Nation nal Cancer Institute Prevalence Projections: The US Experience State of Art Methods for the Analysis of Population- Based Cancer Data January 22-23, 2014 Ispra, Italy U.S. DEPARTMENT OF HEALTH AND HUMAN SERVICES Angela Mariotto National Institutes of Health
Outline Temporal prevalence projections PIAMOD application to SEER data to estimate 2020 projections of US prevalence by phases of care. Geographic prevalence projections Evaluation of biases in projecting SEER prevalelence proportions to other areas
Temporal Projections: Relevance Prevalence estimates (complete or limited duration) lag few years behind current calendar year. In the US we have January 1, 2010 estimates. Projections are needed: To estimate current year prevalence To project into the future for planning and resources allocation
Temporal Projections: Relevance Questions What is projected economic burden of cancer care in the US? NIH Office of the Director What is the demand for oncologists in 2020? American Society of Clinical Oncology (ASCO) What will be the number of cancer survivors aged 65+ in 2020? American Cancer Society (ACS) Is the prevalence of melaonam stage IIB-IV smaller than 200,000? Food Drug Administration (FDA) application of the finantial incentives for drug developement
Projections require modeling: PIAMOD vs. MIAMOD PIAMOD: uses incidence and survival Easier to control future incidence projections. Incidence projections done outside software. Usually needs to project geographically: registry national MIAMOD: uses mortality data to estimate incidence. Mortality data available nationally Less control over incidence projections which will be based on the projected age, period and cohort incidence model. A 2 step approach could be used: Estimate incidence using MIAMOD. Out of MIAMOD project incidence into the future and input in PIAMOD to estimate prevalence
Projections of the US cancer prevalence through 2020 Projections of the future number of incident and prevalent cancer cases, were derived from survival and incidence data from the 9 registries in the SEER program from 1975-2005 (10% of U.S. population). Step1: Incidence Estimate rates US were incidence applied and projections to US Census from Population SEER incidence projections to estimate the Step annual 2: SEER number Survival of new and cancer US cases in the US. incidence rates were used to estimate prevalence using PIAMOD Survival and US incidence rates were used to estimate prevalence.
Input 1: US Population and Projections Male US population born 1945-1954 40-49 50-59 60-69 70-79 80-84 Data source: US Census Bureau
US incidence estimated from SEER data 1975-2010 applying SEER cancer incidence rates to the US populations by age, sex, and race C N = (C R / P R )* P N 2011-2020 (projections) estimated using different assumptions of future SEER cancer incidence rates Base projections: assume future constant incidence rates Current trend projections: estimate trend in the last 5 or 10 years of data and continue the trend
Input: Incidence SEER rates are applied to the US population by age and year to obtain US cases Observed Base (Constant trend) Current (Projected trend) Projected trend estimated by applying last 10 year annual percent change to future rates
Survival Model Projections: Example Male Colorectal Cancer aged 65-74 years Survival is modeled using mixture cure survival model 1-year 5-year Observed - - - Trend Constant 10-year Year at diagnosis
Prevalence Projections 2010 and 2020 Under Different Scenarios Prevalence (No. of People) 2020 Site 2010 Base Incidence Survival Both All Sites 13,772,000 18,071,000 17,465,000 18,878,000 18,229,000 Female Breast 3,461,000 4,538,000 4,275,000 4,597,000 4,329,000 Prostate 2,311,000 3,265,000 3,108,000 3,291,000 3,132,000 Melanoma 1,225,000 1,714,000 1,971,000 1,724,000 1,983,000 Colorectal 1,216,000 1,517,000 1,327,000 1,575,000 1,376,000 Lymphoma 639,000 812,000 803,000 841,000 831,000 Uterus 586,000 672,000 638,000 667,000 634,000 Bladder 514,000 629,000 576,000 640,000 587,000 Lung 374,000 457,000 392,000 481,000 412,000 Kidney 308,000 426,000 487,000 446,000 511,000 Head & Neck 283,000 340,000 308,000 346,000 313,000 Cervix 281,000 276,000 245,000 277,000 245,000 Leukemia 263,000 340,000 328,000 356,000 342,000 Ovary 238,000 282,000 232,000 296,000 241,000 Brain 139,000 176,000 174,000 185,000 182,000 Stomach 74,000 93,000 80,000 103,000 88,000 Esophagus 35,000 50,000 48,000 62,000 60,000 Pancreas 30,000 40,000 40,000 50,000 50,000 Scenarios Base=Impact of changes in population under current cancer control interventions. Both=Continuing trends in incidence and survival
Increase in prevalence from 2010 to 2020 by annual percent change in incidence rates Increase in Prevalence (%) Decreasing trends -6.0-4.0-2.0 0.0 2.0 4.0 Ovary Cervix Kid dney Trend in Incidence (APC) Melan noma 70% 60% 50% 40% 30% 20% 10% 0% -10% -20%
Prevalence by Phase of Care More useful than overall prevalence for planning, resources allocation and costs of cancer Both care and costs vary drastically in the initial and last year of life phases of care compared to the phase in between (continuing care) Prevalence by time since diagnosis, e.g. prevalence of patients diagnosed 0-2, 2-5 and 5+ years from diagnosis can be a surrogate for prevalence by phases of care. Mariotto, Yabroff et al. JNCI, 2011 Mariotto et al. Cancer Causes and Control, 2006
Estimates of Average Annual Costs Brain Pancreas Esophag Stomach Lung Lymphoma Colorectal Other Head/Neck Kidney Leukemia Bladder Prostate Melanoma Initial Continuing Cancer Death Other Cause Death 0 50 100 150 200 250 300 350 Net Costs in Thousands (2010 US Dollars)
Percent of Survivors in Each Phase of Care in 2010 Pancreas 30,000 374,000 35,000 74,000 283,000 Lung Esophagus Stomach Head & Neck Kidney Colorectal Prostate Leukemia Bladder Lymphoma All Sites Brain Melanoma Breast Ovary Uterus Other sites Cervix 0% 20% 40% 60% 80% 100% 308,000 1,216,000 2,311,000 Initial Continuing End of life
Discussions/Conclusions In the US population changes have the largest effect on the 2020 prevalence projections compared to changes in incidence and survival. PIAMOD attractive because allows for different scenarios of population dynamics, incidence and survival Micro-simulation models (type of CISNET) may provide prevalence projections based on assumptions of future trends for particular interventions
Geographical Projections Collaboration with Daniela Pierannunzio (lead), Roberta De Angelis When cancer registries do not have national coverage: national cancer prevalence can be estimated by applying cancer registry prevalence proportions to the respective populations C N = (C R / P R )* P N This method accounts for differences in age, sex and race between registry and nation, but do not account for other factors, such as socioeconomic status, that may biased national prevalence estimates Objective: evaluate biases in prevalence estimates obtained using this naïve extrapolation method to estimate prevalence in different geographic areas, e.g., county and state level.
Data and Methods Data from SEER-18 Registries at county level 5-yrs and 10-yrs limited duration prevalence al 1/1/2010 modelled using different ecological Poisson regression models by county, age sex and cancer site (all sites, prostate, breast, colorectal and lung) and socioeconomic (SES) variables Model 1: Prevalence ~ Mortality + SES Model 2: Prevalence ~ County SEER naïve projected prevalence + Mortality + SES Model 3: Model 2 by age One round of cross-validation splitting the SEER-18 counties in 2 datasets (training and validation) to assess and compare models accuracy
Data: SES, age-adjusted mortality in US and SEER-17 2000 Socio-economic US All counties (N=400) SEER-17 Training counties (N=200) Validation counties (N=200) attributes (N=3142) At least bachelors degree (%) 16.5 17.5 17.4 17.6 Median fam. income 42,154 43,988 44,416 43,560 Persons below poverty (%) 14.2 14.2 13.9 14.4 Unemployed (%) 5.8 6.2 5.9 6.4 Non-white (%) 12.2 8.7 8.5 8.9 Black (%) 9.0 4.7 4.5 4.9 Minority (incl. white Hispanic) ( 18.0 17.3 17.2 17.3 Urban (%) 40.1 46.3 47.2 45.4 Foreign born (%) 3.5 5.2 5.2 5.1 Mortality all causes 870.3 844.9 843.3 846.4 Mortality all malignant cancers 200.2 195.8 195.3 196.3 Colon and rectum mortality 19.8 19.4 19.2 19.6 Lung mortality 61.2 59.5 59.1 60.0 Breast mortality 13.0 13.0 12.9 13.1 Prostate mortality 12.8 12.9 12.8 13.0
Preliminary Results Validation: Goodness of fit All sites Males Data set N Extrapolated Model 1 Model 2 Model 3 Validation 200 3,721 7,494 1,479 1,268 Training 200 3,720 6,905 1,467 1,404 All counties 400 7,441 14,399 2,946 2,672 Lung Male Data set N Extrapolated Model 1 Model 2 Model 3 Validation 199 1,405 862 422 448 Training 198 1,523 745 421 397 All counties 397 2,928 1,608 843 845 Colorectal Male Data set N Extrapolated Model 1 Model 2 Model 3 Validation 194 688 1,351 382 454 Training 186 681 1,010 376 393 All counties 380 1,369 2,361 759 847 Prostate Data set N Extrapolated Model 1 Model 2 Model 3 Validation 191 3,631 5,135 1,841 1,575 Training 186 3,663 4,928 1,382 1,273 All counties 377 7,295 10,063 3,224 2,848 Prevalence counts are compared to the respective observed number of cases using a goodness of fit indicator [ C ˆ j C j ] Cˆ j Area j 2
Geographic Projections Mortality and SES can improve the extrapolated prevalence estimates using a Poisson regression model. Modeling by age can be done for more common cancer Modeling by age can be done for more common cancer sites, with small cells of zero prevalence counts
Acknowledgements National Cancer Institute IMS Rocky Feuer, Robin Yabroff, Joan WarrenJulia Rowland Steve Scoppa, Mark Hachey, Ken Bishop Istitute Superiore di Sanità, Rome, Italy Roberta de Angelis, Daniela Pierannunzio, Riccardo Capocaccia, Lucia Martina
Thank you mariotta@mail.nih.gov