Web appendix: Supplementary material Table 1 Specific diagnosis codes used to identify bladder cancer cases in each dataset Finland and Sweden and UK GP-HOSP datasets Netherlands hospital and cancer registry dataset ICD-9 CM 1 Netherlands GP dataset UK GP datasets READ 3 ICD-10 CM 1 ICPC 2 Malignant neoplasm of the bladder C67 188 U76 B49, 7B2C700, 7B2CE00 Carcinoma in situ of the bladder D09.0 233.7 * B837, Neoplasm of uncertain or unknown behaviour of the bladder (used in sensitivity analysis only) D41.4 236.7 239.4 selected BB4 codes * B917, BA04 1) International Classification of Diseases, 9 th revision or 10 th revision 2) International Classification of Primary Care 3) Coded thesaurus of clinical terms used by general practitioners in United Kingdom * Text mining of the database contents used to identify in situ carcinomas and neoplasms of uncertain or unknown behaviour of the bladder.
Data sources and construction of the study database The study was undertaken using linkage of drug prescribing or dispensing databases to relevant databases including (i) cancer registries, (ii) general practitioner records and/or hospital discharge records, (iii) death records, (iv) reimbursement decision data, and (v) immigration and emigration records. This resulted in the creation of 6 non-overlapping population datasets (1 dataset for Finland, 1 dataset for Sweden, 2 datasets from Netherlands and 2 datasets from UK): Finland: A Finnish nationwide dataset (Finland dataset) was constructed from the Finnish prescription register, the registry for reimbursed medications, the Finnish cancer registry, the Finnish hospital care register, the institutional care register, and the Finnish causes of death register. In Finland dataset morbidity data was based on inpatient hospitalizations and outpatient hospital visits. Sweden: A Swedish nationwide dataset (Sweden dataset) was constructed from the Swedish prescribed drug register, the Swedish cancer register, the National patient register, the Swedish cause of death register, the National diabetes register and the Swedish total population register. In Sweden dataset morbidity data was based on inpatient hospitalizations and outpatient hospital visits. Netherlands: The PHARMO Database Network was used. This was based on probabilistic linkage between the PHARMO pharmacy register, the Dutch hospital admissions register, the Central Bureau for Genealogy registry and the PHARMO GP database. The PHARMO Database Network was used to construct 2 datasets: o the Netherlands Hospital-Pharmacy dataset (Netherlands hospital dataset) and o the Netherlands general practitioner dataset (Netherlands GP dataset). In Netherlands hospital dataset morbidity data was based on hospital admissions from the Dutch Hospital Data Foundation for more than 24 hours and admissions for less than 24 hours for which a bed was required. Morbidities were identified based on discharge diagnoses and performed medical procedures. In Netherlands GP dataset morbidity data was based on GP records only. United Kingdom: The CPRD GOLD was used to create 2 non-overlapping datasets: o The UK GP-hospital dataset included patients from English GP practices within CPRD that permit linkage of the GP data to other medical databases, and which had a valid identifier for linkage. In this dataset CPRD GOLD primary care morbidity, laboratory and prescribing data were linked to Hospital Episode Statistic (HES) data, cancer registry data, and death certificate data. o The UK GP dataset only contained the morbidity, laboratory, and prescribing data from GP practices within CPRD. The majority of the patients in this dataset are from practices that do not permit linkage of their data or did not have valid identifiers for linkage in their GP record. A small number were eligible for linkage but had valid CEDs in CPRD GOLD data and not in the CPRD-HES data. These cases occurred because the GOLD-HES data had a shorter study period. In UK GP-hospital dataset morbidity data was based on GP records and inpatient hospitalizations. In UK GP morbidity data was based on GP records only. Drug usage data were based on outpatient prescription data in UK GP, UK GP-hospital and in the Netherlands GP datasets and on outpatient dispensing data in the Netherlands hospital, Sweden, and Finland datasets.
Table 2 Exact matching variables and propensity score variables used to construct the study cohorts and in the outcome analysis. Exact matching variables (3) Propensity score model variables (3+11) Description Use of other thiazolidinediones (other than pioglitazone) prior to cohort entry date (CED) Type of antidiabetic treatment immediately prior to CED classified as no treatment, metformin only, sulphonylureas (SU) only, metformin and SU, insulin with or without any other antidiabetic medication, or other treatment including use of other thiazolidinediones Type of modification in baseline antidiabetic therapy at CED, classified as treatment switch or add-on treatment. Add-on was defined as either an initiation of a new antidiabetic treatment without any current antidiabetic therapies, or adding a new antidiabetic treatment (from different class) to the current therapy. A switch was defined as either a discontinuation of one antidiabetic treatment class or a change from one antidiabetic treatment class to another. Each of the three exact matching variables separately Duration of treated diabetes mellitus at CED with categories: <1 year, 1-<2 years, 2-<4 yr, 4-<6 years, and 6 years History of diabetic complications at CED defined by the following 5 separate binary (No, Yes) variables: Diabetic retinopathy or maculopathy Ketoacidosis Diabetic coma Diabetic lower limb complications Diabetic renal complications History of myocardial infarction or stroke at CED History of congestive heart failure at CED Year of CED as categorical variable Duration of membership in medication database prior to CED (as categorical variable) Number of different antidiabetic drug classes ever used prior to CED Use in outcome analysis Time-dependent Time-dependent Time-dependent
Cohort entry date (CED) was defined as a time within the study period when a patient either initiated a new antidiabetic medication or modified their current antidiabetic treatment. For those exposed to pioglitazone, the time of initiation of pioglitazone was the only valid CED. For those who were never exposed to pioglitazone, the first initiations of new antidiabetic drugs other than a thiazolidinediones were potential CEDs in UK and Netherlands datasets, and all changes in the antidiabetic drug treatment were potential CEDs in Finland and Sweden datasets. Patients exposed to pioglitazone were linked to patients not exposed to pioglitazone by matching to make sure they were at a comparable stage in treatment and disease severity. Therefore, multiple CEDs were assessed per potential non-exposed patient, but each non-exposed patient was only matched to one patient exposed to pioglitazone. Matching algorithm (1:1 fixed ratio and 1:10 variable ratio) In order to minimize channelling bias for treatment assignment, we separately constructed matched cohorts in each database using the following approach Matching was based on exact matching and a propensity score (PS, probability to initiate pioglitazone therapy conditional on variables that affect pioglitazone initiation). The pre-defined set of exact matching and PS variables are described in Table 2. The PS was estimated using a logistic regression model and evaluated at the time of first prescription of pioglitazone for exposed individuals and at all potential CEDs for unexposed individuals to allow selection of a comparable CED. A weighted logistic regression model was used with the inverse of number of potential CEDs as weights. Patients in the pioglitazone exposed group were first matched exactly within the strata of the exact matching variables to patients in the non-exposed group. Within these strata, in random order of the exposed individuals, the closest match was picked from never-exposed group within a distance of ± 0.05 of the exposed individual s PS. If any pioglitazone exposed patients did not have a match, the matching was repeated for these patients by loosening the matching caliper up to ± 0.1 in order to find one match. For the 1:1 fixed ratio matching the above steps were performed only once. For the 1:10 variable ratio matching the process was continued with the following steps: For those in the pioglitazoe exposed group for whom a match was found in the first round, in random order of the exposed individuals, a new non-exposed patient was picked from the non-exposed group utilizing the exact matching variables and the ± 0.05 matching caliper. This was repeated for at most 9 rounds. No loosening of the matching caliper was applied. For the 1:1 fixed ratio matching the possibility of loosening the matching caliper from 0.05 to 0.1 is equivalent to directly applying the wider caliper.
Weights for multiple matched analyses For the multiple (1: up to 10) matched dataset, to account for imbalance due to the varying number of matched non-exposed patients, it was decided to include balancing weights into the analyses. For the pioglitazone exposed patients the weight is 1 For the matched non-exposed the weights are the inverse of the number of non-exposed within the matching strata multiplied by 10. For example, if there are three non-exposed within one stratum, they are assigned a balancing weight of 10/3. The weights are used to validate the success of matching at baseline (weighted standardized differences) and in the Cox s proportional hazards models.
Table 3 Other explanatory variables utilized in the study. Defined as time dependent variables. History of relevant comorbidities Other urinary tract cancers Other cancers Peripheral vascular disease Congestive heart failure Chronic obstructive pulmonary disease History of other relevant medications Statins or statin combinations Angiotensin receptor blockers (ARB) Angiotensin converting enzyme (ACE) inhibitors Drug for benign prostatic hypertrophy (BPH) History of bladder comorbidities Urinary incontinence Urinary tract infection Pyelonephritis Urolithiasis Hematuria Urinary Retention Neurogenic bladder Catheterization Detailed variable definitions are available on request.
Table 4 Pre-planned sensitivity analyses used to assess the robustness of the study results using the nearest matched cohort Sensitivity Description analysis 1 Analyses excluding all bladder cancers occurring within three and within 12 months after CED to allow for bladder cancer latency. 2 Assessment of the impact of adjusting/not adjusting for smoking status, BMI, and HbA1C information. This analysis was performed using a pooled dataset including the Sweden, Netherlands GP, UK GP, and UK GP hospital datasets. Smoking and BMI were fixed at CED for all four datasets. HbA1C was time dependent for UK datasets and Netherlands GP dataset and fixed at CED for Sweden dataset. 3 Comparison of risk estimates from incident T2DM sub-cohort vs. prevalent-only T2DM sub-cohort. The incident sub-cohort only included patients with at least 12 months of database membership before first diabetes treatment and the prevalent-only sub-cohort only included those with less than 12 months. 4 Analysis of the association of ever- vs. never-exposure to pioglitazone with bladder cancer incidence when insulin use is included as a cumulative duration in the adjusted model. 5 Analysis in which the pioglitazone exposure definition was changed from at least one prescription to at least two prescriptions within a six month period. To ensure immortal time bias was not introduced, the CED was moved to be the original CED plus six months. 6 Analysis in which the definition of incident bladder cancer was broadened to include neoplasms of uncertain and unknown behaviour was performed to assess the specificity of any observed association. 7 Analyses limited to datasets with hospital based morbidity information, performed using a pooled dataset including the Finland, Sweden, UK GP-hospital, and Netherlands hospital datasets.
Table 5 Pre-planned and post-hoc stratified analysis used to assess the robustness of the study using the nearest matched cohort Pre-planned stratified analyses Post-hoc stratified analyses Duration of treated diabetes at CED (categories: <1 year, 1-<2 years, 2-<4 years, 4-<6 years and 6 years), Use of other thiazolidinediones (other than pioglitazone) prior to CED (Yes/No) History of renal complications at CED (Yes/no) Gender (Male, Female) Age at CED (categories: 40-49, 50-59, 60-69, and 70) Calendar year at CED (2000-2003, 2004-2007, and 2008-2011) Cancer events identified from cancer register (Yes/No). Source of morbidity data (GP only, including hospital) Diagnosis of CHF prior to CED (Yes/No) Groups based on quintiles of propensity scores (5 groups)
Table 6 Stepwise variable selection algorithm to define the adjusted model. Step Description 1 The starting point is the base model with the following variables: ever- vs. never-exposure to pioglitazone, dataset origin as a categorical variable, all exact matching variables, the PS quintiles, all individual variables used to generate the PS, age at CED, gender, use of metformin, and use of sulphonylureas, use of insulins, use of other antidiabetic drugs each classified as ever- vs. never-exposed 2 A variable qualified as a candidate covariate for adjustment if it fulfilled the following criteria: At least 5% prevalence in the ever- or never-exposed group and a p-value <0.1 for univariate association between the covariate and bladder cancer incidence. 3 A candidate covariate was considered a potential confounder if when added to the base model the relative change in the HR of pioglitazone exposure was at least 10% compared to the base model. 4 All potential confounders were added simultaneously into the base model. 5 One at a time potential confounders were removed to see if a 10% relative change in the HR of pioglitazone exposure remained. If no potential confounder fulfilled the 10% threshold, the one with the smallest relative change in HR was dropped. This process was repeated until no further changes were possible.
Table 7 Time to bladder cancer after CED Time to incident bladder cancer after CED Exposed to pioglitazone N=130 Not exposed to pioglitazone N=153 nearest matched Not exposed to pioglitazone N=970 multiple matched < 3 months 9 (6.92%) 12 (7.84%) 83 (8.56%) 3-6 months 13 (10.00%) 10 (6.54%) 75 (7.73%) 6-9 months 8 (6.15%) 7(4.58%) 54 (5.57%) 9-12 months 2 (1.54%) 7 (4.58%) 59 (6.08%) 1-2 years 31 (23.85%) 40 (26.14%) 202 (20.82%) 2-3 years 18 (13.85%) 26 (16.99%) 169 (17.42%) 3-4 years 18 (13.85%) 17 (11.11%) 119 (12.27%) 4-5 years 13 (10.00%) 10 (6.54%) 81 (8.35%) 5-6 years 11 (8.46%) 11 (7.19%) 58 (5.98%) 6-7 years 4 (3.08%) 8 (5.23%) 35 (3.61%) 7-8 years 2 (1.54%) 4 (2.61%) 19 (1.96%) 8-9 years 0 (0.00%) 1 (0.65%) 12 (1.24%) > 9 years 1 (0.77%) 0 (0.00%) 4 (0.51%)