.

Annual income inequality indicators and infant mortality across NUTS-2 regions, 2014–2024

Year	Gini coefficient, mean ± SD	P80/P20 ratio, mean ± SD	Infant mortality rate, mean ± SD
2014	0.364 ± 0.025	6.42 ± 0.81	10.81 ± 2.88
2015	0.366 ± 0.023	6.55 ± 0.90	10.00 ± 2.76
2016	0.370 ± 0.022	6.58 ± 0.75	9.67 ± 2.58
2017	0.363 ± 0.026	6.32 ± 0.80	9.23 ± 2.24
2018	0.371 ± 0.025	6.64 ± 0.77	9.27 ± 2.11
2019	0.357 ± 0.024	6.19 ± 0.73	9.02 ± 2.45
2020	0.366 ± 0.030	6.77 ± 0.99	8.50 ± 2.51
2021	0.362 ± 0.026	6.29 ± 0.80	9.12 ± 2.46
2022	0.378 ± 0.029	6.78 ± 0.93	9.03 ± 2.38
2023	0.394 ± 0.032	7.30 ± 1.11	9.75 ± 3.53
2024	0.390 ± 0.031	7.22 ± 0.98	8.86 ± 2.48

For the Gini coefficient and P80/P20 ratio, the 2023 estimates were based on 25 regions because data for TR63 were unavailable. All other annual estimates were based on 26 regions

SD Standard deviation

At the national level, the Gini coefficient increased from 0.379 in 2014 to 0.418 in 2024, while the P80/P20 ratio rose from 6.895 to 8.111. Although both income inequality indicators fluctuated over time, they showed an overall upward trend across the study period, with the highest values recorded in 2023. In contrast, IMR declined from 11.1 per 1,000 live births in 2014 to 8.7 in 2020, increased to 10.1 in 2023, and then decreased to 9.0 in 2024. Overall, while national income inequality indicators showed an upward trend, IMR showed a general decline with marked fluctuations across years (Fig. 1).

Fig. 1.

National trends in income inequality indicators and infant mortality rate in Türkiye, 2014–2024. Panel A presents the annual Gini coefficient, Panel B presents the annual P80/P20 income ratio, and Panel C presents the infant mortality rate per 1,000 live births. Higher Gini coefficients and P80/P20 ratios indicate greater income inequality. Note: The figure presents national aggregate values for Türkiye (TR) obtained from the Turkish Statistical Institute Central Dissemination System (TURKSTAT-MEDAS). These national values are shown for descriptive comparison and differ from the annual means across the 26 NUTS-2 regions reported in Table 1 and used in the regional panel analyses. For income inequality indicators, the income reference period corresponds to the calendar year preceding the survey year

After controlling for year as a categorical fixed effect, IMR differed significantly across NUTS-2 regions, F(25, 48.192) = 33.289, p < 0.001. The overall effect of year was also significant, F(10, 140.385) = 8.288, p < 0.001. Year-adjusted EMMs ranged from 6.82 to 15.03 deaths per 1,000 live births; region-specific estimates are presented in Supplementary Table S1.

In the complementary model including region as a random intercept, the between-region random-intercept variance was 5.678 (95% CI: 3.206–10.055), and the within-region residual variance was 1.154. After accounting for the year effect, 83.1% of the remaining variance was located at the between-region level (BCa 95% CI: 69.0%–89.7%). Adding the regional random intercept significantly improved model fit (Δ − 2RLL = 58.071; boundary-adjusted likelihood-ratio test, p < 0.001). These findings indicated substantial regional heterogeneity in IMR (Table 2).

Table 2.

Regional heterogeneity and variance components of infant mortality rates across NUTS-2 regions, Türkiye, 2014–2024

Analysis component	Estimate or test statistic	Additional information	p value
Panel A. Overall regional and temporal differences
Overall NUTS-2 region effect	F = 33.289	Numerator df = 25; denominator df = 48.192	< 0.001
Overall year effect	F = 8.288	Numerator df = 10; denominator df = 140.385	< 0.001
Panel B. Variance components of the random-intercept model
Between-region random-intercept variance	5.678	SE = 1.656; 95% CI: 3.206–10.055	—
Within-region residual variance	1.154	SE = 0.121; 95% CI: 0.940–1.417	< 0.001
First-order autoregressive correlation, AR(1) ρ	0.282	SE = 0.069; 95% CI: 0.142–0.410	< 0.001
Variance partition coefficient	83.10%	5.678 / (5.678 + 1.154); BCa 95% CI: 69.0%–89.7%	—
Panel C. Contribution of the regional random intercept to model fit
Model including regional random intercept	−2RLL = 922.808	AIC = 928.808; BIC = 939.659	—
Model without regional random intercept	−2RLL = 980.879	AIC = 984.879; BIC = 992.113	—
Likelihood-ratio comparison	Δ − 2RLL = 58.071	Boundary-adjusted test	< 0.001

Infant mortality rates are expressed per 1,000 live births. Year-adjusted estimated marginal means were obtained from the model including NUTS-2 region and year as categorical fixed effects. The variance partition coefficient was calculated as the between-region variance divided by the sum of the between-region and residual variances. The likelihood-ratio test compared restricted maximum-likelihood models with identical fixed effects and was adjusted for the boundary distribution of the variance parameter. . The VPC 95% confidence interval was estimated using a BCa cluster bootstrap with 5,000 resamples at the NUTS-2 region level, retaining each sampled region’s complete annual series

SE Standard error, CI Confidence interval, df degrees of freedom, −2RLL − 2 restricted log likelihood, AIC Akaike information criterion, BIC Bayesian information criterion

A very strong positive Spearman rank-order correlation was observed between the Gini coefficient and the P80/P20 ratio (ρ = 0.962, p < 0.001). IMR showed a weak negative correlation with the Gini coefficient (ρ = −0.180, p = 0.002) and with the P80/P20 ratio (ρ = −0.188, p = 0.001). Because these analyses did not account for repeated observations within regions or the year effect, they were considered exploratory (Table 3).

Table 3.

Spearman correlations between income inequality indicators and infant mortality rate

Variable	1	2	3
1. Gini coefficient	1.000
2. P80/P20 ratio	0.962**	1.000
3. Infant mortality rate	−0.180*	−0.188*	1.000

All correlations involving an income inequality indicator were based on 285 complete region-year observations; infant mortality data were available for 286 region-year observations

*p < 0.01; **p < 0.001

When year was treated as a continuous variable, each 0.01-unit increase in the Gini coefficient was associated with an increase of 0.070 deaths per 1,000 live births in IMR (B = 0.070; 95% CI: 0.007 to 0.134; p = 0.030). A 1-unit increase in the P80/P20 ratio was not significantly associated with IMR (B = 0.080; 95% CI: −0.110 to 0.270; p = 0.410). IMR decreased annually by an average of 0.168 deaths per 1,000 live births in the Gini model and 0.158 in the P80/P20 model (p < 0.001 for both). When year was included as a categorical fixed effect to control for year-specific common changes, neither the Gini coefficient (B = 0.041; 95% CI: −0.025 to 0.106; p = 0.221) nor the P80/P20 ratio (B = 0.054; 95% CI: −0.145 to 0.252; p = 0.594) was significantly associated with IMR. The overall year effect remained significant in both categorical-year models (p < 0.001) (Table 4).

Table 4.

Linear mixed-effects model estimates for the associations of income inequality indicators with infant mortality rate

Panel A. Fixed-effect estimates
Model	Predictor	B	SE	95% CI	p
Gini model, year as continuous variable	Gini coefficient, per 0.01-unit increase	0.070	0.032	0.007 to 0.134	0.030
	Year	−0.168	0.026	−0.219 to − 0.117	< 0.001
P80/P20 model, year as continuous variable	P80/P20 ratio, per 1-unit increase	0.080	0.096	−0.110 to 0.270	0.410
	Year	−0.158	0.026	−0.210 to − 0.105	< 0.001
Gini model, year as categorical fixed effect	Gini coefficient, per 0.01-unit increase	0.041	0.033	−0.025 to 0.106	0.221
P80/P20 model, year as categorical fixed effect	P80/P20 ratio, per 1-unit increase	0.054	0.101	−0.145 to 0.252	0.594

Panel B. Model-level statistics
Model		Overall year effect		AR(1) correlation	Random intercept variance
Gini model, year as continuous variable		F = 42.837; p < 0.001		0.362	5.804
P80/P20 model, year as continuous variable		F = 36.346; p < 0.001		0.382	5.664
Gini model, year as categorical fixed effect		F = 9.792; p < 0.001		0.326	5.744
P80/P20 model, year as categorical fixed effect		F = 9.501; p < 0.001		0.332	5.678

All models included a random intercept for NUTS-2 region and a first-order autoregressive [AR(1)] covariance structure for repeated annual observations. All models were based on 285 region-year observations. The overall year effect represents the test of the linear year term in models treating year as a continuous variable and the omnibus test of year categories in models treating year as a categorical fixed effect

B unstandardized regression coefficient, SE Standard error, CI Confidence interval

Maximum-likelihood comparisons of the alternative time specifications showed that categorical-year models had lower AIC values than continuous-year models for both the Gini model (864.270 vs. 890.262) and the P80/P20 model (865.514 vs. 894.150), whereas BIC modestly favoured the more parsimonious continuous-year specifications (Gini: 912.177 vs. 919.057; P80/P20: 916.065 vs. 920.301). Likelihood-ratio tests favoured the categorical-year specification for both the Gini model (χ²(9) = 43.992, p < 0.001) and the P80/P20 model (χ²(9) = 46.636, p < 0.001). In live-birth-weighted sensitivity analyses, no statistically significant association was detected for either the Gini coefficient (B = 0.013; 95% CI: −0.041 to 0.067; p = 0.640) or the P80/P20 ratio (B = − 0.021; 95% CI: −0.189 to 0.146; p = 0.803). All primary, sensitivity, and model-comparison analyses converged without warnings.

The between-region and within-region components of income inequality were evaluated separately. For the between-region Gini coefficient, the coefficient for a 0.01-unit increase was negative but not statistically significant (B = − 0.350; 95% CI: −0.842 to 0.143; p = 0.156). The within-region Gini deviation was also not significantly associated with IMR (B = 0.048; 95% CI: −0.018 to 0.113; p = 0.154). Similarly, neither the between-region P80/P20 ratio (B = − 1.099; 95% CI: −2.605 to 0.408; p = 0.145) nor the within-region P80/P20 deviation (B = 0.075; 95% CI: −0.125 to 0.275; p = 0.461) was significantly associated with IMR. Thus, no statistically significant association was identified between IMR and either long-term differences in income inequality across regions or annual changes within the same region (Table 5).

Table 5.

Between-region and within-region associations of income inequality indicators with infant mortality rate

Panel A. Fixed-effect estimates
Model	Income inequality component	B	SE	95% CI	p
Gini model	Between-region Gini coefficient, per 0.01-unit increase	−0.350	0.239	−0.842 to 0.143	0.156
	Within-region Gini deviation, per 0.01-unit increase	0.048	0.033	−0.018 to 0.113	0.154
P80/P20 model	Between-region P80/P20 ratio, per 1-unit increase	−1.099	0.730	−2.605 to 0.408	0.145
	Within-region P80/P20 deviation, per 1-unit increase	0.075	0.101	−0.125 to 0.275	0.461

Panel B. Model-level statistics
Model	Overall year effect	AR(1) correlation	Random-intercept variance
Gini model	F = 9.853; p < 0.001	0.324	5.370
P80/P20 model	F = 9.571; p < 0.001	0.330	5.346

The between-region component represents each region’s mean income inequality level across 2014–2024. The within-region component represents the annual deviation from that region-specific mean. Year was included as a categorical fixed effect. Both models included a random intercept for NUTS-2 region and a first-order autoregressive [AR(1)] covariance structure for repeated annual observations. Both models were based on 285 region-year observations

B unstandardized regression coefficient, SE Standard error, CI Confidence interval

Discussion

This study examined the association between regional income inequality and IMR across the 26 NUTS-2 regions of Türkiye from 2014 to 2024. Regional mean Gini coefficients and P80/P20 ratios increased, whereas IMR generally declined. The trend was not linear, and a temporary increase occurred in 2023. After accounting for year-specific common effects and repeated observations within regions, no statistically significant association was detected for either income inequality indicator. In contrast, IMR showed substantial and persistent regional variation. The findings therefore indicate that this study did not detect a statistically significant association between the two income-inequality indicators and IMR, while other health, demographic, and service-delivery factors may contribute to regional differences.

The decline in IMR during a period of increasing income inequality indicates that the measures did not move in the same direction. Globally, child mortality has fallen markedly, although progress has slowed in recent years [5]. The similar direction of the regional means and national series suggests that the observed decline was not confined to a few regions. The pattern in Türkiye may be consistent with expanded vaccination, antenatal care, facility-based delivery, neonatal care, and infant follow-up [12, 14]. The increase in 2023 may represent a temporary deviation from the longer-term trend. It cannot be attributed to a single factor. The significant year effect suggests that shared influences, including economic conditions, changes in health-service use, post-pandemic effects, migration, disasters, and registration practices, may have affected regions to varying degrees.

A key finding was the persistence of regional heterogeneity after adjustment for year. Year-adjusted regional means ranged from 6.82 to 15.03 deaths per 1,000 live births. The improvement in model fit after inclusion of a regional random intercept, together with a VPC of 83.1%, suggests that much of the remaining variation may be related to relatively stable regional characteristics. Previous studies in Türkiye have linked child mortality with maternal education, fertility, women’s social position, economic vulnerability, and place of residence [6, 23]. The Mosley–Chen framework also proposes that socioeconomic conditions influence child survival through maternal factors, nutrition, environmental exposures, and health-service use [30]. Regional differences may therefore reflect prematurity, low birth weight, maternal age, birth spacing, rurality, migration, health workforce, referral capacity, and neonatal care resources. Evidence that geographic proximity to PHC facilities does not necessarily translate into effective access when capacity is considered is also consistent with this interpretation [19].

The national mortality profile also helps contextualise the persistent regional heterogeneity. Given the predominance of neonatal over postneonatal mortality in Türkiye [14], regional differences in prematurity and low birth weight, obstetric and neonatal care capacity, referral pathways, and the detection and management of congenital conditions may plausibly contribute to variation in IMR. These mechanisms were not measured directly and should be regarded as hypotheses for future regional analyses rather than explanations established by the present study.

The very high correlation between the Gini coefficient and P80/P20 ratio supports their evaluation in separate models. Exploratory analyses showed weak negative correlations between IMR and both indicators. However, these analyses did not account for year, repeated observations, or regional dependence. They were therefore interpreted as descriptive rather than inferential.

These two measures also have important interpretive limitations. Both reduce a complex regional income distribution to a single summary value and can mask substantial heterogeneity within NUTS-2 regions. The Gini coefficient captures inequality across the full distribution but does not identify where in the distribution the differences arise, whereas the P80/P20 ratio compares only the highest- and lowest-income quintiles and is insensitive to changes within the middle 60% [28]. Neither indicator directly captures wealth, absolute poverty, employment security, educational disadvantage, housing conditions, or individual access to health care. Consequently, non-significant coefficients for these broad aggregate indicators do not exclude associations between more specific dimensions of socioeconomic disadvantage and infant mortality.

When year was modelled as a continuous variable, a 0.01-unit increase in the Gini coefficient was associated with an increase of 0.070 deaths per 1,000 live births, whereas the P80/P20 ratio was not significantly associated with IMR. In the categorical-year models, neither indicator remained significant. This suggests that the Gini estimate was sensitive to the specification of time. Because IMR did not follow a linear trend, the categorical-year models provided a more cautious basis for interpretation. Maximum-likelihood comparisons further supported this choice: AIC values and likelihood-ratio tests favoured the categorical-year models, although BIC modestly favoured the more parsimonious continuous-year models. The live-birth-weighted sensitivity analyses also yielded no statistically significant association for either income-inequality indicator, indicating that the primary inferential conclusion was robust to differences in regional live-birth counts. Importantly, the 95% confidence interval for the primary categorical-year Gini estimate (− 0.025 to 0.106 per 0.01-unit increase) included the null value but did not exclude small-to-moderate positive effects. Accordingly, this result should be interpreted as a lack of statistically significant evidence of an association rather than evidence that no association exists. Our findings are consistent with Yetim et al., who reported no significant association for the Gini coefficient in NUTS-2 data from 2014 to 2018 [23]. Positive associations have been reported in a provincial panel analysis and a single-year spatial study [24, 25]. In an earlier regional study, income inequality was significant only in the pooled model [22]. These differences suggest that results may depend on the study period, unit of analysis, spatial scale, and treatment of time.

The between-region and within-region analyses supported the primary findings. Neither long-term regional mean inequality nor annual deviations from each region’s own mean were significantly associated with IMR. This does not imply that income inequality is unimportant for child health. Its effects may operate through unmeasured pathways such as maternal education, nutrition, housing, working conditions, and health-service use. Delayed effects or attenuation within regional averages are also possible.

The combination of rising income inequality, declining IMR, and non-significant estimates in the primary models is compatible with a possible buffering role of the health system in Türkiye. This mechanism was not measured directly, and the interpretation is contextual rather than causal. Integration of health insurance programmes under General Health Insurance, nationwide expansion of family medicine, and predominantly public financing of maternal and child health services may have reduced the dependence of essential service use on household income [12]. High reported coverage of hospital delivery and routine childhood vaccination also suggests broad reach [14]. This structure may partly limit income-related barriers to antenatal care, facility-based delivery, neonatal care, vaccination, and infant follow-up.

International evidence offers some support for this interpretation. UHC requires access to needed services without financial hardship, not only insurance coverage [31]. In more than 4.1 million births from 60 low- and middle-income countries, progress in UHC was associated with lower infant mortality, although benefits were not equally distributed across income groups [10]. Higher PHC coverage has also been associated with lower child mortality in Latin America [11]. In Brazil, the joint expansion of PHC, conditional cash transfers, and social pensions was associated with reductions in child mortality [32]. These findings suggest that public health and social protection mechanisms may partly moderate how economic inequality is translated into infant health outcomes.

High overall coverage, however, does not imply full equity. In Türkiye, unmet health-care needs are more common among people with lower income and education, those without social security, and those with greater health needs [15–18]. Income-related inequalities in health-care use also persist [21, 33, 34]. Across Europe, financial hardship from out-of-pocket payments is concentrated among poorer households and is linked to gaps in essential service coverage [35, 36]. Evidence from Türkiye likewise suggests that insurance reduces financial risk but does not provide complete protection for low-income and vulnerable groups [20]. Regional capacity gaps may also prevent nominal coverage from translating into equivalent access and quality [19]. The absence of a significant association in the primary models should therefore not be interpreted as evidence that economic and geographic inequalities in access have disappeared.

The findings point to two complementary policy priorities. Public financing, financial protection at the point of service, and broad maternal and child health coverage should be maintained. Resource allocation should also consider regional infant mortality burden, workforce, referral capacity, and effective access. The increase in the proportion of households experiencing catastrophic health expenditure from 0.31% in 2014 to 0.86% in 2024 warrants monitoring [14]. Microdata for 2014–2023 also indicate a higher relative burden of out-of-pocket spending among lower-income households [37]. Global and regional reports show that such payments can cause financial hardship and delayed or forgone care [9, 35, 38–40]. If financial protection weakens while income inequality rises, the association between inequality and infant health could become more pronounced. Regional IMR should therefore be monitored alongside out-of-pocket spending, unmet need, workforce capacity, service availability, and quality of care.

Limitations

This ecological study cannot support direct inference at the household or individual level. Potential confounders, including maternal education, fertility, prematurity, birth weight, rurality, migration, workforce, neonatal intensive care capacity, and service quality, were not included. The possible buffering role of the health system was not measured directly. Income inequality was assessed only with the Gini coefficient and P80/P20 ratio, which may not capture exposure among specific groups within regions. The income reference period preceded the survey year, and longer lag structures were not examined. Data from 26 regions over 11 years may have limited power, particularly for between-region effects. Missing 2023 income indicators for TR63 caused a small loss of data. The models accounted for temporal but not spatial dependence between regions. Secondary official data may also be affected by changes in registration or measurement practices. The primary mixed-effects models treated all region-year observations equally, although the precision of IMR estimates may vary substantially with the number of live births across regions. Consequently, smaller regions with less precise rate estimates contributed equally to the primary models. However, sensitivity analyses weighting the residual covariance by regional live-birth counts produced the same inferential conclusion for both income-inequality indicators.

In addition, the use of broad regional inequality indices may introduce exposure misclassification at the ecological level: regions with similar Gini or P80/P20 values can have different distributions of poverty or deprivation, and changes concentrated in specific income groups may be obscured by regional averages. Because these indicators are derived from a household survey, sampling error in regional estimates may introduce further measurement error and attenuate associations toward the null.

Conclusions

From 2014 to 2024, regional income inequality indicators in Türkiye generally increased, while IMR declined overall despite annual fluctuations. After accounting for year-specific effects and repeated observations, no statistically significant association was detected between IMR and either the Gini coefficient or the P80/P20 ratio. No statistically significant associations were detected for long-term between-region inequality or annual within-region changes. However, the confidence interval for the primary Gini estimate did not exclude small-to-moderate effects; therefore, these findings should not be interpreted as evidence that no association exists. In contrast, 83.1% of the remaining variation in IMR was located at the between-region level.

These findings should not be interpreted as showing that income inequality is irrelevant to child health. Public financing and broad maternal and child health coverage may have partly limited the translation of economic inequality into infant health outcomes, but this mechanism was not directly assessed. The substantial regional heterogeneity suggests that workforce capacity, referral systems, maternal education, fertility, rurality, and service quality may be important.

Broad maternal and child health coverage and financial protection should be maintained. Resources should also be directed according to need, particularly towards regions with higher IMR and lower service capacity. Regional infant mortality should be monitored together with out-of-pocket spending, unmet need, and service quality.

Abbreviations

AR(1)

First-order autoregressive

AIC

Akaike information criterion

BIC

Bayesian information criterion

CI

Confidence interval

EMM

Estimated marginal mean

IMR

Infant mortality rate

NUTS-2

Nomenclature of Territorial Units for Statistics level 2

PHC

Primary health care

SE

Standard error

STROBE

Strengthening the Reporting of Observational Studies in Epidemiology

TURKSTAT

Turkish Statistical Institute

UHC

Universal health coverage

VPC

Variance partition coefficient

Authors' contributions

O.A.: Conceptualization, methodology, formal analysis, data curation, visualization, writing—original draft, and writing—review and editing. F.K.: Conceptualization, methodology, visualization, writing—original draft, and writing—review and editing. Both authors read and approved the final manuscript.

Funding

This research received no external funding.

Data availability

The datasets analysed during the current study are publicly available through the Turkish Statistical Institute Central Dissemination System, as cited in the reference list. The analytical dataset supporting the findings of this study is available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Ethics committee approval was not required because the study used publicly available, aggregate, and non-identifiable secondary data and did not involve human participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.