Data regarding the incidence of cervical cancer between 1993 and 2012 were derived from the Korea National Cancer Incidence Database (KNCIDB) of the Korea Central Cancer Registry (KCCR). The KNCIDB was launched by the Korean Ministry of Health and Welfare in 1980, and is a nationwide hospital-based cancer registry. The KCCR collects data regarding approximately 80%-90% of cancer cases each year from > 180 training hospitals throughout the country [7]. Since 1999, the KCCR has covered the entire population under the population-based cancer registry program. The KNCIDB includes information regarding age, sex, diagnosis date, geographical region, histological type, primary site, and first treatment modality. Detailed information regarding the KNCIDB have been previously reported [8]. Data regarding cervical cancer-related mortality between 1993 and 2012 were obtained from Statistics Korea (http://kosis.kr). Patient records were anonymized and de-identified prior to analysis.

Ethical approval for the research protocol was provided by the Institutional Review Board of the National Cancer Center, Goyang, Korea (NCC2015-0185).

Although the incidence data has high quality and completeness, the majority of mortality data for cervical cancer has been recorded as unspecified uterine cancer deaths [6,9]. Therefore, it is necessary to redistribute unspecified uterine cancer deaths to cervical cancer deaths using a redistribution algorithm estimation of more accurate cervical cancer mortality trends. To accomplish this, we linked the incidence database with the mortality database, then extracted registered cases of malignant neoplasms of the cervix uteri (C53) among deaths from unspecified malignant neoplasms of the uterus (C55) based on the tenth edition of the International Classification of Diseases [10]. For the present study, we adjusted the numbers of cervical cancer-related deaths (C53) using the proportion of registered cases of cervical cancer (i.e., incidence, ‘C53’) to the number of deaths due to unspecified uterine cancer (C55) in each age group. The detailed redistribution algorithm has been documented in a previous report [6].

We excluded the incidence and mortality data from < 20-year-old women because the numbers of cases and cancer-related deaths were very low. The incidence and mortality rates of cervical cancer were grouped into thirteen age groups (most age groups covered 5 years; 20-24 years old, …, ≥ 80 years old). The incidence and mortality data were also arranged in four 5-year periods from 1993-1997 to 2008-2012. The age-standardized incidences and mortality rates were calculated using Segi’s world standard population [11]. The Joinpoint regression analysis program was employed to estimate trends in the age-standardized incidences and mortality rates [12], and the results were expressed as the annual percent changes (APCs) and their 95% confidence intervals (CIs) for each period. Age-period-cohort effects were assessed using a log-linear model to describe trends in the incidence and mortality rate using the intrinsic estimator method [13]. The goodness of fit for each model was estimated using the likelihood ratio test and the Akaike information criterion (AIC). p-values of < 0.05 were considered statistically significant. All statistical analyses were performed using the Stata software ver. 12.0 (StataCorp LP, College Station, TX) and Joinpoint software ver. 4.1.1 (Surveillance Research Program, National Cancer Institute, Bethesda, MD).

The age-standardized incidence of cervical cancer among ≥ 20-year-old women decreased from 32.8 per 100,000 in 1993 to 15.9 per 100,000 in 2012 (APC, –3.9%; 95% CI, –4.2% to –3.6%) (Table 1). The peak age for the incidence of cervical cancer steadily increased from 60-64 years during 1993-1997 to 75-79 years during 2008-2012 (Fig. 1A). The age-specific incidence of cervical cancer exhibited a decreasing trend for all age groups, with the exception of women who were < 30 years old and ≥ 75 years old (Fig. 1B). The incidence of cervical cancer among < 30-year-old women exhibited an increasing trend, and the APC among 20-24-year-old women (APC, 5.6%; 95% CI, 2.9% to 8.3%) was greater than that among 25-29-year-old women (APC, 2.5%; 95% CI, 1.4% to 2.5%) (data not shown).

The age-standardized mortality rate for cervical cancer among ≥ 20-year-old women decreased from 5.2 per 100,000 in 1993 to 2.1 per 100,000 in 2012 (APC, –4.8%; 95% CI, –5.1% to –4.4%) (Table 1). As shown in Fig. 2A, the mortality rate for cervical cancer increased with age, except among elderly people (≥ 70 years old) during 1993-1997. The age-specific mortality rate for cervical cancer exhibited a decreasing trend for all age groups, with the exception of women who were 20-24 years old and ≥ 75 years old (Fig. 2B).

The age-period-cohort model of incidence had the lowest AIC value from the goodness of fit tests (Table 2). The age-period-cohort model of mortality had the lowest AIC value, and was selected as the model with the best fit based on the goodness of fit test results (Table 2). In the age-period-cohort models, the period effects for the incidence and mortality rate of cervical cancer both decreased from 1993 to 2008 (Fig. 3A and B). The risk ratio for cervical cancer incidence and mortality decreased among women who were born between 1928 and 1973 after we adjusted for age and period effects. However, the risk ratio for cervical cancer incidence and mortality increased among women who were born after 1973.

The decreasing cohort effects on the incidence and mortality rate among women who were born from 1928-1973 could be explained by improvements in their socioeconomic status, access to hospital-based treatment, and cervical cancer screening program [15]. Furthermore, the cohort effect that we observed was similar to that reported in Mumbai [16]. In this context, lower socioeconomic status is highly associated with the risk of HPV infection, which is a major risk factor for cervical cancer [17]. Thus, the economic growth in Korea may have resulted in improved access to hospital-based treatments and greater coverage for the cervical cancer screening program. However, the risk ratio for cervical cancer incidence and mortality exhibited an increasing trend among younger cohorts. There are some possible explanations for the increased cervical cancer incidence and mortality among young women in Korea. Specifically, sexual behavior in younger cohorts has changed compared with older cohorts and the age at first intercourse has become earlier [3]. These changes in sexual behavior may have led to the increased prevalence of HPV infection in more recent Korean birth cohorts [4]. Additionally, there have been changes in the nature of cervical cancer in young women. For example, in situ adenocarcinoma or adenocarcinoma has increased among young women in western countries [18-20]. In Korea, adenocarcinoma increased from 1993 to 2001 among young women aged 20-29 years old [1]. Adenocarcinoma is usually located in the cervical canal; therefore, it was relatively hard to detect in cervical cancer screening compared to squamous cell carcinoma [21]. Indeed, it was reported that the proportion of the lymph node metastasis and tumor size have increased among young women aged < 35 years old in Korea [22]. Finally, young women have a relatively low screening rate for cervical cancer and this rate showed slight decreasing trends among women aged 30-39 years old [23]. Taken together, these factors have likely contributed to the rise in incidence and mortality of cervical cancer among young women in Korea.

Cervical cancer screening programs are effective and powerful tools that can reduce the risk of cervical cancer and prevent cervical cancer-related deaths. For example, the implementation of cervical cancer screening programs has decreased the incidence and mortality rate of cervical cancer in Europe and North America [14,24], as well as in Asia [25,26]. Based on the age-period-cohort analysis results, period effects imply that this screening has beneficial effects [14]. However, using period effects as a surrogate marker for the screening effects assumes that the cervical cancer screening was implemented at the same time in all age groups. Although the Korea screening program was implemented in 1996, not all age groups received free access to the program at the same time [1]. In addition, the participation rate varies according to age group [6,23]. Thus, differences in the participation rates may have resulted in the cohorts affecting the cervical cancer incidence and mortality rather than the periods, which were considered a surrogate marker for the screening effects [14].

Our results were internally consistent with the incidence and mortality rate of cervical cancer. In addition, our study population represents the entire population of Korea and our results are biologically plausible based on the results from other studies [16,27]. Nevertheless, this study has several limitations that warrant consideration. First, the completeness of the incidence data from before 1999 may be lower than that from after 1999, and this discrepancy may have introduced bias into our estimated trends and age-period-cohort effects. However, the incidence data from after 1993 were collected by the Gynecologic Oncology Committee of Korea, which launched a gynecological cancer registry in 1991, and it is likely that there were very few discrepancies between the data from 1993-1998 and the data from 1999-2012. Second, the early mortality data are less reliable, and the validity of the cervical cancer-related mortality data would be expected to increase over time. Therefore, we adjusted these data by matching the mortality data with the cancer registry data using a redistribution algorithm. Nevertheless, it is possible that residual errors were included in the cervical cancer-related mortality data, despite the use of the redistribution algorithm. Third, age-period-cohort models always have an internal limitation that is related to identifiability. However, the intrinsic estimator method is a variant of the principal components regression method that provides an unbiased estimation and has a smaller variance than constrained models [28].