Cancer risks among patients with type 2 diabetes: a 10-year follow-up study of a nationwide population-based cohort in Taiwan

A national health insurance program was implemented in March 1995 [24]. In 2007, 22.6 million individuals from a total population of 23.0 million in Taiwan were enrolled in this insurance program. The Bureau of National Health Insurance (BNHI) contracted with 97% of hospitals and 92% of clinics in Taiwan. The datasets of the study consisted of registry for beneficiaries, ambulatory and inpatient care claims, and Registry for Catastrophic Illness from 1996 to 2007 from NHIRD. BNHI performs quarterly expert reviews on random samples of every 50 to 100 ambulatory and inpatient claims in each hospital and clinic. False diagnosis reports entail a high penalty.

Every individual in Taiwan has a unique personal identification number (PIN) code. To protect privacy, data on patient identities are scrambled cryptographically by NHIRD. All the datasets can be interlinked through each individual PIN. Ambulatory care claims contain individual’s gender and birthday, date of visit, and codes for the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes, or A-codes for three primary diagnoses. Inpatient claims contain ICD-9-CM codes for principal diagnosis up to four secondary diagnoses. Registry for Catastrophic Illness database contains data from insurers who suffer from major diseases and are granted exemption from co-payment. All cancer cases registered in the catastrophic illness database should be confirmed by pathological reports. Our study using these data was exempted from institutional review board approval of Public Health, Social and Behavioral Science Committee Research Ethics Committee, China Medical University and Hospital.We conducted a population-based cohort study of two groups. Patients with type 2 diabetes (aged ≥ 20 years) were identified in 1997 to 1998 and followed up until December 31, 2007 or until the first manifestation of any cancer type. Population with type 2 diabetes should have at least three ambulatory claims or at least one inpatient claim with diagnosis of ICD-9-CM code 250 or A-code A181 from 1997 to 1998. To exclude those individuals with type 1 diabetes, we have done two steps. First, we identify all individuals with type 1 diabetes from Registry for Catastrophic Illness database. Second, we excluded those individuals with type 1 diabetes identified in the first step from our study cohort with diabetes. We initially excluded subjects with type 1 diabetes (N = 3,750), any cancer type (N = 135,060), and those aged <20 years (N = 17,679) at baseline from 633,680 patients with type 2 diabetes aged ≥ 20 years. We further excluded those with incomplete information for gender and area registered for NIH program (N = 4,212). The other group was general population, which comprised all insured individuals of the same age without any diabetes. The baseline date or index date for type 2 diabetes group was date of the first outpatient visit or inpatient admission. For general population, the index date was randomly assigned between January 1, 1997 and December 31, 1998 according to index date distribution of type 2 diabetes group. A total of 21,680,686 subjects were obtained from insured population in the data (Figure 1), and we excluded subjects with type 1 diabetes (N = 8,910), any cancer at baseline (N = 2,401,786), any diabetes (N = 2,735,586) from 1996 to 2007, and individuals aged < 20 years (N = 7,055,840). Those with incomplete information for gender and area registered for NIH program (N = 67,315) from 1997 to 1998 were also excluded. Thus, 472,979 patients with type 2 diabetes and 9,411,249 individuals without any diabetes from 1997 to 1998 were included in final analysis.

Sociodemographic factors include age, gender, insurance premium, and urbanization degree of area registered for NIH program. Age was divided into 17 groups with five-year intervals from 20 to >90 years. Gender was categorized into male and female. Insurance premium was categorized according to median of the amounts of insurance premiums, in which median value for these two groups was both 19,200 NT dollars from 1997 to 1998. We used an urbanization indicator developed by Liu et al. [25], who categorized 365 Taiwan towns into seven degrees of urbanization: high- and medium-density urban areas, newly developed area, general area, aging-society area, rural area, and non-developed area.

Cancer cases were identified from ambulatory and inpatient care claims of NHIRD from 1999 to 2007. Incidence rates of lung cancer (ICD-9 code 162; A-code A101), liver cancer (ICD-9 codes 155; A-code A095), colorectal cancer (ICD-9 codes 153, 154; A-code A093, A094), breast cancer (ICD-9 code 174; A-code A113), gastric cancer (ICD-9 code 151; A-code A091), oral cancer (ICD-9 codes 140 to 141, 143 to 146, 148 to 149; A-code A08), prostate cancer (ICD-9 code 185; A-code A124), esophageal cancer (ICD-9 code 150; A-code A090), pancreatic cancer (ICD-9 code 157; A-code A096), cervical cancer (ICD-9 codes 179, 180; A-code A120), nasopharyngeal cancer (ICD-9 code 147; A-code A08-01), small intestine, including duodenum cancer (ICD-9 code 152; A-code A092), gallbladder cancer (ICD-9 code 156; A-code A099-02), retroperitoneum and peritoneum cancers (ICD-9 code 158; A-code A099), laryngeal cancer (ICD-9 code 161; A-code A100), respiratory and intrathoratic organ cancers (ICD-9 codes 160, 163 to 165; A-code A109), bone cancer (ICD-9 code 170; A-code A110), connective and other soft tissue cancers (ICD-9 code 171; A-code A114), skin cancer (ICD-9 code 172; A-code A111), placenta cancer (ICD-9 code 181; A-code A121), endometrial cancer (ICD-9 code 182; A-code A122-01), ovarian cancer (ICD-9 code 183; A-code A123), testicular cancer (ICD-9 code 186; A-code A125), penile cancer (ICD-9 code 187; A-code A129-02), bladder cancer (ICD-9 code 188; A-code A126), kidney cancer (ICD-9 code 189; A-code A129-04), brain cancer (ICD-9 code 191; A-code A130), Hodgkin’s disease (ICD-9 code 201; A-code A140), leukemia (ICD-9 codes 204 to 208; A-code A141), and carcinoma in situ (ICD-9 codes 230 to 234; A-code A16) were estimated for type 2 diabetes group and general population. The incidence rates of cancers were estimated using number of new cancer cases identified by NHIRD from 1999 to 2007 as numerators and total person-years from individuals with type 2 diabetes and without any diabetes during follow-up period as denominators.

Person-years of two populations were calculated from baseline to the occurrence of specific cancers or closing date (December 31, 2007). SIRs and 95% confidence intervals (CI) were estimated for cancers by using Poisson regression analysis and gender, area registered for NIH program, and age were adjusted. A sensitivity analysis was performed under two conditions. For the first condition, same comparisons were made except in the excluded cancer cases identified in 1999. These cancer cases were excluded because patients are very likely to have cancers at baseline and have not been diagnosed, thereby ruling out the possibility of effect–cause relationship between diabetes and cancer. The second condition included the use of Registry for Catastrophic Illness database to identify cancer cases confirmed by pathological reports to estimate SIRs. PAFs for site-specific cancer incidence caused by diabetes were calculated for each gender by using previously published prevalence estimates of diabetes in Taiwan [2] using the same dataset as the current study with the following formula [26]:

PAF = diabetes prevalence × (RR - 1)/[1 + diabetes prevalence × (RR - 1)]. All statistical analyses were performed using SAS version 9.2 software (SAS Institute, Inc., Cary, NC).

This study has several merits. First, this study is considered a large study that involved estimation of SIRs for cancer patients with type 2 diabetes. Thus, this study has sufficient capability to detect the effect of type 2 diabetes and to adjust according to several risk factors, such as age, gender, insurance premium, and urbanization degree of area registered for NIH program through standardization. Although Asia Pacific Cohort Studies Collaboration (APCSC) has examined associations between diabetes and cancer mortality with a large sample size (Lam et al., [28]), our study has two advantages. One is that participants of APCSC are from thirty-six cohort Asian and Australasian studies with various ethnic origins, which may modify associations between diabetes and cancer incidences. The other is that APCSC has focused on cancer mortality, and cancer incidence has not been considered. Second, NHIRD included all diagnosed records. Thus, we can accurately determine cancer incidence and minimize the number of subjects in the cohort who were lost during follow-up period. Third, data with one-year left-censored for exploring the possibility of reverse causality had a negligible effect on original estimates. In addition, most of estimated SIRs are similar to those obtained from analysis, in which cancer incidences obtained from Registry for Catastrophic Illness database were used. The consistent findings from our sensitivity analysis indicated that our results are robust.

Several limitations of the study were also observed. First, we cannot obtain data of behavioral factors, such as smoking, alcohol consumption, obesity, body mass index, and physical activity. In addition, we cannot determine familial risks for diabetes to explain effects of genetic and environmental factors. Thus, independent effect of type 2 diabetes on cancer cannot be established. However, our study allows for rate comparison by adjusting for population structure of age, gender, insurance premium, and area registered for NIH program, which can be performed as the first step of this line of research. Second, diabetic patients may have taken medicine that affected cancer risks. Previous studies have also indicated that glucose-lowering medicines, such as metformin, may reduce risks of cancers in diabetic patients. On the contrary, sulfonylurea drugs or insulin are associated with increased cancer risks [46]. Thus the strength of association between type 2 diabetes and cancer estimated for different populations depend on prevalence of anti-diabetes medication in population with diabetes. Although we did not have information regarding glucose-lowering medications, it won’t confound our estimation for association between type 2 diabetes and cancer in our population.