Impact of SARS-CoV-2 vaccination on cancer trends: a nationwide interrupted time series study

Abstract

Background

A recent population-based cohort study of Seoul residents in South Korea reported increased incidence of several cancers within one year of SARS-CoV-2 vaccination. We examined whether a short-term, population-level change in cancer incidence trends was detectable at the national vaccination rollout, using a design that does not compare vaccinated with unvaccinated individuals.

Methods

We conducted an interrupted time series analysis of Korea National Health Insurance Service data for 36,024,394 individuals born between 1955 and 2004, covering February 2011 to September 2025 (176 months). The intervention point was August 2021, when cumulative first-dose coverage exceeded 50%. Monthly incidence of 13 cancer categories defined by ICD-10 code ranges and of four site-specific cancers (stomach, colorectum, lung and prostate) was compared with counterfactual forecasts from ARIMA/SARIMA models, and segmented regression estimated immediate level and trend changes (significance threshold P < 0.01). The placement of the intervention point was varied from 0 to 18 months, and the resulting estimates were compared with those obtained at 30 placebo points in the pre-pandemic period.

Results

Overall, 89.3% of the population received at least one dose. No immediate level change (relative risk 0.923–1.072) or trend change (0.995–1.005) reached significance for any of the 13 categories, and none reached significance for the four site-specific cancers. Across the lag grid, 2.0% of estimates met the threshold, compared with 4.1% at placebo points. The 99% confidence limits excluded immediate increases larger than 8.9% to 23.0% across the 13 categories and 8.1% to 23.4% across the four site-specific cancers.

Conclusions

The national SARS-CoV-2 vaccination rollout in South Korea was not accompanied by a detectable short-term, population-level change in cancer incidence trends over the 50 months that followed.

Trial registration

Not applicable.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s40364-026-01027-6.

Background

The SARS-CoV-2 pandemic prompted an extraordinary global effort to accelerate vaccine development, leading to the deployment of vaccines in a remarkably short timeframe. These vaccines have been instrumental in reducing infection rates and disease severity [1–3]. However, the accelerated timeline also raised public concerns regarding potential adverse effects, including long-term outcomes [4–6].

These concerns have been amplified by reports suggesting associations between vaccination and cancer. A prominent example is a recent cohort study by Kim et al., which used Korean National Health Insurance data for residents of Seoul to report increased incidence of several cancers, including gastric, colorectal, lung, breast, thyroid and prostate cancer, among vaccinated individuals within one year of vaccination [7]. An excess of newly diagnosed cancers within so short an interval is difficult to reconcile with the induction periods generally observed in carcinogenesis, which are measured in years to decades. The finding therefore calls for examination with a design that does not share the vulnerability of the individual-level comparison on which it rests.

Observational cohort studies comparing vaccinated with unvaccinated individuals are susceptible to confounding and selection bias [8]. A well-recognized example is the healthy vaccinee effect, whereby individuals who choose vaccination differ from those who do not in baseline health, health-seeking behavior and access to care [9, 10]. Such differences can produce associations in either direction that are unrelated to vaccination, and they cannot be fully removed by adjustment.

Interrupted time series (ITS) analysis offers a different approach. It evaluates whether the incidence trend of an outcome in the whole population changes at the time of a population-wide intervention, using the pre-intervention trend as the counterfactual [11]. Because vaccinated and unvaccinated individuals are not compared, selection on individual characteristics cannot bias the estimate; the design remains sensitive to other events that coincide with the intervention, and these must be addressed separately.

In this nationwide ITS study using the Korea National Health Insurance Service database, we examined whether a short-term change in the incidence trends of 13 cancer categories defined by ICD-10 code ranges occurred at the national vaccination rollout in South Korea and over the 50 months that followed. The same analysis was applied separately to the four site-specific cancers among those reported by Kim et al. that fall within these categories.

Methods

Study design and data sources

We employed an ITS analysis, a quasi-experimental design used to evaluate the impact of population-level interventions using longitudinal data. Data for this study were obtained from the Korea National Health Insurance Service (NHIS) database, a comprehensive nationwide claims database covering nearly the entire South Korean population, and records for the study cohort were linked with the national SARS-CoV-2 vaccination registry maintained by the Korea Disease Control and Prevention Agency (KDCA).

Study population and observational period

The study population included individuals born between 1955 and 2004. The observational period spanned from February 2011 to September 2025, comprising 176 monthly observations; the end of observation was set to precede the extraction date by a margin sufficient for claims submission, so that incomplete recent months would not appear as a decline in incidence. Foreign residents were excluded from the analysis. To reduce the inclusion of prevalent cases, so that the cancers analyzed would be incident, we applied a diagnosis-free washout period covering the full history available in the database before the start of observation (January 2002 to January 2011); individuals diagnosed with any type of cancer during that period were excluded.

Outcome ascertainment and cancer classification

The primary outcomes were the monthly incidence rates of 13 cancer categories defined a priori by ICD-10 code ranges (eTable 1). To ensure high diagnostic accuracy, case ascertainment relied on registration in the national co-payment reduction program for critical illnesses, which requires certification of the diagnosis by a physician on the basis of diagnostic test results, such as histopathology, or imaging findings with a specialist's confirmation when biopsy is not feasible. Because four of the cancers for which short-term associations have been reported [7] fall within broader code ranges (breast and thyroid cancers each correspond to a single category), we additionally analyzed four site-specific cancers (stomach, C16; colorectum, C18–C20; lung, C33–C34; prostate, C61) using the same case definition, washout period, denominator and modeling framework (eTable 1).

Intervention definition

The intervention was defined as the population-wide rollout of SARS-CoV-2 vaccination. The primary intervention point was set at August 2021, the month in which cumulative first-dose coverage in the general population exceeded 50% according to national vaccination statistics published by the Korea Disease Control and Prevention Agency (eTable 2). Because eligibility was phased in by age group, the month in which this threshold was reached differed across birth cohorts (June to September 2021); in the birth-cohort-stratified analyses the interruption was therefore placed at the cohort-specific month. The 50% threshold was adopted because the Korean rollout was compressed rather than gradual: within each cohort, coverage rose by 33 to 72 percentage points in a single calendar month, so that coverage curves approximate a step rather than a slowly rising function. The months in which coverage first exceeded 40%, 60% and 70% are given for comparison (eTable 3).

Statistical analysis

Outcome measures

Monthly incidence rates were calculated as the number of newly diagnosed cancer cases per 100,000 population at risk. For the denominator, we used the number of individuals remaining at risk of a first diagnosis of the cancer concerned at the start of each month; individuals were removed from the denominator upon diagnosis of that cancer. For sex-specific cancers (breast, female genital and male genital organs, and prostate), the respective sex-specific population sizes were used. Losses from the cohort due to mortality and emigration were not incorporated

Time series modeling

For each cancer category, autoregressive integrated moving average (ARIMA) or seasonal ARIMA (SARIMA) models were fitted to the pre-intervention period (February 2011 to July 2021), comprising 126 monthly observations and thus more than ten complete annual cycles, and were used to generate counterfactual forecasts for the post-intervention period (August 2021 to September 2025; 50 months) with 95% and 99% prediction intervals. To estimate the association directly, segmented regression using ARIMAX or SARIMAX models was fitted to the full 176-month series, with exogenous terms representing an immediate level change (a step function at the intervention point, estimated as the change in the intercept of the fitted trend; eAppendix 1) and a change in the post-vaccination trend (a slope change). Coefficients (β) were converted to relative risks as exp(β) with 99% confidence intervals

Birth-cohort stratification

The 13 primary categories were additionally analyzed by 5-year birth cohort, with the interruption placed at the cohort-specific month (eTable 2); cohort-specific series with sparse monthly counts were identified by prespecified criteria and are marked in the corresponding tables (eAppendix 1)

Sensitivity analyses

Alternative intervention points

To evaluate the potential impact of the COVID-19 pandemic itself, such as disruptions in healthcare utilization or screening, the analyses were repeated with the interruption placed at February 2020 (the onset of the pandemic in Korea) and March 2022 (the peak of the Omicron wave) (eTable 11)

Interval between exposure and diagnosis

The segmented model was refitted with the interruption shifted by 3 and 9 months, with model orders re-selected at each placement (eTables 8 and 12). Because no such interval is identifiable a priori and any single choice is to some degree arbitrary, we additionally scanned the full range, placing the interruption at every month from 0 to 18 after the primary intervention point and retaining for each cancer category the ARIMA orders selected in the primary analysis, so that variation reflects the placement of the interruption rather than repeated model selection (eFigure 3). The upper bound was fixed a priori at 18 months, which exceeds the one-year window over which the reported short-term associations were observed [7]; the grid was extended to 24 months and its summaries recalculated at upper bounds of 12, 15, 18, 21 and 24 months to examine sensitivity to this choice (eTable 14). When the interruption is shifted later, the months between the original and the shifted placement are treated as part of the pre-intervention period. These estimates therefore show whether the results depend on where the interruption is placed; they do not test for an effect that emerges only after a delay

Negative control

To assess whether a systematic pattern across the grid could arise from the analytic procedure itself rather than from vaccination or other events after the intervention point, the entire grid was repeated on the pre-pandemic segment of each series (February 2011 to January 2020), with the interruption placed at each of 30 placebo intervention points spaced one month apart between January 2015 and June 2017, a period with neither vaccination nor pandemic-related disruption of health care. At each placebo point the model order was selected by the same procedure and then held fixed across the grid, mirroring the treatment of the actual intervention point. The placebo estimates provide an empirical reference for the actual estimates: the proportion of placebo estimates meeting the significance threshold and the distribution of placebo estimates at each placement were compared with the corresponding values at the actual intervention point (eFigures 3 and 4 and eTable 14). The lag grid and the negative control were applied to the 13 primary categories and, separately, to the four site-specific cancers (eFigures 6 and 7 and eTable 14, panel B)

Magnitude of change that can be excluded

To distinguish a change that did not occur from one that the data could not have detected, we calculated, for each of the 13 primary categories and the four site-specific cancers, the limits of the 99% confidence interval for the immediate level change at the primary intervention point, expressed as percentage changes, 100 × [exp(β ± 2.576 × SE) − 1]. These limits indicate the largest immediate increase and the largest immediate decrease compatible with the data; the corresponding limits at selected placements across the lag grid were also calculated (eTable 13). We additionally counted, for each series, the number of post-intervention months in which the observed incidence fell outside the 99% prediction interval of the counterfactual forecast (eTable 13).

Data reporting and software

All statistical tests were 2-sided. To reduce the likelihood of spurious associations arising from multiple comparisons across cancer categories and birth cohorts, a P value < 0.01 was considered statistically significant. Model specification, order selection and residual diagnostics are described in eAppendix 1, and diagnostics for the counterfactual, primary, cohort-stratified, alternative-intervention-point and lagged models are reported in eTables 5–8. Cohort construction and extraction of monthly case counts and denominators were performed in SAS Enterprise Guide, version 8.3 (SAS Institute Inc., Cary, NC), within the NHIS analysis environment. Time-series analyses were performed using R software, version 4.6.1 (R Foundation for Statistical Computing), with the forecast, lmtest, readxl, dplyr, tidyr, purrr and ggplot2 packages.

Ethical considerations

This study was approved by the institutional review board of Korea University Guro Hospital (No. 2021GR0304) and was conducted in accordance with the principles of the Declaration of Helsinki. All standard ethical guidelines were followed. The requirement for informed consent was waived as the research was conducted using fully anonymized and de-identified data.

Clinical trial registration number

Not applicable.

Results

Characteristics of the study population

The study population comprised 36,024,394 individuals (52.5% men) (Table 1), of whom 89.3% had received at least one dose of a SARS-CoV-2 vaccine, with coverage ranging from 84.7% to 93.8% across birth cohorts (eTable 4). Digestive organs (571,921 cases) and thyroid and other endocrine glands (361,021 cases) were the most frequently diagnosed categories over the observation period (Table 1).

Table 1.

Baseline characteristics of the study population, 2011–2025

Characteristic	No. (%)
Total	36,024,394 (100.0)
Sex
Men	18,907,787 (52.5)
Women	17,116,607 (47.5)
Birth cohort
1955–1959	3,268,632 (9.1)
1960–1964	3,429,768 (9.5)
1965–1969	3,159,965 (8.8)
1970–1974	3,867,060 (10.7)
1975–1979	3,764,683 (10.5)
1980–1984	4,306,384 (12.0)
1985–1989	4,116,915 (11.4)
1990–1994	4,014,886 (11.1)
1995–1999	3,466,414 (9.6)
2000–2004	2,629,687 (7.3)
Cumulative incidence rate 1
Lip, oral cavity and pharynx	30,103 (0.08)
Digestive organs	571,921 (1.59)
Respiratory and intrathoracic organs	156,091 (0.43)
Bone and articular cartilage	9,247 (0.03)
Melanoma and other malignant neoplasms of skin	28,693 (0.08)
Mesothelial and soft tissue	21,719 (0.06)
Breast cancer (Among Women) 2	246,504 (1.44)
Female genital organs 2	134,281 (0.78)
Male genital organs 3	85,391 (0.45)
Urinary tract	78,841 (0.22)
Eye, brain and other parts of central nervous system	29,301 (0.08)
Thyroid and other endocrine glands	361,021 (1.00)
Lymphoid, hematopoietic and related tissue	85,055 (0.24)
Site-specific cancers⁴
Stomach	168,688 (0.47)
Colorectum	274,146 (0.76)
Lung	137,212 (0.38)
Prostate³	80,648 (0.43)

Data are presented as number (%) unless otherwise specified. ¹Cumulative incidence represents the percentage of the population diagnosed with each cancer category between February 2011 and September 2025. ²Calculated using the female population (n = 17,116,607) as the denominator. ³Calculated using the male population (n = 18,907,787) as the denominator. ⁴Stomach (C16) and colorectum (C18–C20) fall within the code range for digestive organs, lung (C33–C34) within that for respiratory and intrathoracic organs, and prostate (C61) within that for male genital organs; they are shown separately for reference and are not additional to the total. Analytic results for these four cancers are given in eTable 15

Observed versus predicted cancer incidence trends

The monthly incidence trends for the 13 cancer categories are shown in eFigure 1. Prior to the intervention point (August 2021), most categories demonstrated gradually increasing incidence trends over time, with some exhibiting seasonal variations; thyroid and other endocrine glands declined between 2014 and 2016 and increased gradually thereafter. The fit and residual diagnostics of the counterfactual models are provided in eTable 5.

After the intervention point, the observed monthly incidence rates lay within the 99% prediction intervals of the counterfactual forecasts in all but two of the 850 post-intervention cancer-months examined (13 primary categories and four site-specific cancers over 50 months; Fig. 1 and eFigures 2 and 5): eye, brain and other parts of the central nervous system in November 2021 and female genital organs in October 2022, both above the upper limit.

Fig. 1.

Comparison of observed and predicted monthly incidence of major cancers before and after the intervention point. The solid line represents the observed monthly cancer incidence rate per 100,000 population. The dotted line indicates the counterfactual forecast based on the pre-intervention trend (February 2011 to July 2021), derived from ARIMA/SARIMA models. The shaded areas represent the prediction intervals for the forecast, with the darker area indicating the 95% prediction interval and the lighter area the 99% prediction interval. The vertical dotted line marks the intervention point (August 2021), the month in which cumulative first-dose coverage exceeded 50%. (A) Digestive organs; (B) respiratory and intrathoracic organs; (C) breast; (D) female genital organs; (E) male genital organs; (F) thyroid and other endocrine glands. The remaining seven categories are shown in eFigure 2

Immediate and trend changes at the intervention point

Estimates from the segmented regression are shown in Fig.2; β coefficients and standard errors are provided in eTable 9, the corresponding relative risks by birth cohort in eTable 10, and model diagnostics in eTable 6. No immediate level change reached statistical significance at the intervention point (all P ≥ 0.01): the relative risk (RR) estimates ranged from 0.923 for bone and articular cartilage (99% CI, 0.714–1.195) to 1.072 for eye, brain and other parts of the central nervous system (99% CI, 0.934–1.230); the RR for thyroid and other endocrine glands was 0.958 (99% CI, 0.790–1.161). Similarly, no change in the post-vaccination trend reached significance compared with the pre-intervention trend (all P ≥ 0.01); the RRs for the monthly trend change ranged from 0.995 to 1.005.

Fig. 2.

Relative risks for the immediate level change and the change in the post-vaccination trend in the 13 cancer categories. The forest plot displays the relative risks (RRs) and 99% confidence intervals (CIs) for the (A) immediate change in level and (B) change in the post-vaccination trend for each cancer category. Estimates were derived from segmented regression analysis using ARIMAX/SARIMAX models. An RR of 1.0 indicates no change compared with the pre-intervention trend. The x-axis is presented on a log scale. Abbreviation: CNS, central nervous system

Stratified and sensitivity analyses

Birth-cohort stratification

When stratified by 5-year birth cohorts, with the interruption placed at the cohort-specific month (eTable 2), no immediate level change met the predefined threshold for statistical significance (P < 0.01) in any cohort. Three of the 260 estimates met the threshold, all for the monthly trend change: digestive organs in the 1965–1969 cohort (RR, 0.996; 99% CI, 0.992–1.000; P = 0.0078) and breast cancer in the 1970–1974 (RR, 0.990; 99% CI, 0.982–0.999; P = 0.003) and 1975–1979 cohorts (RR, 0.992; 99% CI, 0.984–0.999; P = 0.004) (eTables 9 and 10). Twenty-nine of the 130 cohort-specific series met the prespecified criteria for sparse data (eAppendix 1) and are marked in eTables 9 and 10; 27 of these were in the cohorts born in 1985 or later

Alternative intervention points

Varying the coverage threshold between 40% and 70% shifted the population-wide intervention point by at most one month and no cohort-specific point by more than one month, except that the youngest cohort did not reach the 70% threshold within the months covered by the national statistics (to September 2021) (eTable 3). When the interruption was placed at the pandemic onset (February 2020) or the Omicron peak (March 2022), no cancer category showed a significant change in the post-intervention trend, and one immediate level change met the threshold: a decrease in female genital organ cancers at the pandemic onset (RR, 0.836; 99% CI, 0.717–0.973; P = 0.002). At the pandemic onset the level RR was below 1.00 for 11 of the 13 categories, and at the Omicron peak it was above 1.00 for 10 of 13 (eTable 11)

Interval between exposure and diagnosis

With the interruption placed 3 and 9 months after the primary intervention point, the only estimate with P < 0.01 was an immediate increase in female genital organ cancers at the 9-month placement (RR, 1.167; 99% CI, 1.001–1.362; interruption placed in May 2022) (eTable 12). Across the lag grid, 5 of 246 estimates (2.0%) met the threshold, four of them for female genital organs at lags of 9 and 12 to 14 months (interruptions placed in May and in August to October 2022) and one a decrease for bone and articular cartilage at a lag of 16 months (eFigure 3 and eTable 16). Beyond the prespecified bound, two further estimates met the threshold, both decreases for female genital organs at lags of 20 and 21 months (eTable 16). Extending the upper bound of the grid from 12 to 24 months did not alter this pattern: the proportion of estimates meeting the threshold ranged from 1.2% to 2.5% (eTable 14)

Negative control

In the negative control, 4.1% of placebo estimates met the threshold over the same lag range, compared with 2.0% at the actual intervention point, and the mean relative risk across the 13 categories at each placement lay within the range of the placebo estimates (eFigure 4 and eTable 14). At upper bounds of 12 to 24 months the placebo proportion ranged from 3.9% to 4.2% (eTable 14)

Site-specific analyses

Stomach and colorectal cancers accounted for 29.5% and 47.9% of the digestive organ category, lung cancer for 87.9% of respiratory and intrathoracic organs, and prostate cancer for 94.4% of male genital organs (Table 1). None of the four site-specific cancers showed a statistically significant immediate level change or trend change (eTable 15). The RRs for immediate level change were 0.984 for stomach (99% CI, 0.850–1.139), 0.945 for colorectum (99% CI, 0.826–1.081), 0.982 for lung (99% CI, 0.781–1.234) and 0.968 for prostate (99% CI, 0.815–1.151); the 99% CIs overlapped those of the parent categories. The observed and counterfactual series for these four cancers are shown in eFigure 5. Across the lag grid, none of the 76 estimates for the four site-specific cancers met the threshold at any lag from 0 to 18 months, compared with 5.1% of the corresponding placebo estimates, and none met it at any upper bound from 12 to 24 months (eFigures 6 and 7 and eTable 14).

Magnitude of change that can be excluded

At the primary intervention point, the 99% confidence limits excluded immediate increases larger than 8.9% to 23.0% and immediate decreases larger than 6.6% to 28.6% across the 13 cancer categories; for the four site-specific cancers the corresponding limits were 8.1% to 23.4% and 15.0% to 21.9% (eTable 13).

Discussion

In this nationwide interrupted time series analysis covering 176 months and more than 36 million individuals, the population-wide rollout of SARS-CoV-2 vaccination was not accompanied by a detectable change in the incidence trends of any of the 13 cancer categories or of the four site-specific cancers examined. Observed incidence remained within the 99% prediction intervals of the counterfactual forecasts in 848 of 850 post-intervention cancer-months, no immediate level or trend change reached the prespecified threshold in the primary analysis, and the few estimates that did so in the birth-cohort-stratified and sensitivity analyses are examined below. The question this design can answer is narrower than whether vaccination is carcinogenic: it is whether the rapid, population-level increase implied by the reported individual-level associations is detectable in national incidence trends over the 50 months that followed the rollout. It was not.

This study was designed as a population-level test of the short-term associations reported by Kim et al. [7]. If those associations reflected an effect of vaccination that began at vaccination and persisted over the following year, the rapid uptake by nearly 90% of the population would, at the reported point estimates, have produced an increase in population incidence larger than the limits excluded here. The hazard ratio of 1.35 reported for thyroid cancer, for example, implies an increase of about 31% in population incidence at the observed coverage of 89% (coverage × [HR − 1]), well above the 16% excluded here (eTable 13). No such increase was observed.

Stomach, colorectum, lung and prostate, four of the six sites reported by Kim et al., were analyzed separately because each falls within a broader ICD-10 category that could mask a change confined to one site. The site-specific estimates were consistent with those of their parent categories: the 99% confidence intervals for each site-specific estimate overlapped those of the category containing it, and across the lag grid none of the 76 estimates for these four cancers met the prespecified threshold at any placement of the interruption from 0 to 18 months, compared with 5.1% of the corresponding placebo estimates. Breast and thyroid cancers, the remaining two sites for which short-term associations have been reported, are represented as single categories among the 13, with relative risks for the immediate level change of 0.993 (breast) and 0.958 (thyroid and other endocrine glands; this category includes rare adrenal and other endocrine cancers, though thyroid cancer accounts for the great majority).

Because the design contrasts population-level incidence before and after the rollout rather than vaccinated with unvaccinated individuals, it is not exposed to the selection mechanisms that affect individual-level comparisons, of which the healthy vaccinee effect is the most widely documented [9, 10]. It is, however, subject to any influence that acts on the whole population at the same time, and the pandemic itself is one such influence, as discussed below.

That influence is visible across the outcome set. When the interruption was placed at two milestones of the pandemic rather than at the vaccination threshold, the direction of the immediate level change was the same in most categories: at the onset of the pandemic in February 2020 the relative risk was below 1.00 for 11 of the 13 primary categories, and at the peak of the Omicron wave in March 2022 it was above 1.00 for 10 of 13. The pandemic-onset placement precedes the vaccination program and cannot reflect it, and consistency in the direction of the immediate level change across cancer categories with different etiologies and induction periods is more readily explained by a population-wide influence on diagnosis than by vaccination; the pattern is consistent with the widely documented disruption of cancer screening and diagnostic services during the pandemic [12, 13] and with their subsequent resumption in Korea, described below.

The influence of the pandemic is most visible in female genital organ cancers, which accounted for most of the departures identified outside the primary analysis; these fall in temporal order. Placing the interruption at the onset of the pandemic (February 2020) gave an immediate decrease of 16% (RR, 0.836; 99% CI, 0.717–0.973), twelve months before the national vaccination program began. Incidence then rose above the 99% prediction interval in October 2022, and in the lag grid the estimates were elevated at lags of 8 to 14 months, corresponding to interruptions placed between April and October 2022, before returning to unity by a lag of 17 months (eTable 16). Because a shifted interruption absorbs the intervening months into the baseline, these estimates contrast the period after mid-2022 with a baseline that already includes the pandemic trough. Beyond the prespecified bound the direction reverses: at lags of 19 to 24 months the estimates fall below 1.00, meeting the threshold at 20 and 21 months (RR, 0.818 and 0.855; eTable 16), so the elevation was not sustained across later placements; this reversal also reflects the placement of the interruption rather than an effect of vaccination. At lags of 8, 9 and 12 to 14 months the estimates also exceeded the highest value at any of the 30 placebo points (eFigure 3, panel H); the placebo points, however, lie entirely in the pre-pandemic period and cannot show what a disruption of care followed by a rebound produces, so this exceedance does not by itself indicate an effect of vaccination. The excursion is consistent with a change in gynecological care around mid-2022.

Taken together, the sequence — a decline coinciding with the onset of the pandemic and a rebound in late 2022 — is more readily explained by the disruption and subsequent recovery of healthcare utilization, including gynecological care, than by the primary vaccination series, which had been completed for most of the population more than a year before the rebound; later doses overlapped the recovery of care in time and cannot be seperated from it by this design (see Limitations). Participation in organized cervical cancer screening in Korea fell from 57.8% in 2019 to 52.2% in 2020, with a 49% drop in March 2020 alone [14]; healthcare utilization recovered progressively through 2022 [15]; and nationally the number of new cancer diagnoses fell in 2020 and rose by 10.8% in 2021 as access to care returned [16]. Cervical cancer accounts for only part of the C51–C58 category, however, and cancers of the uterine corpus and ovary are detected through clinical contact rather than organized screening, so the explanation rests on disruption of gynecological care in general rather than on screening alone.

Outside female genital organ cancers, few estimates departed from the null, and those that met the threshold were all decreases: 3 of 260 birth-cohort estimates, against 2.6 expected by chance, all trend changes indicating slightly slower growth after the rollout; and one estimate for bone and articular cartilage at a lag of 16 months, the category with the fewest cases (9,247), where the estimates at adjacent lags did not meet the threshold (eFigure 3, panel D, and eTable 16). The only other month outside a 99% prediction interval, November 2021 for eye, brain and other parts of the central nervous system, lay above the upper limit; the immediate level change for this category was 1.072 (99% CI, 0.934–1.230) at the primary intervention point and did not meet the threshold at any placement (eFigure 3, panel K).

A null result is informative only to the extent that it excludes effects of a stated size. At the primary intervention point the 99% confidence limits excluded immediate increases larger than 8.9% to 23.0% across the 13 primary categories and larger than 8.1% to 23.4% for the four site-specific cancers, the limits being narrowest for the categories with the most stable pre-intervention series. Smaller changes cannot be ruled out, and effects that would emerge after the observation period are not addressed by this analysis. The same applies to the prediction intervals: their width differed substantially between categories (Fig. 1 and eFigs. 2 and 5) and was greatest for thyroid and other endocrine glands, whose pre-intervention series fell between 2014 and 2016 and rose thereafter (Fig. 1, panel F), so that remaining within the interval excludes less than the confidence limits do.

The negative control addresses whether the estimates in the lag grid that met the threshold go beyond what the analytic procedure alone would produce. Repeating the entire grid at 30 placebo intervention points in the pre-pandemic period, when no vaccination had occurred, yielded a higher proportion of estimates meeting the threshold (4.1%) than the actual intervention point (2.0%). Consecutive threshold-meeting estimates are likewise expected: adjacent placements of the interruption share most of the series, so their estimates are correlated and significant results occur in runs, and the longest run at the placebo points was 13 consecutive placements compared with three at the actual intervention point (eFigure 4). By both summary measures, the estimates that met the threshold at the actual intervention point lie within the range that the procedure generates in the absence of vaccination. The placebo distribution for each cancer category is shown in eFigure 3.

This study has several strengths. First, the use of the Korean NHIS database linked with the national vaccination registry allowed for the analysis of nearly the entire target population. Second, the diagnostic accuracy of the outcomes is high, as cancer cases were ascertained through a national co-payment reduction program requiring physician certification based on diagnostic test results, rather than relying solely on claims codes. Third, the ITS design does not depend on comparing vaccinated with unvaccinated individuals and is therefore not affected by the selection mechanisms that limit such comparisons. Fourth, the pre-intervention period of 126 months provided more than ten complete annual cycles for the counterfactual models, and the sensitivity of every estimate to the placement of the interruption and to the analytic procedure itself was examined explicitly.

Despite these strengths, our findings must be interpreted in light of several limitations.

First, this study utilized an ecological design based on aggregate time-series data. Therefore, the findings represent associations at the population level and cannot be used to infer causality or estimate individual-level risks. We could not account for individual-level risk factors (e.g., smoking, obesity, genetic predisposition) or behaviors. Vaccination entered the analysis only through population coverage, which defined the intervention point; the number of doses, the time since vaccination and the vaccine type were not represented, and exposure was necessarily approximated as a step at a single month even though coverage rose over several months and later doses followed (a third dose from late October 2021, mostly between December 2021 and February 2022; a fourth dose for adults aged 60 years or older from April 2022; and bivalent booster doses from October 2022).

Second, while ITS analysis is robust against many confounders, it cannot exclude the influence of concurrent time-varying factors. The COVID-19 pandemic caused disruptions in healthcare utilization, including cancer screening delays. These disruptions could mask or mimic changes in cancer incidence trends; for instance, an initial decline in diagnoses due to missed screenings followed by a subsequent rebound. The alternative intervention points show that such disruption did occur, and the female genital organ findings are most plausibly a consequence of it; because the rollout and the recovery of health care overlapped in time, their influences cannot be fully separated. The negative control does not resolve this: the placebo points lie entirely in the pre-pandemic segment of each series and therefore cannot reproduce a disruption of care followed by a rebound, so the negative control cannot distinguish an effect of vaccination from the recovery of care, and the attribution of the female genital organ findings to the latter rests on their timing and on the external evidence cited above.

Third, the 50-month post-intervention period is adequate for the short-term question addressed here but not for effects with long induction periods. For most solid cancers, induction periods are measured in years to decades. The shifted-interruption analyses examine the placement of the intervention point, not induction periods, and cannot substitute for longer follow-up. Ascertainment in the final months of the series may also be marginally less complete than earlier because of the interval between diagnosis and claims submission; observation was closed well before the extraction date to limit this, but any residual incompleteness would lower, not raise, the observed incidence relative to the counterfactual.

Fourth, the reliance on administrative data means we lacked detailed clinical information, such as cancer stage at diagnosis, histological subtypes, and biomarkers. Consequently, the categories could not be subdivided by histology, and we were unable to assess whether vaccination influenced the severity or specific subtypes of cancer.

Fifth, residual diagnostics indicated a departure from at least one model assumption for five of the 13 primary segmented models and three of the four site-specific models (eTable 6); where residual autocorrelation remained (digestive organs, lymphoid tissue and colorectum), the confidence intervals, and hence the excluded magnitudes, may be narrower than nominal. For the counterfactual models, residuals departed from normality for two series (eTable 5), and because prediction intervals depend on the normality of residuals, the intervals for those series should be read with caution. In the birth-cohort-stratified analyses, 29 of the 130 series met prespecified criteria for sparse data, 27 of them in cohorts born in 1985 or later, and the corresponding estimates are numerically unstable.

Finally, the denominator was reconstructed each month as the number of individuals still at risk of a first diagnosis of the cancer concerned, but losses from the cohort through death or emigration were not incorporated. Such losses are gradual and monotonic and are therefore absorbed into the pre-intervention trend rather than producing a change at the intervention point, but they leave an inaccuracy of unknown size in absolute rates.

Conclusions

In this nationwide interrupted time series analysis, the SARS-CoV-2 vaccination rollout in South Korea was not accompanied by a detectable short-term, population-level change in the incidence trends of 13 cancer categories or of the four site-specific cancers for which short-term associations have been reported, over the 50 months that followed. Immediate increases larger than roughly 8% to 23%, depending on the category, are excluded at the 99% confidence level. The population-level increases implied by the reported individual-level associations were not observed; effects with longer induction periods lie beyond the scope of this analysis.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ITS

Interrupted time series

NHIS

National Health Insurance Service

KDCA

Korea Disease Control and Prevention Agency

ICD-10

International Classification of Diseases, 10th Revision

ARIMA

Autoregressive integrated moving average

SARIMA

Seasonal autoregressive integrated moving average

PI

Prediction interval

CIs

Confidence intervals

RR

Relative risk

CNS

Central nervous system

Author contributions

JJ conceived and designed the study. YN performed data acquisition and executed the research. All authors contributed to the data interpretation. JJ and YN had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. JJ and YN drafted the manuscript and prepared the initial version. All authors attest they meet the ICMJE criteria for authorship. All authors approved the final version of the manuscript.

Funding

This research was supported by a Korea University Grant and the Bio&Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean government (MSIT) (No. RS-2024-00439160). The funder of this study had no involvement in the direction, content, or conclusions of the analysis.

Data availability

The data used in this study are not publicly available due to legal and privacy restrictions. Researchers may request access to the original raw data through the official data application process of the NHIS and KDCA in accordance with relevant regulations. However, the derived dataset containing monthly aggregated cancer incidence rates used for the interrupted time series analysis, which does not include any individual-level information, is available upon reasonable request from the corresponding author.

Declarations

Ethics approval and consent to participate

This study was approved by the institutional review board of Korea University Guro Hospital (No. 2021GR0304) and was conducted in accordance with the principles of the Declaration of Helsinki. All standard ethical guidelines were followed. The requirement for informed consent was waived as the research was conducted using fully anonymized and de-identified data.

Consent for publication

Not applicable. This study utilized fully anonymized and de-identified secondary data; therefore, informed consent for publication was not required.

Use of AI in publication and research

A generative AI tool was used to correct grammar, spelling, and language refinement. The authors subsequently reviewed and revised the content as needed and take full responsibility for the final version of the manuscript.

Competing interests

The authors declare no competing interests.