Estimating Cancer Burden in India

A detailed description of methods to estimate cancer mortality, incidence, prevalence, and disability-adjusted life-years (DALYs), and the analytical approaches used in GBD 2016 have been reported elsewhere, and are summarised in the appendix (pp 3–16).1 (link), 25 (link), 26 (link), 27 (link), 28 (link), 29 (link) Briefly, the major data inputs to determine cancer mortality in India included the nationwide Sample Registration System (SRS) cause of death data, the Medically Certified Cause of Death data, and 42 population-based cancer registries (appendix pp 6–8). SRS verbal autopsy cause of death data on 455 460 deaths covering the rural and urban populations of every state of India from 2004 to 2013 were included.⁴ (link) For states with at least one population-based cancer registry, the incidence data were transformed to mortality by multiplying incidence data with an independently modelled urban or rural mortality-incidence (MI) ratio for the respective states. Cancer registry data were used as the gold-standard against which other data sources (SRS or Medically Certified Cause of Death data) were compared. If the other data sources differed substantially from the registry data, they were excluded.1 (link), 30 (link) Because of limitations associated with the Medically Certified Cause of Death mortality data, these were used only when the cancer type was not captured by the SRS cause of death data. The combined data for cancer mortality were used in a modelling approach (CODEm), where an ensemble of plausible models is selected. The CoDCorrect algorithm was used to adjust cancer subtypes to the parent cause and to adjust the sum of predicted deaths from these models for each type of cancers in an age–sex–state–year group to be consistent with the results from all-cause mortality estimation.
The estimation of cancer incidence was driven by registry data from India. The mortality estimates that were derived from transformation of incidence data using the MI ratios, as noted above, were transformed back to incidence after the CODEm and CoDCorrect model adjustments.¹ (link) 10-year cancer prevalence was estimated by modelling survival using the MI ratio as a surrogate for access to cancer care. Incidence cohorts were scaled between a theoretical best and worst case survival using the MI ratio scaling factor. Lifetime prevalence was only estimated for 10 years post incidence¹ (link) and for long-term sequelae from procedures (mastectomy, laryngectomy, stoma, incontinence cystectomy, and prostatectomy). Disability for each cancer was estimated by splitting the prevalence into four sequelae: diagnosis and primary treatment, controlled phase, metastatic phase, and terminal phase. Each prevalence sequela was multiplied with specific disability weights to determine years lived with disability (YLDs). We computed years of life lost (YLLs) from the age-specific mortality estimates and a reference life expectancy for that age group. DALYs, a summary measure of total health loss, were computed by adding YLLs and YLDs for each cancer type for location, year, age and sex.1 (link), 28 (link) The appendix provides a list of data inputs used for these estimations (pp 17–29).
A description of estimation of risk factor exposure and its contribution to disease burden in GBD is available elsewhere.²⁹ (link) Briefly, this includes determination of risk exposure and disease outcome pairs based on available evidence and inclusion criteria, assessment of risk exposure from all accessible data sources, and estimation of disease burden attributable to risks based on published relative risks. Estimates of DALYs for specific types of cancers that were attributable to each risk factor were produced by location, age, sex, and year.
GBD uses covariates, which are explanatory variables that have a known association with the outcome of interest, to arrive at the best possible estimate of the outcome of interest when data for the outcome are scarce but data for the covariates are available.25 (link), 26 (link), 27 (link), 28 (link), 29 (link) This approach was part of the estimation process for the findings presented in this report.