THE ISSUE OF STAGE AT DIAGNOSIS IN SURVIVAL ANALYSIS Pamela Minicozzi Analytical Epidemiology and Health Impact Unit Department of Preventive and Predictive Medicine, Fodazione IRCCS Istituto Nazionale dei Tumori, Milan January 23, 2014 EPAAC WP9 Satellite Meeting, Ispra (Italy)
EUROCARE-5: Between-country variation and time-trends in 5-year age-adjusted relative survival 2 De Angelis et al, LO, 2014
Objectives Overview of Staging systems Determinants of stage Stage comparability Stage incompletness 3
What is stage? 4
Stage describes the size, the extent and the diffusion of cancer. For solid tumours, the stage is based on: Location of the primary tumour Tumour size (T) Lymph node involvement (N) Presence or absence of distant metastasis (M) 1. ptnm, ctnm What is stage? BUT 2. Condensed TNM (each coded as 0, 1, X; ENCR reccomandation; when T or N or M have not been explicitly recorded in the clinical/pathological records) 5 3. Extent of disease (SEER; In situ, localised, regional, metastatic, unknown)
What stage system? 6
In what manner do differences in stage distribution help with the survival interpretation? 7
Stage-standardisation: some cautions EUROCARE-4: Stage distribution and 10-year age-std. relative survival for breast cancer women, 2000-02, 02, by CR Pr roportion 90 80 70 60 50 local 40 regional 30 metastatic 20 unknown 10 age-std 10-y RS 0 8 Allemani C et al, Int J Cancer, 2013
To be considered Stage migration - Strictly related to more sensitive clinical investigations -Tumour of limited metastatic activity may be re-allocated to the metastatic category -Affects stage-specific survival Early detection - Detection of malignant lesions that would have never progressed and presented clinically - Time between diagnosis and death is extended - Too optimistic survival estimates -Affects both overall and stage-specific survival 9
Differences in survival confounded by the stage migration (1) SEER-Medicare: proportions and 2-year survival, NSCLC, 1998-03, by stage. 10 Dinan MA et al, J Clin Oncol 2012
Differences in survival confounded by the stage migration (2) EUROCARE-3/SEER: RER of death, by stage and number of examined lymph nodes, colorectal cases, 1990-91. Multivariate regression. considering the number of examined lymph nodes ( [or lymph node 11 ratio] determinant of stage) partially improves the comparability of cases classified in the same stage Ciccolallo L et al, Gut, 2005
Differences in survival confounded by the stage migration (3) Randomised prospective trial: survival and HR of death, of any event and of distant metastasis at 5-year after diagnosis, breast cancers, enrolled between May, 1999, and February, 2004. 3,887 clinically node negative tumours 11% isolated tumour-cell; 5% micro and macro metastases 95.8% 94.6% 89.2% 92.5% 86.4% 86.4% HR occult vs no occult =1.40 (1.05-1.86) HR occult vs no occult =1.31 (1.07-1.60) HR occult vs no occult =1.30 (1.02-1.66) 12 Weaver DL et al, New Engl J Med, 2011
To be considered Stage migration - Strictly related to more sensitive clinical investigations -Tumour of limited metastatic activity may be re-allocated to the metastatic category -Affects stage-specific survival Early detection - Detection of malignant lesions that would have never progressed and presented clinically - Time between diagnosis and death is extended: the lead-time bias - Too optimistic survival estimates -Affects both overall and stage-specific survival - 13
The reduction in mortality between screen-detected and those who were not screen-detected is the right measure of benefit Breast Cancer Screening Programme UK: age 50 70, screening every 3 years, since 1988. 20% relative risk reduction in breast cancer mortality (screen-d vs. non screen-d) 14 19% frequency of overdiagnosis
Other issue... The unknown stages 15
Dealing with unknown stage 100 90 80 70 60 50 40 30 20 10 0 10-year RS (%) EUROCARE-4: 10-year relative survival for breast cancer women, 2000-02, by CR and stage local regional metastatic unknown Stage at diagnosis - Unknown category is a combination of patients distributed across all stage categories - survival estimates lays in the middle we can ONLY suppose that unstaged cases are not selected Amsterdam Finland Geneva Allemani C et al, Int J Cancer, 2013 -Cases with unknown stage are selected (untreated, advanced cases, old) -Exclusion of unstaged cases overestimates survival in the population as whole - Survival for public health evaluation should include all cancer patients Osaka CR: 5-year survival for stomach cancer, 1975-94 Five-year survival (%) 40 35 30 25 20 15 10 5 0 Staged Unstaged 16 Ito Y et al, Jpn J Clin Oncol 2007
Dealing with unknown stage Multiple imputation - Aims: to identify a replacement for a missing value to use the available information to preserve the relationships in the dataset to obtain a robust estimates of excess hazard ratio of death -Assumptions: Data are missing at random Stage depends on age at diagnosis, censoring, survival time, and the interaction between survival time and censoring - Logistic regression models for each variables with missing values include: vital status age at diagnosis excess hazard or follow-up time covariables predicting stage (i.e. sub-site, year of diagnosis, sex, etc) any interactions on the excess hazard or the follow-up time 17 Meringe et al, Acta Oncol 2013
Dealing with unknown stage Multiple imputation and the outcome No missing (truth) Imputation 1, outcome Imputation 1, no outcome Imputation 2, outcome Imputation 2, no outcome Complete case Outcome Imput missing data Is it a circular analysis? Bias in the estimation of the intercept and other coefficients of variables imputed Moons KGM et al, J Clin Epidemiol 2006 No missing (truth) Imputation 1, outcome Imputation 1, no outcome Imputation 2, outcome Imputation 2, no outcome Complete case 18
Conclusions Stage is a parameter directly involved in cancer prognosis Stage-specific and stage-standardised survival estimates are not sufficient for interpretation of survival differences Staging processes should be carefully examined and comparability of staging be verified before any analysis High resolution studies can help with this issue (detailed clinical information on stage and determinants of stage at diagnosis, not routinely collected by cancer registries, is abstracted by clinical records) Possible biases occurring through stage completeness should be taken into consideration 19
Thank you 20
Summary extent of disease In situ: Abnormal cells are present only in the layer of cells in which they developed Localized: Cancer is limited to the organ in which it began, without evidence of spread Regional: Cancer has spread beyond the primary site to nearby lymph nodes or tissues and organs Distant: Cancer has spread from the primary site to distant tissues or organs or to distant lymph nodes Unknown: There is not enough information to determine the stage 21
MCAR: missing completely at random missing cases are no different than non-missing cases MAR: missing at random Missing data missing cases are different from the complete cases but the probability of missing data of being missing is independent of the unseen values, given other observed variables and/or outcomes MNAR: missing not at random cases are missing in a unmeasured fashion 22