Net costs of breast cancer in Colombia: a cost-of-illness study based on administrative claims databases | Cost Effectiveness and Resource Allocation

Setting and design

This study uses a retrospective COI approach based on prevalent cases to establish the net costs of BC borne by the contributory regime in 2019. For services included in the HBP (HBP services), net costs are estimated using individual-patient data from existing administrative databases. For services not included in the HBP (non-HBP services), quantities and prices are extracted from aggregated data using the public information system provided by the Ministry of Health (SISPRO from the Spanish Sistema Integrado de Información de la Protección Social) and available tariff manuals. Finally, total net costs are calculated as the sum of the costs of HBP and non-HBP services. All costs are transformed to 2019 International US dollars using the Purchasing Power Parity conversion factor for Colombia published by The World Bank (1,343.6 Colombian pesos per international dollar) [22]. All costs are expressed as the mean monetary values paid by the health system for the provision of health services to patients with BC during 2019, regardless of the time since diagnosis or the severity of the disease. Given previous reports of substantial differences between costs estimates for patients with recent diagnosis or in proximity to death, this study explores 2019 mean costs for patients in the study cohort who received a recent diagnosis (defined as those with a first diagnosis of BC during 2018), and for patients with proximity to death (defined as those who died during the first semester of 2020). All analyses are made using Stata MP^® 14.0 and Microsoft^® Excel^® for Microsoft 365 MSO licensed to the National University of Colombia.

Data sources

This study uses data from existing administrative databases within the health system. The main source of data is the Database for the Study of the Sufficiency of the Capitation Unit (UPC from the Spanish Base de datos para el estudio de la Unidad por Capitación). The UPC database is used by the Ministry of Health to adjust the capitation payments received by HMOs according to the level of use of HBP services of their affiliates. The available UPC database reports patient-level data on the use of HBP services from approximately 80% of the contributory regime and includes the costs paid by the health system, the diagnosis associated with each service using the International Classification of Diseases (ICD-10), and the 10th revision jointly with the basic demographic characteristics. The UPC database is reported by HMOs to the Ministry of Health, and its data are collected at the point of delivery by health care providers. The process of data cleaning and validation of the UPC database is described in detail by Bolivar et al. [23].

All individuals in the UPC database are identified using an anonymized identification number that allows us to link their data to the Unique Registry of Affiliates (RUAF from the Spanish Registro Único de Afiliados) and the Unique Database of Affiliation (BDUA from the Spanish Base de Datos Única de Afiliación). RUAF allows health professionals to report data on all births and deaths, and its data are key to estimating, among others, mortality indicators of interest for the Colombian Government. The BDUA contains data from all individuals affiliated with the health system, and its data are essential for all reimbursements made by the Ministry of Health to HMOs [24]. Finally, this study uses aggregated data from the My Prescription Database (MIPRES from the Spanish Mi Prescripción) available through SISPRO [25]. MIPRES is a web tool that allows health professionals to report the prescription of non-PBS services, and it supports all claims for non-PBS services made by the HMOs.

Identification of patients

To identify patients with BC affiliated with the contributory regime, this study used the electronic algorithm previously validated by Saldaña et al. [26]

Saldana et al. compared the incidence estimates of breast, stomach, and colorectal cancer obtained by several electronic algorithms in the UPC database with the estimates published by a prospective, multicentric, cancer registry in Colombia known as Infocancer. The electronic algorithms that yielded the incidence estimates nearest to the ones published by Infocancer were selected as the recommended algorithms for research within the UPC database.

For BC, the persistence of ICD-10 codes for at least four months and at least one BC-specific procedure (the “specific” algorithm) was selected as the recommended electronic algorithm and was used in this study for the main analysis. To estimate the robustness of these results to changes in the electronic algorithm, we calculated the prevalence estimates and the number of cases detected using the persistence of ICD-10 codes for at least four months without the criteria for BC-specific procedures (the “sensitive” algorithm). The full electronic algorithms used in this study are described in Table S1 in the additional file.

All women who fulfilled the specific algorithm at any time from 2015 to 2019 and received at least one health care service during 2019, regardless of the stage of the disease or its histological classification, were included in the main cohort and classified as “exposed”. Unexposed individuals were selected among all individuals affiliated with the contributory regime who did not receive any ICD-10 code of BC from 2015 to 2019 and received at least one health care service for any other reason during 2019. Data regarding unexposed individuals without any consumption of health services during 2019 (i.e., non-users), and who did not receive any ICD-10 code of BC from 2015 to 2019, were collected from the corresponding BDUA database.

Finally, prevalence estimates were calculated as follows:

where \({P}_{ar}\) is the prevalence estimate per age group a, and region r, \({n}_{ar}\) is the number of exposed individuals (i.e. women who received at least one BC related health service per month for at least four different months at any time from 2015 to 2019), who were alive on January 1st, 2019 and identified in the UPC database; and \({N}_{ar}\) is the number of women affiliated to the EAPBs with data available in the UPC database and identified using BDUA.

All costs accrued by exposed and unexposed individuals and borne by the health system, from January 1st to December 31st, 2019, were considered in the analysis. Given that, Colombia provides compulsory insurance coverage for all its citizens and that the UPC database contains data on most individuals affiliated with the Contributory regime, no administrative censorship or losses to follow-up were considered in the cohort.

To explore plausible differences in costs during 2019 between women within the first year of diagnosis of BC or at the end of life, this study identified the subgroups of patients who received their first diagnosis of BC during 2018 or who died during the first semester of 2020 within the main cohort. Given that the available UPC database contains data regarding approximately 80% of the individuals affiliated with the contributory regime, this study assumes that the individuals in the remaining 20% of the contributory regime present a similar distribution of disease conditional prevalence, demographic characteristics, and costs.

Estimation of net costs associated to non-HBP services

To estimate the net costs associated to the delivery of non-HBP services, this study uses data on quantities from the MIPRES database through SISPRO and data on prices from the national tariff manuals [25, 27, 28]. Equation 4 summarizes the approach used to calculate the costs of non-HBP services:

where \(\mathop q\limits^ -\)_iras is the mean number of non-HBP services delivered per patient by region, age group and service, p_s is the price of each service according to the data published by the Ministry of Health in national tariff manuals, and n_ra is the expected number of patients per region and age group. To decrease the risks of a misallocation of services to the main disease, this study only considers the costs associated with the delivery of targeted therapies (that includes immunotherapy, hormonal therapy, and standard chemotherapy) in this category. We believe that this is a valid assumption since most non-HBP costs are caused by the delivery of targeted therapies, and these are rarely used for other conditions unrelated to cancer. The full list of included non-HBP therapies is shown in Table S2 in the additional file.(20). All estimates are reported with their corresponding 95% confidence intervals (95% CI), assuming a uniform distribution for prices and quantities identified in the MIPRES database as the population parameters.

Estimation of net costs associated to HBP services

A failure to adjust for comorbidities increases the risk of double counting due to a misallocation of all expenditures to the main diagnosis. To decrease this risk, this study uses an approach based on first estimating the marginal costs of delivering HBP services after adjusting for comorbidities (τ) and then multiplying them by the expected number of individuals (n) per geographical region r and age group a. Equation 1 summarizes this approach:

“To estimate τ, this study uses a nearest neighbor propensity score matching (PSM) within a caliper of 1 × 10 − 3% points without replacement in the main analysis.”. PSM allows us to decrease the risk of a misallocation of costs by comparing individuals with and without a given condition, but otherwise, a similar risk of use of resources [19]. The methods used to estimate attributable costs in COI studies using PSM have been described previously [29]. In summary, propensity scores per individual i are defined as the conditional probability of assignment to a particular treatment (W = 1) versus no treatment (W = 0) given a vector of observed covariates x_i:

To estimate the propensity scores, this study uses logistic regression analysis with a binary variable that indicates the disease status as the dependent variable and comorbidities (using the Charlson Comorbidity Index), age, region of residence and employment status (defined as employed or unemployed during the last 12 months) as explanatory variables. Given that data regarding income during the last 12 months is available for approximately 80% of the entire cohort only, we performed a sensitivity analysis to compare our PSM estimates for the full cohort and for the subgroup with available data regarding average income (see Table S3 in the additional file). Finally, the marginal cost (τ) is defined as the mean of the expected differences across all the matched pairs. Equation 3 summarizes this approach:

where M_ra is the number of matched pairs and c_ira and c’_ira are the costs of delivering health services per individual i in the age a and region r group, with and without BC, respectively.

The quality of the matching was assessed by calculating the remaining differences across the covariates. Standardized differences between individuals with and without BC after matching of greater than 20% were considered unacceptable (see Table S4 in the additional file).

The robustness of estimations of marginal costs were tested using multilevel ordinary least squares (OLS) mixed models, with a random intercept for department at level two and individual affiliates at level one. The fitted mixed OLS regression model is as follows:

where COSTS_ij represents the costs accrued throughout 2019 by the individual i in the department j where most health services were delivered; BREAST_ij represents a dummy variable that can take values of one if the individual is classified with diagnosis of breast cancer, or zero otherwise; AGE represents the age in years, INS is a nominal variable that represents the insurer; JOB represents a dummy variable that can take values of one if the individual i was employed during the last 12 months, or zero otherwise. The vector Z includes dummies for all comorbidities defined by the Charlson Comorbidity Index; u_j represents a random intercept that varies according to the department where most services were delivered during the study; and \(\in\)_ij an error term clustered by individual. The OLS model included all variables available in the dataset to describe baseline sociodemographic and clinical characteristics, therefore no variable selection strategies were used to specify the final model.

All estimates are reported with 95% confidence intervals (95% CIs) using robust standard errors and are reported by category of service and setting. Exploratory estimations of costs are made for the subgroup of patients who received the first diagnosis of BC during 2018 and for the subgroup of those who died during the first semester of 2020.

Estimation of total net costs

Finally, total attributable costs by age group and region are calculated by summing the costs due to the delivery of HBP and non-HBP services as calculated in Eqs. 1 and 4. All estimates of total attributable costs and their corresponding 95% CI are presented by category, setting of delivery and exploratory subgroups of interest.