Tracing echoes from the past: a district-level analysis of thyroid cancer incidence and radiation exposure in post-Chornobyl Ukraine

Abstract

Introduction

Increases in thyroid cancer have been reported as one of the major public health consequences of the Chornobyl accident. Using an excess relative risk (ERR) model and detailed district-level data across Ukraine, we aimed to estimate the association between iodine-131 exposure and thyroid cancer risk while accounting for spatial confounding.

Methods

This ecological study used data on all diagnosed thyroid cancer cases in 2001 among individuals aged 15 years or older and settlement-level iodine-131 thyroid doses from 1986, covering the entire territory of Ukraine and encompassing 627 spatial units (districts and cities). District doses were estimated by sex and age using areal-weighted interpolation. District-level standardised incidence ratios (SIRs) were calculated and smoothed with negative binomial regression and spatial P-splines. The ERR was assessed using an intrinsic conditional autoregressive negative binomial (ICAR-NB) model adjusting for sex, age at exposure and district-specific effects.

Results

Excess thyroid cancer incidence was observed not only in northern districts near the Chornobyl plant but also in several central districts. The ICAR-NB model produced lower ERR estimates than the corresponding model without district-level adjustment, suggesting that risks may be overestimated when spatial dependencies are ignored (ages at exposure 0–4: 24.85 (95% CI 7.78 to 41.78) vs 48.09 (95% CI 17.05 to 79.11)). ERRs were significant for both sexes although sex differences were not statistically significant. Risk decreased strongly with age at exposure.

Conclusions

This study underscores the value of small-area data analysis and spatial modelling for assessing radiation-related health risks and offers methodological guidance for future studies following nuclear accidents.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Exposure to radiation after nuclear accidents is associated with an increased risk of thyroid cancer, particularly in children who ingest high doses of radioactive iodine.

• Prior studies have explored the association between radiation and thyroid cancer, primarily focusing on heavily contaminated areas in northern Ukraine and other parts of the former USSR, or using aggregated data at the regional (oblast) level.

WHAT THIS STUDY ADDS

• This is the first district-level analysis of thyroid cancer incidence covering the whole of Ukraine using spatial smoothing methods. It revealed elevated incidence not only in northern districts near Chornobyl but also in several districts of Central Ukraine.

• Although the study confirms a strong ecological association between iodine-131 intake in 1986 and thyroid cancer incidence in 2001, it also warns that ignoring spatial dependence may lead to overestimation of radiation-related risks because of unmeasured spatially varying factors.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The results provide valuable contextual evidence and methodological guidance for future research on the public health consequences of nuclear accidents using aggregated spatial data.

• Our findings underscore the importance of continued monitoring of long-term health and environmental risks, especially amid rising nuclear safety concerns related to the ongoing war in Ukraine, other military conflicts and climate disasters.

Introduction

The Chornobyl (Chernobyl) accident in 1986 in Ukraine remains the most serious nuclear accident in history, resulting in the largest release of radioactive material into the environment and unprecedented consequences for populations living in the affected regions.^1–7 Decades later, the Chornobyl plant returned to public attention when it was temporarily occupied by Russian troops in early 2022. The ongoing Russian occupation of the Zaporizhzhia nuclear power plant, located near the frontline, has further heightened concerns about the risk of nuclear disaster. These distressing events highlight the vulnerability of nuclear power sites during armed conflicts, and emphasise the need for an immediate international response to any armed attacks on nuclear facilities.^8–10

Present-day concerns about the long-term effects of radiation exposure are rooted in the lessons learnt from the 1986 Chornobyl accident, which abruptly placed 2293 Ukrainian settlements with a total population of more than 2.4 million within contaminated territory, subjecting them to varying levels of radiation exposure.⁶ Compared with other nuclear accidents, such as the Fukushima Daiichi disaster in 2011, thyroid radiation doses from Chornobyl were substantially higher, particularly in children who ingested radioactive iodine (iodine-131) through contaminated milk and food.^7,11 Over the past four decades, extensive research has examined health repercussions linked to this accident. Among these, a sharp rise in thyroid cancer incidence has been the most notable health consequence at the population level.^4,5,12 The increased risk of thyroid cancer was also observed in Belarus and some regions of Russia.^13–15 While significant progress has been made in understanding the health impacts of the Chornobyl accident, most epidemiological studies have focused on heavily contaminated areas in northern Ukraine, Belarus and western Russia or have relied on aggregated data at the regional (oblast) level.^4,16–20 Many of these studies are based on cohorts with individual-level dose assessments and standardised thyroid screenings conducted in the most exposed areas, providing critical insights into high-risk groups. At the same time, much less is known about the broader spatial patterns of thyroid cancer incidence across Ukraine, especially in areas with moderate or lower exposure levels. Ecological studies at the national level have been rather scarce, largely due to the lack of detailed small-area data and concerns about differential screening bias, where heavily exposed areas may have undergone more intensive monitoring and thus report higher cancer rates.^21–23

To address these limitations, we conducted the first nationwide district-level analysis of thyroid cancer incidence across the whole territory of Ukraine using cancer registry data linked with district-level estimates of thyroid radiation doses from iodine-131 (¹³¹I; half-life: 8 days) exposure. Our study addressed two questions: (1) What were the spatial patterns of thyroid cancer incidence across Ukraine 15 years after the Chornobyl accident, and how did they vary between districts? (2) To what extent were district-level thyroid radiation doses from iodine-131 exposure associated with thyroid cancer risk after accounting for spatial dependence?

Unlike previous ecological studies, we combined spatial smoothing of standardised incidence ratios (SIRs) with a spatially adjusted excess relative risk (ERR) model that accounts for correlations between neighbouring districts. This approach enabled us to identify clusters of high and low thyroid cancer incidence and to estimate the association between thyroid radiation dose and thyroid cancer risk while accounting for residual spatial variation that may reflect differences in healthcare access, diagnostic intensity, post-accident population mobility, iodine nutrition and other contextual factors.

By combining small-area analysis with spatial smoothing and spatially adjusted ERR modelling, our study demonstrates the potential of ecological data to complement findings from individual-level cohort studies. It provides additional insights into the long-term population-level consequences of the Chornobyl accident by characterising the spatial distribution of thyroid cancer across areas with different levels of radiation exposure.

Methods

Data

This nationwide ecological study used aggregated data and relied on the following datasets. Thyroid cancer data were obtained from the National Cancer Registry of Ukraine (NCRU), which has operated nationally since 2000.²⁴ We included all new incident thyroid cancer cases diagnosed in 2001 among individuals aged 15 years or older (born in 1986 or earlier), totalling 2269 totalling 2269 cases across 675 districts and cities; 18% were male.

The population at risk was derived from the 2001 Ukraine census population, stratified by age and sex, covering 663 districts and cities across 27 regions, with a cumulative population at risk of 40.5 million for individuals aged 15 years and older (online supplemental table S2). The analysis was restricted to 2001 because nationally consistent district-level population denominators by age and sex were available only from the 2001 Ukraine population census, the only census conducted since Ukraine’s independence. To harmonise differences in spatial units between cancer cases and population data, certain cities were merged with their districts. This resulted in a total of 627 spatial units (with two missing cancer data). Cases were stratified by 5-year age groups, sex and district/city. We also used revised thyroid dose estimates from ¹³¹I intake in 1986, produced by the National Research Center for Radiation Medicine.²⁵ Doses were estimated by sex and age group (0, 1–4, 5–9, 10–14, 15–19, 20+ years) for 22 632 settlements in Ukraine in 27 regions. The data on thyroid cancer cases and thyroid doses received are part of the MAPA: Digital Atlas of Ukraine project being developed by the Harvard Ukrainian Research Institute.²⁶ Details can be found in online supplemental material.

Patient and public involvement

Patients and members of the public were not involved in the design, conduct, reporting or dissemination plans of this research.

Statistical analysis

Our analysis was conducted at the district level using data from 627 spatial units (districts and cities) on population and thyroid cancer incidence in 2001. First, we harmonised spatial units across datasets, assigned settlement-level thyroid doses from 1986 (22 632 settlements) to these units and applied areal-weighted interpolation to obtain district-level average doses by age and sex.²⁷ To estimate relevant exposure for the 2001 study population (aged 15 years or older), we calculated weighted averages of district- and sex-specific doses using census population data.

Second, we calculated district-level SIRs of thyroid cancer in 2001 for individuals aged 15 and older. Expected cases were derived by applying national age- and sex-specific incidence rates to district populations. Global Moran’s I²⁸ and mapping were used to assess spatial autocorrelation. Because 24% of districts had zero cases, raw SIRs were unstable. To address this, we applied a negative binomial regression with spatial P-splines, which accounts for overdispersion and provides smoothed SIR estimates across Ukraine.²⁹ Analyses were implemented in R (mgcv package).³⁰ To assess the robustness of the smoothing, we compared spatial clusters in raw SIRs, identified using the local indicators of spatial association (LISA),³¹ with those in the smoothed estimates (online supplemental figure S3).

Finally, we assessed the association between radiation exposure and thyroid cancer incidence. We estimated the ERR per gray (Gy) of district-level average thyroid dose from iodine-131 (¹³¹I) using a conditional autoregressive negative binomial model (ICAR-NB) that accounts for spatial dependence between neighbouring districts. Baseline rates were modelled as a function of sex, age at exposure and a district-specific ICAR spatial random effect (ζᵢ). The district coefficient is a spatial neighbourhood effect representing a weighted average of spatial effects in neighbouring districts, and thus accounts for spatial autocorrelation due to unobserved shared characteristics across districts. ERR was modelled as a function of thyroid dose, sex and age at exposure. Bayesian inference (10 000 posterior draws) was implemented using the inlabru package in R,³² and results are presented as posterior medians with 95% credible intervals. For comparison and to address potential sources of bias, we also fitted a non-spatial NB model. More model specification details are provided in online supplemental material.

Results

District average thyroid radiation doses received in 1986: geographical distribution

Figure 1A illustrates the distribution of absorbed thyroid radiation doses in Ukrainian districts in 1986, interpolated from settlement-level data by age and sex to 627 harmonised districts and weighted by the 2001 population structure. The national population-weighted average dose was 46.2 milligray (mGy), with brown and green shades indicating values above and below this baseline. The mapped doses correspond closely to the pattern of ¹³¹I ground deposition caused by shifting winds after the accident.^25,26 Traces of Chornobyl releases are evident in the northwestern, northeastern and central districts, with the highest doses near the plant. Interestingly, while the central areas show a less pronounced impact than the ‘northern belt’ (Kyiv, Zhytomyr, Rivne and Chernihiv oblasts), the thyroid doses there were still higher than the national average in Ukraine. Overall, 47.8% of the population aged 15 and older lived in 2001 in areas with 20–40 mGy and 31.4% in regions with 40–80 mGy. Smaller percentages resided in lower- or higher-dose areas (online supplemental table S1).

Figure 1. District-level average thyroid radiation dose from iodine-131 (¹³¹I) in 1986 (A), raw standardised incidence ratios (SIRs) of thyroid cancer (B) and smoothed SIRs at ages 15 years and older in 2001 (C). Panel A shows the estimated district-level average absorbed thyroid dose (mGy) from ¹³¹I following the Chornobyl accident. Panel B presents the observed (raw) SIRs of thyroid cancer in 2001. Panel C shows spatially smoothed SIRs that highlight districts with statistically significantly higher or lower thyroid cancer incidence relative to the national average (SIR=1). Prepared by the authors. Base maps: Ministry for Regional Development, Building and Housing of Ukraine; Natural Earth. GAM, generalised additive model.

Figure 1B shows the raw SIRs of thyroid cancer in individuals aged 15 and older in 2001, with considerable variability due to small or zero case counts in many districts. Global Moran’s I confirms spatial autocorrelation in thyroid cancer distribution, with significant coefficients for both sexes: moderate for women (0.31, p<0.001) and weaker but significant for men (0.10, p<0.001).

Next, using NB regression smoothing with 2D P-splines, we reduced the large variability in the raw SIRs by incorporating information from neighbouring areas and considering spatial dependencies in the data. The resulting figure 1C with smoothed SIRs of thyroid cancer highlights areas with statistically significant higher or lower incidence rates compared with the national average level, which is taken as one. Our map revealed elevated incidence levels not only in the northern regions but also in some parts of Vinnytsia and Cherkasy oblasts, which are situated significantly south of the Chornobyl plant. Regions with significantly lower SIRs of thyroid cancer incidence are observed in far-western regions, particularly Ivano-Frankivsk and Chernivtsi. The smoothed SIR model was tested for zero excess and spatial autocorrelation, and both were found to be non-significant.

Association between district average thyroid dose and thyroid cancer incidence at ages 15 years and older in 2001

Table 1 presents estimates from the NB model and the conditional autoregressive NB (ICAR-NB) model. The effect of radiation exposure on the risk of thyroid cancer was found to be significant for both sexes. Specifically, the estimated ERR per grey (Gy) for exposure at ages 0–4 was 48.09 (95% CI 17.05 to 79.11) in the NB model and 24.85 (95% CI 7.78 to 41.78) in the ICAR-NB model. ERR estimates were higher for females than males in both models. However, the female-to-male ERR ratios (1.53 and 1.23, respectively) had wide credible intervals, indicating that sex differences were not statistically significant.

Table 1. Estimates of the effect of district average absorbed ¹³¹I dose to the thyroid gland on thyroid cancer among Ukraine’s population in 2001.

Description	Parameter*	Non-spatial excess relative risk model (NB)		Excess relative risk model with spatial ICAR extension (ICAR-NB)
		Median estimate	95% CI	Median estimate	95% CI
Baseline incidence rate for exposure at ages 0–4 per million person-years	$\mathrm{e}\mathrm{x}\mathrm{p}\left({\alpha }_0\right)\times {10}^6$	2.60	(1.59 to 4.33)	3.20	(2.08 to 4.95)
Female	$\mathrm{e}\mathrm{x}\mathrm{p}\left({\alpha }_0+\frac{{\alpha }_s}{2}\right)\times {10}^6$	4.65	(2.81 to 7.79)	6.14	(3.97 to 9.46)
Male	$\mathrm{e}\mathrm{x}\mathrm{p}\left({\alpha }_0-\frac{{\alpha }_s}{2}\right)\times {10}^6$	1.46	(0.87 to 2.49)	1.67	(1.07 to 2.62)
Ratio female to male	$\mathrm{e}\mathrm{x}\mathrm{p}\left({\alpha }_s\right)$	3.18	(2.60 to 3.88)	3.67	(3.13 to 4.31)
Linear coefficient of age at exposure effect on log incidence	${\alpha }_{a1}$	0.15	(0.12 to 0.17)	0.14	(0.12 to 0.16)
Quadratic coefficient of squared age at exposure effect on log incidence	${\alpha }_{a2}$	−0.002	(−0.002 to −0.001)	−0.002	(−0.002 to −0.001)
Excess relative risk per Gy for exposure at ages 0–4	${\beta }_d$	48.09	(17.05 to 79.11)	24.85	(7.78 to 41.78)
Female	${\beta }_d\times \mathrm{e}\mathrm{x}\mathrm{p}\left(\frac{{\beta }_s}{2}\right)$	58.76	(21.95 to 98.55)	27.43	(9.15 to 46.85)
Male	${\beta }_d\times \mathrm{e}\mathrm{x}\mathrm{p}\left(\frac{{-\beta }_s}{2}\right)$	38.48	(12.60 to 72.46)	22.23	(6.24 to 42.94)
Ratio female to male	$\mathrm{e}\mathrm{x}\mathrm{p}\left({\beta }_s\right)$	1.53	(0.90 to 2.60)	1.23	(0.70 to 2.19)
Multiplicative change in excess relative risk with 5-year increase in age at exposure	$\mathrm{e}\mathrm{x}\mathrm{p}\left({\beta }_a\times 5\right)$	0.79	(0.72 to 0.88)	0.76	(0.67 to 0.87)
Negative binomial overdispersion	$\varphi$	0.459	(0.344 to 0.602)	0.01	(<0.001 to 0.085)
Test for 0 inflation†		1.003	(p=0.338)	1.004	(p=0.127)
Test for spatial autocorrelation‡		0.047	(p<0.001)	−0.008	(p=0.071)
Test for overdispersion§		2.021	(p=0.004)	0.9121	(p=0.311)
Deviance information criterion		9630.9		9305.6

^*See online supplemental material for the full model equation and definitions of parameters.

^†Ratio of observed to expected zeros under negative binomial model (H0: no excess 0’s).

^‡Moran’s I (H0: no residual autocorrelation under model).

^§SD of residuals (H0: no remaining overdispersion under model).

ERR, excess relative risk; ICAR-NB, intrinsic conditional autoregressive; NB, negative binomial.

The radiation-related risk decreased with age at exposure: a 5-year increase in age at exposure was associated with a multiplicative change in ERR of 0.79 (95% CI 0.72 to 0.88) in the NB model and 0.76 (95% CI 0.67 to 0.87) in the ICAR-NB model, which corresponds to a 21% and 24% reduction in ERR, respectively. The highest ERR was observed among individuals exposed at ages 0–4, followed by a steady decline with increasing age, with no significant differences between males and females (online supplemental figure S1).

Baseline thyroid cancer incidence rates estimated by the ICAR-NB model had a median of 3.20 cases per million person-years (95% CI 2.08 to 4.95) at ages 0–4 years. Incidence was substantially higher in females than in males, as reflected by a significant female-to-male ratio of 3.67 (95% CI 3.13 to 4.31) (table 1). Unlike the ERR, which decreased with increasing age at exposure, the baseline incidence was very low at early ages, peaked around ages 35–39 and declined thereafter (online supplemental figure S2).

Including spatial random effects in the ICAR-NB model did not change the overall pattern of associations but substantially reduced the estimated ERR compared with the non-spatial NB model. In the ICAR-NB model, the ERR per Gy for exposure at ages 0–4 was 27.43 (95% CI 9.15 to 46.85) for females and 22.23 (95% CI 6.24 to 42.94) for males, whereas the non-spatial model produced much higher estimates of 58.76 (95% CI 21.95 to 98.55) for females and 38.48 (95% CI 12.60 to 72.46) for males.

The district-level ICAR spatial random effects (ζᵢ) represent residual spatial variation in thyroid cancer incidence after adjustment for thyroid dose, sex, age at exposure and person-years (figure 2). Positive spatial effects were observed primarily in Kyiv and surrounding districts, with more moderate positive effects in parts of Central Ukraine (Cherkasy and Vinnytsia oblasts), whereas negative spatial effects were observed in southwestern Ukraine. Importantly, these residual spatial effects likely reflect various unmeasured district-level characteristics, rather than any single underlying mechanism. Compared with the non-spatial model, the ICAR-NB model provided a better fit, reduced overdispersion and effectively eliminated residual spatial autocorrelation (Moran’s I: 0.047, p<0.001 vs −0.008, p=0.071), indicating that the model adequately accounted for residual spatial clustering.

Figure 2. District-level ICAR spatial effects exp(ζᵢ) from the ICAR-NB model. The map shows the estimated residual spatial effects after adjustment for thyroid dose, sex, age at exposure and person-years. Values above 1 indicate districts with higher thyroid cancer incidence than expected under the fitted model, whereas values below indicate lower-than-expected incidence. These spatial effects summarise residual spatial dependence associated with unmeasured district-level characteristics shared by neighbouring districts rather than the effect of any single factor. Prepared by the authors. Base maps: Ministry for Regional Development, Building and Housing of Ukraine; Natural Earth. ICAR-NB, intrinsic conditional autoregressive negative binomial.

Discussion

This study is, to our knowledge, the first nationwide study using small-area (district-level) data for all of Ukraine to examine the spatial distribution of thyroid cancer incidence and its association with thyroid radiation dose received in 1986. Compared with previous studies, which primarily focused on heavily contaminated districts in northern Ukraine or used oblast-level data, our analysis provides a more detailed picture of the geographical distribution of thyroid cancer across the country.

First, our small-area analysis revealed considerable within-oblast variation in thyroid cancer incidence. We identified significantly elevated incidence not only in the northern districts surrounding the Chornobyl plant, as documented in previous studies, but also in several districts of Cherkasy and Vinnytsia oblasts, although at substantially lower levels. We also identified a cluster of significantly lower incidence in southwestern Ukraine. To our knowledge, district-level maps of thyroid cancer incidence for the whole territory of Ukraine have not previously been published.

Second, applying spatial smoothing substantially improved the stability of SIRs as it reduced random variation arising from very small case counts in some districts. Even though smoothing may result in a loss of some local details, the maps generated with smoothed SIR estimates effectively revealed spatial trends and enabled us to identify areas with both high and low incidence rates. To assess the efficacy of the applied smoothing technique, we conducted a robustness sensitivity check. Comparing the LISA of raw SIRs with the smoothed SIRs, we observed clear similarities between (1) the high-high and low-low clusters in raw SIRs (online supplemental figure S3) and (2) the corresponding localities in smoothed SIRs (figure 1C).

Third, our findings align with previous studies showing a strong positive dose–response relationship between radioiodine exposure and thyroid cancer risk.^16,17,20,33 However, the magnitude of the ERR/Gy varies across studies due to differences in design, population coverage and modelling. Our estimates are generally higher than most cohort or case–control studies,^18,20,34 while being consistent with other ecological studies. For example, Jacob et al¹⁷ reported a linear ERR per Gy of 18.9 (95% CI 11.1 to 26.7) at age of exposure 10 and attained age 20, and Likhtarov et al¹⁶ found 15.52 (95% CI 3.55 to 27.48) for age at exposure 1–4. Both studies focused on individuals exposed as children and adolescents in the most contaminated oblasts. In contrast, our ICAR-NB model covers the entire Ukrainian population, includes adults at exposure, uses spatially interpolated district-average doses and accounts for overdispersion and spatial clustering. It predicted ERRs/Gy of 24.88 (95% CI 8.08 to 41.99) at ages 0–4, 18.90 (95% CI 6.37 to 31.18) at ages 5–9, 14.43 (95% CI 4.93 to 24.03) at ages 10–14 and 10.90 (95% CI 3.80 to 19.29) at ages 15–19. At ages 5–9 years, our estimate is close to Jacob et al’s¹⁷ linear coefficient, while at older ages (10–14 years) our ERRs remain higher. Compared with Likhtarov et al¹⁶ our continuous age effect produces higher estimates but captures the same sharp decline with increasing age at exposure.

While it is generally acknowledged that thyroid cancers are more common in women, evidence regarding sex differences in radiation-related thyroid cancer risk remains mixed. In our data, females accounted for a substantially larger proportion of thyroid cancer cases than males (female-to-male ratio 4.7:1). This aligns closely with estimates from Bray et al,³⁵ who reported age-standardised incidence rates in Eastern Europe of 2.4 per 100 000 for males and 9.7 for females (ratio of 4.04:1).³⁵ Thyroid cancer incidence can vary by population, age structure and diagnostic practices. Differences in healthcare utilisation and diagnostic intensity may also contribute to a higher female-to-male ratio, given that women generally undergo thyroid examinations and ultrasounds at higher rates, which in turn leads to increased detection of small thyroid cancers.

Studies of radiation-related thyroid cancer have also reported inconsistent findings regarding sex differences in ERR, with some reporting higher ERRs in males^16,17 and others in females,^18,20 although the differences were generally not statistically significant. In our district-level models, the dose–response relationship was statistically significant for both males and females, with higher coefficients observed in females. However, consistent with earlier studies, the lack of a significant difference in ERR between sexes suggests that the available data are insufficient to establish a robust sex-specific effect.

Fourth, by using a spatially adjusted ERR model, we accounted for geographic clustering. The ICAR spatial random effects capture residual spatial variation after accounting for thyroid dose, sex, age at exposure and person-years. They summarise the influence of unmeasured district-level characteristics shared by neighbouring districts, including differences in healthcare access, thyroid cancer screening,^22,23 post-accident population mobility,^36–38 the geographical distribution of Chornobyl survivors (evacuees and liquidators) (online supplemental figures S4 and S5), iodine nutrition^39–41 and other contextual factors. The concentration of positive spatial effects around Kyiv and, to a lesser extent, parts of Cherkasy and Vinnytsia oblasts is consistent with this interpretation. However, the ICAR spatial random effects capture the combined influence of multiple unmeasured district-level factors. Because these effects represent residual spatial dependence rather than individual variables, they should not be interpreted as evidence for any single causal mechanism. Compared with the non-spatial model, the ICAR-NB model produced substantially lower ERR estimates, indicating that non-spatial models may overestimate radiation-related risk by ignoring residual spatial dependence. These findings underscore the importance of incorporating spatial dependence into ecological analyses of post-Chornobyl thyroid cancer risk.

Finally, the observed spatial patterns provide additional insights into the long-term consequences of the Chornobyl accident. The elevated thyroid cancer incidence identified in parts of Cherkasy and Vinnytsia oblasts suggests that the geographical distribution of disease cannot be explained only by proximity to the Chornobyl plant or by district-average thyroid dose. Instead, it likely reflects the combined influence of radioactive fallout together with other unmeasured regional characteristics.⁴² In contrast, the lower incidence observed in southwestern Ukraine is consistent with substantially lower contamination in this region, limited to several districts. Although iodine deficiency may modify susceptibility to radiation-induced thyroid cancer, it is unlikely to explain the lower incidence in southwestern Ukraine, where mild-to-moderate iodine deficiency has been common.⁴¹ Notably, prior studies in Belarus and the Russian Federation^34,43 found higher radiation-related thyroid cancer risk in iodine-deficient areas, while Ukrainian studies^18,20 did not show a significant effect. Furthermore, despite the high thyroid doses occurring in northern Ukraine, not all northern districts, particularly those in northwestern Rivne and Volyn oblasts, formed statistically significant high-incidence clusters. This likely reflects the combination of sparse populations, relatively few thyroid cancer cases, and the smoothing procedures applied to stabilise district-level estimates. These issues are discussed further in the limitations section.

Limitations

Our study has several limitations. First, due to the risk of ecological fallacy, our findings cannot be used to infer causal relationships at the individual level. District-level thyroid dose was significantly associated with thyroid cancer incidence; however, individual radiation doses, residential histories and other personal risk factors were unavailable.

Second, several potentially important district-level determinants could not be incorporated directly into the models, such as post-accident population relocation, the geographical distribution of evacuees and Chornobyl clean-up workers (liquidators), regional differences in iodine nutrition, healthcare access, thyroid cancer screening intensity and other contextual characteristics. Instead, these sources of spatial heterogeneity were addressed indirectly through the ICAR spatial random effects, which account for residual spatial autocorrelation shared by neighbouring districts. Although this approach substantially reduced spatial dependence and lowered ERR estimates compared with the non-spatial model, some residual confounding cannot be excluded. The elevated incidence observed in parts of Cherkasy and Vinnytsia oblasts likely reflects a combination of direct radioactive fallout and various unmeasured contextual factors rather than any single explanation. A more detailed discussion of these potential confounders is provided in online supplemental discussion.

Third, our analysis was restricted to thyroid cancer incidence in a single calendar year (2001). This was necessary because the district-level spatial analyses required population denominators by age and sex, which are available nationally only from the 2001 population census. As Ukraine has not conducted a subsequent national census, and no officially published district-level population estimates by age and sex are available for later years, we were unable to extend the analysis to multiple calendar years. This limitation reduced the statistical power to detect spatial clusters in sparsely populated districts, particularly in the northwestern Polissia regions (Rivne and Volyn oblasts), which experienced radioactive fallout following the Chornobyl accident but had relatively few thyroid cancer cases in 2001. It also limited our ability to assess the joint effects of age at exposure and time since exposure. Future studies incorporating multiple years of thyroid cancer incidence together with harmonised district-level population denominators by age and sex would provide additional insights into temporal changes in radiation-related risk, as well as the spatial patterns of thyroid cancer incidence, particularly in sparsely populated contaminated districts.

Fourth, for population-weighted district average doses for 1986, we used the 2001 population census data to estimate the ages of individuals in 1986. This approach was used because detailed district-level population data for 1986 were unavailable. District average thyroid doses by sex and age were derived via spatial areal-weighted interpolation, assuming even distribution of population and dose within each polygon. However, this assumption is not entirely accurate, as dose likely exists as a gradient across polygons, and populations are typically clustered in settlements rather than uniformly distributed. Despite this, our dose estimates should not be biased because there were significantly more spatial units measuring dose (22 632) than spatial units measuring population and incidence (627). Supporting this, our calculated average dose of 46.16 mGy is close to the oblast-specific arithmetic mean of 54 mGy reported by Masiuk et al.²⁵

Next, our analyses relied on routinely collected data from the NCRU. Previous evaluations have shown that the NCRU meets international quality standards and demonstrates generally good completeness based on indicators such as the mortality-to-incidence ratio, the proportion of morphologically verified cases, the proportion of death-certificate-only cases and timeliness of registration.²⁴ Ryzhov et al reported some under-reporting in older populations. However, this is unlikely to alter our study results, as no evidence of systematic sex-specific bias in registration completeness was found. Furthermore, because our analyses were conducted approximately 25 years after the 2001 diagnosis year, delayed case registration is unlikely to have affected the study findings. Nevertheless, as with any population-based cancer registry, some degree of under-reporting cannot be excluded.

Finally, to our knowledge, the NCRU does not contain information on migration history, body mass index, environmental pollution, lifestyle, dietary habits or other individual-level characteristics that may influence thyroid cancer risk.⁴⁴ Consequently, these factors could not be directly evaluated or incorporated into our analysis. Future cohort or case–control studies gathering detailed individual-level information, together with residential histories and repeated follow-up, would provide important complementary evidence to the ecological associations observed in the present study.

Conclusions

Using district-level data for the entire territory of Ukraine, this study confirms a significant ecological association between thyroid radiation dose and thyroid cancer risk 15 years after the Chornobyl accident. The outcomes, based on the more detailed geographical coverage, provide valuable contextual evidence for future individual-level studies. Our findings also emphasise the importance of using small-area data to detect localised impacts of nuclear disasters, as analyses of larger regional units may overlook some affected areas and their specific contexts. From a methodological perspective, our findings highlight the need to account for spatial autocorrelation when evaluating associations between radiation exposure and thyroid cancer risk, which we addressed through a spatially adjusted ERR model. Nearly 40 years later, lessons from the Chornobyl incident continue to be important for the international community. The population health and environmental consequences still affecting Ukraine and neighbouring countries serve as a stark reminder, particularly in the context of the current occupation by Russian forces of Europe’s largest nuclear power plant in Ukraine’s Zaporizhzhia region.