Cancer Survival Query System (CSQS): Making Survival Estimates from Population-Based Cancer Registries More Timely and Relevant for Recently Diagnosed Patients Sept. 20-21, 2010 Methods and Applications for Population-Based Survival Workshop Fascati, Italy Eric J. (Rocky) Feuer, Ph.D. Chief, Statistical Methodology and Applications Branch Division of Cancer Control and Population Sciences National Cancer Institute Some Questions • When someone calls 1-800-4CANCER and asks about the prognosis of a family member who was newly diagnosed, where should the information come from? • How can physicians get a better understanding of the potential impact of competing risks for newly diagnosed cancer patients with significant comorbidities? • Can population-based cancer registry data play a role in answering these questions? Outline I. Statistical Methodology II. Application to Prostate Cancer III. Demonstration IV. Testing Usefulness in Real World Situations I. Statistical Methodology Competing Risks Analysis (Discrete Time) Pi Probability of surviving all causes in interval (i) given alive at (i-1) hci Crude probability of death from cancer in interval (i) given live at (i-1) hoi Crude probability of death from other causes in interval (i) given alive at (i-1) GcM = Cumulative probability of dying of cancer through time interval M M x 1 Pi hcx x 1 i 1 GoM = Cumulative probability of dying of other causes through time interval M M x 1 Pi hox x 1 i 1 Two Data Situations Competing Risks Analysis All of the relevant patient Cancer and other cause of characteristics for both death characteristics are in cancer and other causes are separate data sets in the same data set I. Everything in A Single Data Set • Example: co-morbidity added to SEER through SEER-Medicare linkage – Standard competing risks analysis methods can be used – No assumption of independence of competing risks is necessary – Some restrictions on the parameterization may be necessary • (Example: complicated if the time scales for both causes of death are not the same – e.g. time since dx for cancer and age for other causes) – Minjung Lee will present II. Cancer and Other Cause Mortality Derived from Separate Data Sets • Examples: – Other cause mortality derived from combination of SEER-Medicare and 5% non-cancer matching patients (Angela’s talk) – Other-cause mortality derived from mortality follow-up of National Health Interview Surveys (NHIS) as a function of general health status, functional status, and self- reported conditions – (all ages available!) • Conditional independence is required (conditional on covariates) • Parameterization for each cause is flexible • Covered in this talk! Competing Risks Under Independence d ci Net probability of dying of cancer in interval (i) given alive at (i-1) d oi Net probability of dying of other causes in interval (i) given alive at (i-1) GcM = Cumulative probability of dying of cancer through time interval M x 1 Pi d cx d cx d ox M 1 x 1 i 1 2 GoM = Cumulative probability of dying of other causes through time interval M x 1 Pi d ox d cx d ox M 1 x 1 i 1 2 Assuming uniform deaths from cancer and other causes in the interval. Hakulinen T, Net Probababilities in the Theory of Competing Causes, Scan Actuarial Journal , (1977) Using Relative Survival* Pi d ci 1 Ri (1- interval relative survival for time interval i, i.e. 1 ) Ei d oi 1- Ei (1 - interval expected probability of surviving interval i) GcM = Cumulative probability of dying of cancer through time inteval M x 1 Pi 1 Rx 1 Rx 1 Ex M 1 x 1 i 1 2 GoM = Cumulative probability of dying of other causes through time interval M x 1 Pi 1 Ex 1 Rx 1 Ex M 1 x 1 i 1 2 * Cronin and Feuer, “Cumulative Cause-Specific Mortality for Cancer Patients in the Presence of Other Causes – A Crude Analogue of Relative Survival”, Statistics in Medicine, 2000. Moving from Cohort to Individual • Up to now the equations apply to estimating competing risk survival for a cohort of individuals (e.g. age 60+, Stage II CRC, both genders, all races) • We are interested in customizing the estimates for individual (j) with – Cancer characteristics (zj ) • E.g. Gleason’s score, stage, age, race, comorbidity – Other cause characteristics ( wj ) • E.g. age, race, co-morbidity Customized for individual ( j ) with cancer characteristics ( z j ) and other cause characteristics ( w j ) GcM (z j ,w j ) = Cumulative probability of dying of cancer through time interval M for an individual (j) with cancer characteristics (z j ) and other cause characteristics (w j ) x 1 Ri ( z j ) Ei ( w j ) 1 Rx ( z j ) 1 Rx ( z j ) 1 Ex ( w j ) M 1 x 1 i 1 2 GoM (z j ,w j ) = Cumulative probability of dying of other causes through time interval M for an individual (j) with cancer characteristics (z j ) and other cause characteristics (w j ) x 1 Ri ( z j ) Ei ( w j ) 1 Ex ( w j ) 1 Rx ( z j ) 1 Ex ( w j ) M 1 x 1 i 1 2 Analogue When We Use Cause of Death Information Si ( z j ) net cause-specific cancer survival through interval (i) for an individual with cancer characteristics (z j ), given alive at start of interval (i) GcM (z j ,w j ) = Cumulative probability of dying of cancer through time interval M for individual (j) with cancer characteristics (z j ) and other cause characteristics (w j ) x 1 Si ( z j ) Ei ( w j ) 1 S x ( z j ) 1 S x ( z j ) 1 Ex ( w j ) M 1 x 1 i 1 2 GoM (z j ,w j ) = Cumulative probability of dying of other causes through time interval M for individual (j) with cancer characteristics (z j ) and other cause characteristics (w j ) x 1 Si ( z j ) Ei ( w j ) 1 Ex ( w j ) 1 S x ( z j ) 1 Ex ( w j ) M 1 x 1 i 1 2 II. Application to Prostate Cancer* *Colorectal cancer also available Basics Models fit using SEER 13 + entire state of CA (20.3% of US) from 1995-2005 to allow consistent modern staging over time Si ( z j ) or Ri ( z j ) is estimated using discrete time Cox regression* from SEER, but stratified to accurately capture baseline survival for appropriate subgroups Ei ( w j ) is estimated using the methods described in Angela's talk (but other co-morbidity calculators could be substituted) *Prentice RL and and Glockeler LA "Regression Analysis of Grouped Survival Data with Application to Breast Cancer, Biometrics, 1978. Hakulinen T and Tenkanen L "Regression Analysis of Relative Survival Rates, Applied Statistics, 1987. 3 Staging Groups • Pre-Treatment Clinical – For patients who have not yet been treated – Estimable because for prostate cancer SEER maintains data on both clinical and pathologic staging • Pure Clinical – For patients who elected not to have surgery • Pathologic – For patients who had surgery Prostate Cancer – Extent of Disease • T1 (Clinical Staging only) – T1a: Tumor incidentally found in 5% or less of resected prostate tissue (TURP). – T1b: Tumor incidentally found in > 5% of resected prostate tissue (TURP). – T1c: Tumor found in a needle biopsy performed due to elevated PSA. • T2: Tumor confined within prostate. • T3: Tumor extends through prostatic capsule. • T4: Tumor is fixed, or invades adjacent structures other than seminal vesicles, e.g., bladder neck, external sphincter, rectum, levator muscles, and/or pelvic wall. Prostate Cancer • Inclusion Criteria – Age 94 and under – First Cancer • Staging – Localized (Inapparent) - T1a,T1b,T1c N0 M0 (Clinical only) – Localized (Apparent) - T2 N0 M0 – Locally Advanced I – T3 N0 M0 – Locally Advanced II - T4 N0 M0 – Nodal Disease I - T1-T3 N1 M0 – Nodal Disease II – T4 N1 M0 – Distant Mets – Any T, Any N, M1 (Clinical Only) Strata and Sample Sizes Pre-treatment Clinical Pure Clinical Stage Co-morbidity Co-morbidity All All (Age 66+) (Age 66+) Localized (Inapparent) 34839 109079 25516 63222 Localized (Apparent) 49706 137518 35714 79418 Locally Adv and Nodal 3649 9455 2757 6669 Distant Metastases 3997 9756 3486 8592 Totals 92191 265808 67473 157901 Path Stage Co-morbidity All (Age 66+) Localized 11063 60338 Locally Adv and Nodal 5490 27116 Totals 16553 87454 Prostate Covariates • Substages of Localized (Inapparent) • Substages of Locally Advanced and Nodal Disease • Gleason’s Score (2-7 and 8-10) • Substages x Gleason's Score • Age (cubic spline – flat under age 50 and after age 90) • Race (white, black, other) • Marital Status (married, other) • Co-morbidity – age 66+ (linear – flat at high values ) • Calendar year (linear) – Projected to most recent data year (2005) and then flat to (conservatively) represent prognosis of recently dx patient – Mariotto AB, Wesley MN, Cronin KA, Johnson KA, Feuer EJ. Estimates of long-term survival for newly diagnosed cancer patients: a projection approach. Cancer. 2006 May 1;106(9):2039-50. III. Demonstration Website http://www16.imsweb.com/ Username: imsdev Password: website CSQS Home Page Prostate, Pre-Trt Clinical T3 N0 M0 Gleasons 8-10 73 White Married 73 Chronologic Age, 67 Health Adjusted Age Show Diabetes, Congestive Heart Failure Show Health Adjusted Age at 82, Then Add 3 Years Subjective 85 People Chart for 1, 5, 10 Years People Chart for 1, 5, 10 Years Pie Chart for 1, 5, 10 Years Pie Chart for 1, 5, 10 Years Summary Chart – Alive Summary Chart – Death From Other Causes Summary Chart – Death From Cancer IV. Testing Usefulness in Real World Situations Questions • Should this system be public, or only for use by clinicians? • How can the results of this system be best used to contribute to health care provider-patient communications? • Can this system contribute to tumor board discussions? • For what medical specialties is this system best suited? Oncologist, Surgical Oncologist, Primary Care Physician? • Can modifiable risk factors (such as treatment) be added to the system? Example of Adjuvant! Online Output (http://www.adjuvantonline.com/) No Additional Therapy 32.3 alive in 5 years 55.5 die due to cancer 12.2 die of other causes With Selected Additional Therapy Additional Slides 32.3 alive in 5 years 13.8 alive due to chemotherapy 39.9 die due to cancer 14.0 die of other causes Future Directions • Testing in clinical settings (tumor board and patient perceptions) – Supplemental grant to the Centers for Excellence in Communications (Kaiser HMO setting) • Validation • Potential new cancer sites – Head and neck cancers – Breast cancer • Adding new comorbidity calculators (NHIS –based) • Adding ecologic covariates Collaborators • NCI – Angela Mariotto, Minjung Lee, Kathy Cronin, Laurie Cynkin, Antoinette Percy-Laurry • IMS – Ben Hankey, Steve Scoppa, Dave Campbell, Ginger Carter, Mark Hachey, Joe Zou • Advisory – Dave Penson (Urologist, Vanderbilt) – Deborah Schrag (CRC Oncologist, Dana Farber) – (Consultants - User Interface) • Scott Gilkeson, Bill Killiam One Dataset Dataset 1 Dataset 2 Cancer Patients Non-cancer Cox Cox Model 1 Model 2 Cox Cox Model 1 Model 2 Net probability Net probability Net probability Net probability of dying of of dying of of dying of of dying of Cancer Other Causes Cancer Other Causes Equations are the same Crude probabilities dying of Cancer Crude probabilities dying of Cancer and Other Causes and Other Causes No need for independence assumption Needs independence assumption of Minjung used a continuous time model where competing risk and that populations are estimates are computed using counting process* similar* Estimates and SE’s of cumulative incidence Can take advantage of the richness of are identical if independence is assumed or alternative different data sources. not (Nonidentifiability: Tsiatis,1975) Use discrete time model – CI’s of cumulative *Cheng SC, Fine JP, Wei LJ, “Prediction of the Cumulative incidence computed using delta method Incidence Function under the Proportional Hazards Model”, Biometrics, 54, 1998.
Pages to are hidden for
"PowerPoint Presentation"Please download to view full document