Cancer survival in ﬁ ve continents: a worldwide population-based study (CONCORD)

Background Cancer survival varies widely between countries. The CONCORD study provides survival estimates for 1·9 million adults (aged 15–99 years) diagnosed with a ﬁ rst, primary, invasive cancer of the breast (women), colon, rectum, or prostate during 1990–94 and followed up to 1999, by use of individual tumour records from 101 population-based cancer registries in 31 countries on ﬁ ve continents. This is, to our knowledge, the ﬁ rst worldwide analysis of cancer survival, with standard quality-control procedures and identical analytic methods for all datasets. Methods To compensate for wide international diﬀ erences in general population (background) mortality by age, sex, country, region, calendar period, and (in the USA) ethnic origin, we estimated relative survival, the ratio of survival noted in the patients with cancer, and the survival that would have been expected had they been subject only to the background mortality rates. 2800 life tables were constructed. Survival estimates were also adjusted for diﬀ erences in the age structure of populations of patients with cancer. and substantially lower than in white men and women in all 16 states and six metropolitan areas included. Relative survival for all ethnicities combined was 2–4% lower in states covered by NPCR than in areas covered by the Surveillance Epidemiology and End Results (SEER) Program. Age-standardised relative survival by use of the appropriate race-speciﬁ c and state-speciﬁ c life tables was up to 2% lower for breast cancer and up to 5% lower for prostate cancer than with the census-derived national life tables used by the SEER Program. These diﬀ erences in population coverage and analytical method have both contributed to the survival deﬁ cit noted between Europe and the USA, from which only SEER data have been available until now. relative c precision the square of their standard errors in each the target, as the pooled 5-year relative survival for all participating US populations, age-standardised to the same weights. Raw survival esti mates were not plotted. The 99·8% control limits super imposed on each plot represent about three standard deviations from the pooled US survival at each level of precision.


Introduction
International comparisons of population-based cancer survival have been rare, 1-5 but large and unexplained diff erences in survival have been reported for many cancers from individual studies and cancer registries in Europe and North America. 6 For example, 5-year relative survival for women diagnosed with breast cancer during 1985-89 was 73% in Europe (weighted mean for 17 countries) 7 and 84% in the USA. 8 The CONCORD study provides a systematic comparison of survival between Europe and North America, [9][10][11][12][13][14][15][16] extended to countries in all other continents.
The fi rst international comparison of cancer survival, published in 1964, 17 was a study of patients diagnosed with one of 15 common cancers in Denmark, England, Finland, France, Norway, Sweden, and the USA, mainly during 1945-54. It was the fi rst study in which relative survival techniques, fi rst described in the 1950s, [18][19][20] were used to correct the survival estimates for diff erences in background mortality between participant countries. The fi ndings are mainly of historical interest, but survival in the USA (represented by Connecticut) was generally higher than in the European countries.
Cancer survival is known to vary between the regions of the USA covered by the US National Cancer Institute's (NCI) Surveillance, Epidemiology and End Results (SEER) Program, 21 but the range of survival in Europe is much wider. Furthermore, survival from breast cancer during 1985-94 was higher in each of the nine SEER areas than in any of the 22 countries participating in the European study of cancer survival (EUROCARE). 7,22 The diff erences were often more marked in elderly patients: 9 for several cancers, 5-year survival for patients diagnosed aged 75 years or older during the 1990s was nearly 20% higher in the USA than in Europe. 23 The CONCORD study began in 1999 as an extension of the EUROCARE-3 study, then just starting. EUROCARE has published systematic comparisons of survival for most adult and childhood cancers in Europe since 1995. 24 The fi rst EURO CARE study involved patients diagnosed in 1978-84 in 12 countries; 25 EUROCARE-2 covered patients diagnosed during 1985-89 in 17 countries, 26 and EURO CARE-3 involved 22 countries, with patients diag nosed in 1990-94 and followed up to 1999. 27,28 More recently, EURO CARE-4 has included patients diagnosed in 23 coun tries dur ing all or part of 1995-2002 and followed up to 2003. 29,30 CONCORD was originally designed to assess the survival of adults (aged 15-99 years) diagnosed with cancer of the breast (women), colorectum, or prostate during 1990-94 in Europe and the USA, using population-based data and standardised quality control, and with identical analysis for all datasets, adjusted for diff erences in general population (background) mortality by country, region, race, and calendar period, and also for diff erences in the age structure of patient populations. CONCORD also enables comparison of cancer survival between fi ve states and four metropolitan areas in the USA covered by the SEER Program (SEER-9) and 11 states covered by the Centers for Disease Control and Prevention's (CDC) National Program of Cancer Registries (NPCR). It also provides a wider comparison of cancer survival between black and white patients in the USA than has previously been possible.
CONCORD includes data from one or more countries on all fi ve continents. To our knowledge, it is the fi rst attempt at a global comparison of cancer survival.

Cancer registries
In 1999, we identifi ed at international cancer meetings in Atlanta (USA) and Lisbon (Portugal), and from published studies, population-based cancer registries that had published survival data and were operational during 1990-99. Registries that had met the quality criteria for inclusion in Cancer Incidence in Five Continents (volume VII, 1988-92) 31 were eligible. We obtained data from 19 other registries. Most had met comparable criteria, such as those in the EURO CARE-3 study (patients diagnosed during 1990-94 with follow-up to 1999). 28 North American regist ries were eligible if they had met the standards required for Cancer Incidence in North America, 1991-95, 32 and could provide com plete follow-up to the end of 1999. In total, we identifi ed 112 registries, but 11 were withdrawn or excluded: no response (one); withdrawal for legal reasons (one); incomplete registration before 1995 (four); follow-up activity stopped be fore 1999 (two); data not supplied by the September, 2005 deadline (three).
In Europe, the 53 cancer registries that contributed data to EUROCARE-3 28 on cancers of the breast, colon, rectum, or prostate all participated in the CONCORD study. Six other registries also provided data: two national registries (Northern Ireland and Ireland) and four regional registries from the Netherlands (North) and Switzerland (Graubunden-Glarus, St Gallen-Appenzell, Valais). As in the EURO CARE study, the UK is considered as its four constitu ent countries (England, Scotland, Wales, Northern Ireland), each of which has a national registry. In England, both the national cancer registry and eight of the regional cancer registries submitted datasets.
Oceania was represented by the national cancer registry of Australia, with data from each of the eight populationbased state or territorial registries.

Quality control
Procedures used in the EUROCARE-3 study were applied to all datasets. Tumour records were supplied with the anatomical site coded to the ninth revision of the International Classifi cation of Diseases (ICD-9 39 ) for four index tumours: cancers of the breast (women) (ICD-9 174.0-174.9), colon (153.0-153.9), rectum (including the anus, 154.0-154.9), and prostate (185). Tumour morph ology and behaviour were coded to the fi rst or second revision of ICD-Oncology (ICD-O, 40 ICD-O-2 41 ). Only invasive malignant tumours (behaviour code 3) were included. Patients with an index tumour had sometimes been registered with another malignancy, either before or after the index tumour. Data on those other cancers in index patients were also submitted. Only the fi rst, primary, invasive, malignant tumour diagnosed in each patient was retained for analysis. Patients registered with a malignant neoplasm before the index tumour were excluded, although non-melanoma skin cancer was not counted as a previous tumour for this purpose. Bilateral breast cancers and multiple colon cancers were included as a single tumour if synchronous; otherwise, only the earliest tumour was considered. The duration of survival was taken from the date of diagnosis of the index tumour until death from any cause, or until the patient was censored from the analysis as alive, either at loss to follow-up or after Dec 31, 1999, whichever came fi rst; any subsequent tumour occurring in the same patient during that period was ignored.
Standard quality-control routines, based on those developed by the International Agency for Research on Cancer, 42 were applied to each tumour record. Records with invalid codes, impossible sequences of dates, or improbable combinations of tumour site and morphology were re turned to the registry for checking. Usually, the registry provided a correction or an explanation. Corrected tumour records were checked again: those which still had missing, invalid or inconsistent values for sex, site, morphology, or dates were fl agged as major errors and excluded from analysis. Records for which an unlikely combination of age, site and morphology had nonetheless been confi rmed as correct were fl agged as minor errors, and included in the analyses. Details of the approach have been published else where. 43 Detailed quality-control fi ndings are available online. 34

Follow-up
All registries used more than one mechanism of follow-up to ascertain the vital status (alive, dead, emigrated, lost to follow-up) and the date of the last vital status for each registered patient. The mechanisms varied between countries, usually linkage between the registry's database and a variety of other data sources, especially the national index of deaths. Secure linkage of a tumour record and a record of death, based on a set of identifi ers such as name, sex, date of birth, and personal identity number, enabled the registry to update the tumour record accordingly. Direct contact with the patient or their family to establish vital status was unusual, although home visits by registry staff were done in Algeria. Enquiries to the patient's primary care physician or hospital consultant were frequently used. A wide variety of administrative data bases was also used, such as social insurance, health insurance, motor vehicle records, drivers' licences, hospital discharge records, national primary-care databases, electoral registers (those eligible to vote), and voter registration records (those who voted in the last election). The presence of a person's record in such administrative databases on a given date is taken as evidence that the person was alive on that date. This is subject to administrative error (failure to remove in timely fashion the record of a person known to be dead) and fraud (by someone seeking to retain access to benefi ts received by the deceased), but in most instances the risks are small. If coverage of the databases was known to be high, and especially if a person was present in more than one such database, the risk of error decreased further.
In the USA, a match to an administrative database might show that an event occurred during a certain quarter of a year (eg, an insurance claim paid, a licence renewed), but the exact date might not be known; the date of last vital status was then set to the fi rst day of the quarter, ie, Jan 1, April 1, July 1, Sept 1. This approach can give rise to irregu lar distribu tions of the day of last known vital status, but it is a conservative approach to establishing when patients were last known to be alive, because patients are censored from sur vival analysis on the latest of any such dates in the record.
The proportion of patients not known to be dead and for whom the registry could not be certain that the date of last vital status was at least 5 years after diagnosis was less than 1% overall. The proportion was often zero (follow-up for at least 5 years was established for every patient not known to be dead), the highest proportion was 4%, and only in a

America (Central and South)
Brazilian registries

America (North)
North American registries  very few registries was it greater than 1% (available online 34 ). Such patients are described as censored from the analysis.

Statistical analysis
We estimated relative survival up to 5 years after diagnosis from the individual tumour data, using the Hakulinen approach 44 embedded in the US National Cancer Institute's publicly accessible SEER*Stat software. 45 SEER*Stat is the standard tool used for cancer-survival estimation by the SEER Program cancer registries, and we used it to ensure that survival estimates for US registries would be seen as comparable with those already published by the SEER Program. Survival estimates were also derived by race for the USA (black and white). Relative survival is the ratio of the survival noted in the patients with cancer and the survival that would have been expected had they been subject only to the mortality rates of the general population (background mortality). It is a measure of the excess mortality in patients with cancer over and above the background mortality, and can be interpreted as survival from the cancer after correction for other causes of death. This approach is crucial for inter national comparisons of cancer survival, because the back ground risks of death from all causes in adults often diff er very widely. Background mortality was taken from life tables developed specially for the CONCORD study, specifi c for sex, calendar year, region, and race. 46 The probability of survival in successive years after diagnosis was estimated in survivors to the start of each year. We report the cumulative relative survival at 5 years. Survival was not estimated if fewer than fi ve patients with a given cancer were available for analysis in any category defi ned by age, sex, and race. Relative survival was adjusted RS=relative survival. R=raw (not age-standardised) survival estimate: too few cases in one or more age groups. *International Cancer Survival Standard (see text). †No state-wide data available for this city. ‡Survival truncated if greater than 1·0 (100%). 95% CIs were calculated by use of a logarithmic transformation (see text). for heterogeneity in the withdrawal of patients from followup and consequent changes in the age-sex-race distribution of patients with cancer in successive calendar years, by use of the exact method. 44 Expected survival was derived from complete life tables that contained the probabilities of death or the central death rates for the general population of the registry's territory, by single year of age, sex and (where possible) race, and single calendar year between 1990 and 1999. Many registries provided complete life tables. For some registries, complete life tables were constructed from raw data obtained from published sources on the numbers of deaths by age, sex, and race in the relevant year(s) or period, and the corresponding populations. For the remaining registries, abridged (5-year or 10-year age groups) life tables from published sources were smoothed to produce complete life tables. In some registries, life tables were interpolated, as required, to provide life tables by single calendar year throughout the decade 1990-99. Details are provided in an accompanying paper. 46 Cancer survival is known to vary with race, 47-55 and we assessed racial diff erences in survival where possible. Individual tumour records were coded by race only in the data from the USA (black, white, other). Race-specifi c estimates of relative survival were produced with separate life tables for each race, constructed from the raw data on populations and the number of deaths. 46 In the USA, race-specifi c mortality in the general population also varies between states. 36 We developed separate sets of complete life tables for each state and metropolitan area and for each sex. This approach was designed to enable the closest possible adjustment of relative survival estimates in the USA for geographic variation in background mortality in both blacks and whites, by age, sex, and calendar period. Race-specifi c life tables for both blacks and whites were developed for 11 of the 16 states and all six metropolitan areas. Where racespecifi c life tables were available, they were used in the estimation of relative survival for patients of that race. For other patients, the all-races life table for that population was used. For fi ve less populous states (Hawaii, Idaho, New Mexico, Utah, and Wyoming: 6% of the 109 million population covered by participating registries; webtable), only the life tables for whites were suffi ciently robust, and relative survival estimates for blacks are not separately presented.
Relative survival measures the extent to which patients with cancer have a higher death rate than the general population of the country or region in which they live. 56 Occasionally, despite use of the most appropriate life table, this excess death rate can be negative in a given time interval since diagnosis, implying that the death rate of cancer survivors during that interval is actually lower than that of the general population. This situation can arise from random variation in the death rate when the number of deaths in the interval is small, 57 either because the interval is very short, or because survival is poor and most patients have already died before the start of the interval, or because survival is high and there are very few deaths. In such situations, we present by default the estimate derived by use of the SEER*Stat option to constrain the excess mortality rate to zero, which imposes a plateau in the relative survival curve. The unconstrained estimate was also obtained for comparison.
Even though relative survival is already adjusted for agespecifi c diff erences in background mortality, robust international comparison of relative survival requires agestandard isation, 23 because the age distribution of patients with cancer varies between countries, and because relative survival also varies widely by age, at least in Europe. 27 Conventional age-specifi c weights used to standardise incidence or mortality rates (eg, the national population or the hypothetical world standard population 58 ) are unsuitable because patients with cancer have a very diff erent age profi le from that of the general population.
A cancer-survival comparison of such wide scope has not been done before and the choice of weights for agestandardisation was not straightforward. International stan dard cancer-patient populations have been proposed, with diff erent sets of weights in 5-year or 10-year age bands for each of 20 common cancers, derived from their worldwide distribution. 59 The weights used for the EUROCARE-3 study were derived from the age distribution of all patients included in that study for each cancer, and were thus cancer-specifi c. 43 The disadvantage of these standards is either that a unique set of weights is required for each cancer (cancer-specifi c), or else that the standards are arbitrary (study-specifi c), vitiating comparison between studies.
We chose the recently developed International Cancer Survival Standard (ICSS) weights. 60 These comprise just three sets of age weights, derived from discriminant analysis to fi nd the smallest number of sets of standard age weights that enable adequate standardisation of survival. Each standard is applicable to a range of diff erent cancers, and provides age-standardised survival estimates that are not too diff erent from the unstandardised estimates. The fi rst ICSS standard applies to cancers for which incidence rises rapidly with age, and we used this in all analyses. For cancers of the breast, colon, and rectum, we used fi ve age groups: 15-44, 45-54, 55-64, 65-74, and 75-99 years. For prostate cancer, which occurs mainly in older men, we used four age groups: 15-54, 55-64, 65-74, and 75-99 years. Where data were too sparse for standardisation, the raw (unstandardised) survival estimate is presented, fl agged with "R".
The same age weights were used for men and women, and for each race, enabling direct comparison of agestandardised relative survival between patient groups defi ned by sex and race. Because identical weights were used for breast, colon, and rectal cancer, the agestandardised estimates of survival for these cancers can also be directly compared. This would not be possible if cancer-specifi c weights were used.  For countries represented by more than one regional cancer registry, we provide a survival estimate derived from the pooled data for all contributing registries, agestandardised in the same way. This is an overall estimate of survival in the combined territories providing data from that country, not a weighted mean of the various regional estimates. The combined estimate should not be considered as necessarily representative of survival in the country as a whole, except where the regional registries cover the entire country.  Hawaii (all races), S 90·2 (88·1-92·3) The proportion of survivors is constrained in the range zero to one (or 0 to 100%), but confi dence intervals (CIs) for relative survival derived in the usual way, from the Normal approximation, can produce implausible values (<0 or >1). SEER*Stat provided the standard error of the cumulative relative survival based on the Greenwood formula, 61 but did not provide CIs. We used these standard errors to estimate CIs on the logarithmic scale (webpanel).
For the USA, we constructed funnel plots of relative survival for each cancer and sex, to obtain further insight into the variability of survival by race and state, and to avoid spurious ranking of the survival estimates. 62 The plots show how much a particular survival estimate deviates from the pooled US value, given the precision of each estimate. The precision depends on the number of deaths included in the analysis, which depends in turn on the size of the popu lation and the frequency and lethality of the cancer in that population. 5-year relative survival estimates for each popu lation, age-standardised and adjusted for race-specifi c and state-specifi c background mortality, were plotted against the precision of the estimates, taken as the inverse square of their standard errors (webpanel). The horizontal line in each plot, the target, was estimated as the pooled 5-year relative survival for all participating US populations, agestandardised to the same weights. Raw survival esti mates were not plotted. The 99·8% control limits super imposed on each plot represent about three standard deviations from the pooled US survival at each level of precision.    Diff erences between survival estimates are presented as the absolute value, eg, 15% is given as 5% (not 50%) higher than 10%.
We analysed individual data for almost 2 million adults who were diagnosed with a fi rst, primary, malignant, invasive neoplasm of the breast (women), colon, rectum, or prostate during the period 1990-94 and who had been followed up to ascertain their vital status for at least 5 years after diagnosis until the end of 1999 or later. Data were contributed by 101 population-based cancer regis tries covering a combined population of almost 300 million persons living in 31 countries (table 1 and webfi gure 1). Canada and the USA contributed 1·07 million patients (54% of the total) from a population base of 125 million. The 24 European countries contri buted 740 000 patients (37%) from a population base of 126 million, indicating lower mean incidence of cancer than in North America.
The smallest dataset came from Sétif (Algeria), covering a population of 1·1 million, some 4% of the national population. The registry could only provide data for the period 1992-94, the population is young, and cancer risks are currently low on the global scale. 63 The dataset was therefore small, a total of 300 patients. This decreases the statistical precision of survival estimates, but no patient was detected solely at death certifi cation or autopsy, and the vital status of every patient was ascertained at a home visit by registry staff , something no other registry could deliver. Some of the datasets for black patients in US states were of similar size (webtable). California provided the largest single dataset of 240 000 patients diagnosed during 1990-94 in a population of 31 million (12% of the US population), with a very high cancer risk on the global scale (table 1).
For 16 of the 31 countries, the data covered 100% of the national population (table 1). The proportion of the national population covered by the data for the other 15 countries ranged from less than 10% (Algeria, Brazil, Japan, Austria, Czech Republic, France, Germany, Poland) to 10-29% (Italy, Portugal, Spain, Switzerland) and 30% or more (Canada, USA, the Netherlands).
Most registries provided data on patients diagnosed during the entire 5-year period 1990-94, but ten registries provided data for shorter periods (table 1).
Data for all four index cancers were provided by 89 of the 101 registries. Two specialised registries in Côte-d'Or (France) only collect data on cancers of the breast or colorectum, respectively, whereas ten general registries that collect data for all cancers only contributed data for selected cancers: breast (Isère, France; northern Netherlands; all fi ve Swiss registries); breast, colon, and prostate (Campinas, Brazil; Nova Scotia, Canada), or breast, colon, and rectum (Granada, Spain; table 1).
Ethical approval for the CONCORD study 33 was obtained from the Istituto Superiore di Sanità, Rome, Italy (CE-ISS 02/03, May 20, 2002) and from the Institutional Review Board of the CDC, Atlanta, GA, USA (IRB #3551, July 24, 2002). SEER data were obtained from the public-use dataset. 38 For other registries, anonymised data were trans-mitted to the CONCORD Data Analysis Centre at the Istituto Superiore di Sanità by use of special courier delivery of encrypted and password-protected CD-ROMs with separate email transmission of the password, or preplanned deposition of password-protected fi les on a specially created File Transfer Protocol (FTP) site from which the data were immediately removed in Rome. Each tumour record included a serial number for the purposes of quality control with the originating cancer registry.

Results
The background risk of death in the general population varied widely between the participating countries and regions. The mean life expectancy at birth during the decade 1990-99 ranged from 63·7 to 77·6 years in men and from 70·9 to 83·7 years in women. 46 In the USA, the range of life expectancies in white and black populations did not overlap at all in the states and metropolitan areas for which life tables could be con structed for both groups. The ranges for men were 64·0-70·1 years in blacks and 71·1-75·9 years in whites, whereas the ranges in women were 73·3-76·5 years in blacks and 78·8-80·9 years in whites.
The cumulative risk of death from all causes over the age range 15-59 years in the general population of the participating countries and regions ranged widely, from 9% to 34% in men and from 5% to 17% in women. Over the age range 60-84 years, the cumulative risk of death ranged from 60% to 86% in men and from 40% to 75% in women. 46 Of 785 255 records of breast cancer submitted for analysis, 45 020 (6%) related to women registered with a previous primary cancer, and were excluded (available online 34 ) Of the 740 235 eligible fi rst primary invasive breast cancers, 9215 records were excluded as death-certifi cateonly (DCO) registrations (1%), 239 as autopsy-detected tumours (<1%) and 2064 with major errors (<1%), leaving 728 717 patients for analysis (98% of those eligible), of whom 370 000 (51%) were resident in North America and 304 000 (42%) in Europe (table 1). Almost all (97%) of the tumours included in the analyses were microscopically verifi ed, less than 1% of patients were censored from the analysis within 5 years of diagnosis, and 2·3% died within 1 month of diagnosis.
Relative survival at 5 years, age-standardised to the International Cancer Survival Standard weights, ranged from 80% or over in North America, Sweden, Japan, Finland, and Australia to less than 60% in Brazil and Slovakia, and below 40% in Algeria (table 2 and fi gure 1). Survival in the 24 European countries that contributed to CONCORD was mostly in the range 70-79%.
The survival estimate of 38·8% (95% CI 31·4-46·2) for Sétif (Algeria) is based on 180 patients, and it is not agestandardised because there were too few patients for analysis in some age groups, but age-standardisation for breast cancer in other datasets rarely altered the raw estimate by more than 5% in either direction (data not shown), and survival from breast cancer was undoubtedly much lower in Algeria than in all the other countries.
The pooled estimate of 5-year survival for the two Brazilian registries was 58·4%, but the estimate for Goiânia (65·4%) is more reliable than the very low fi gure for Campinas (36·6%), where high proportions of patients were excluded as DCO or with major errors (available online 34 ). The propor tion of metastatic tumours was higher in Campinas, however. The 5-year survival estimate for Cuba was 84·0%, but this was likely to be an over-estimate: some 28% of records were excluded because they were registered solely from a death certifi cate.
The pooled estimate of 5-year survival for Canada was 82·5%, with a narrow range from 79·3% in Nova Scotia to 85·4% in British Columbia (table 2 and fi gure 1). In the USA, 5-year relative survival for all races combined ranged from 78-81% in New York City, New York State, and Louisiana to 89-90% in Hawaii and Seattle, WA (table 2), but most of the estimates were within a fairly narrow range, from 82% to 87% (fi gure 2). Survival in metropolitan areas covered by SEER registries was similar to that in the respective states: Detroit, MI 83·0% and Michigan State 82·3%; San Francisco, CA 86·2% and California State 84·6%. 5-year survival was 77·4% for residents of New York City, NY (with 40% of the state population), slightly lower than for New York State as a whole, 81·0%.
Survival was lower for blacks than for whites in all 17 populations in the USA for which this could be assessed with race-specifi c life tables (webfi gure 2). The age-adjusted pooled estimate of 5-year survival was 84·7% for whites (range 80-90%) and 70·9% for blacks (table 3). The range in survival was wider for blacks (60-83%), but the values at both extremes of the range were based on relatively few patients and have wider CIs. Within a given US population, the absolute diff erence in age-adjusted relative survival between blacks and whites ranged from 2% (Rhode Island, Nebraska) to 25-27% (Iowa and Seattle, WA; table 3 and webfi gure 2). Even in areas where blacks comprise 25% or more of the population, survival for black women was 8-14% below the lowest estimate for white women (79·6%) in any of the participating areas: Atlanta, GA (71·1%), Detroit, MI (71·7%), New York City, NY (65·8%), and Louisiana (69·9%). The pooled estimate of 5-year survival for the USA was 84·0%, with 86·1% in areas covered by SEER and 83·1% in areas covered by NPCR (table 3).
Survival in black women was always lower than the mean survival for all US populations included, and often more than three standard deviations below it (below the 99·8% control limits), after controlling for the precision of the estimates. Survival in white women is generally within or above the upper control limits, especially in territories covered by the SEER Program. A notable exception is for white women in New York State, including New York City, where the survival estimates are precise, but well below the lower control limits (webfi gure 3).
5-year relative survival for breast cancer in Europe, agestandardised to the ICSS weights, ranged from 57·9% in Slovakia to 82·0% in Sweden (table 2 and fi gure 1), whereas the pooled estimate derived from the data of 58 registries in the 24 participating European countries was 73·1%. Survival estimates for most of these countries have been reported. 27 The CONCORD study includes additional data from four countries: 5-year survival was 69·6% in Ireland and 72·0% in Northern Ireland, similar to the UK mean value of 69·7% (table 2). In Switzerland, 5-year survival in the cantons of St Gallen-Appenzell, Graubunden-Glarus, and Valais was 72-75%, about 4-7% lower than in Geneva or Basel. 5-year survival was 77·8% in northern Netherlands, similar to that in Amsterdam and southern Netherlands (76-78%).
The national estimate of 5-year survival for breast cancer in Australia was 80·7%. Survival was virtually identical in the six largest states (96% of the national population), in the range 80-82%, but notably lower in the two smallest regions: 71·9% in Northern Territory (1·0% of the population) and 77·1% in Tasmania (2·6%).
Relative survival at 5 years, age-standardised to the ICSS weights, ranged from about 60% in North America, Japan, Australia, and some western European countries down to See Online for webfi gures 2 and 3 40% or less in Algeria, Brazil, Czech Republic, Estonia, Poland, Slovenia, and Wales (table 2 and fi gure 3).
The estimate of 5-year relative survival for Goiânia (48·1% in men, 44·8% in women) was more plausible for Brazil than the low estimates for Campinas, where 26% of patients had to be excluded with errors. 34 5-year survival in Cuba was about 60% in both sexes, although more than half the patients were excluded from analysis as DCOs. 34 In Canada, the pooled estimate of 5-year survival was 56·1% for men and 58·7% for women. Variation between provinces was small, from 54-57% in men and 58-60% in women (table 2 and fi gure 3). In the USA, 5-year relative survival for all races combined was 60·1% in both sexes, with a range from 53·6% for women in New York City to 67·9% for men in Hawaii (table 2 and fi gure  2). Again, most of the estimates were within a narrow range, 59-64%.
5-year survival for colon cancer among blacks in the USA was lower than among whites. In 34 paired observations of this diff erence in survival (men and women in 17 populations), only three exceptions were noted, in men and women in Iowa and men in Nebraska. The estimates for blacks in those three areas were based on fewer than 50 patients, have wide confi dence intervals and were not age-standardised (table 3 and webfi gure 4). The pooled estimate of age-adjusted 5-year relative survival for the USA was 61% for white men and women, and 51-52% for black men and women. Within a given population, the absolute diff erence between blacks and whites ranged from 2·6% in men and 7·3% in women in Los Angeles, CA to 14·3-17·2% in Colorado. The geographical range in black-white diff erences in survival is aff ected by small populations to some extent, but even in areas where blacks comprise 25% or more of the population (Atlanta, GA, Detroit, MI, New York City, NY, Louisiana), 5-year survival from colon cancer in blacks was 6-12% lower than for whites in the same population (table 3). The pooled estimate of 5-year survival in areas covered by NPCR was 59·8% for men and 59·6% for women, and 61·9% for men and 62·1% for women in SEER areas.
Age-standardised survival in whites ranged from 54·9% to 67·9% (table 3 and webfi gure 4). The range of agestandardised survival for blacks was 45·0% to 59·9%. Survival in blacks was generally lower than the mean value for all included US populations and often more than three standard deviations below the mean, after controlling for precision of the survival estimates. Survival in whites was generally within the control limits. The main exception was for white women in New York State, including New York City, NY, where survival estimates were precise, but more than three standard deviations below the lower control limits around the pooled US estimate (webfi gure 3).
In Japan, the pooled survival estimate was 63·0% in men and 57·1% in women, although survival was about 10% lower in Osaka prefecture than in Fukui or Yamagata (table 2 and fi gure 3).
In Europe, 5-year relative survival for colon cancer in men ranged from 28·5% in Poland to 54-57% in Spain, Finland, Austria, and France. In women, the lowest estimate was also for Poland (30·9%), while survival was in the range 55-60% in nine countries (table 2 and fi gure 3). The pooled estimates for the 51 contributing registries in 23 European countries were 46·8% in men and 48·4% in women. Data on colon cancer were not available for the fi ve Swiss registries, Isère (France), or northern Netherlands. Survival estimates for most of these countries have been published elsewhere. 27 This study included additional data from two countries. 5-year survival in Ireland was 49·1% in men and 48·5% in women. The estimates for Northern Ireland were 47·3% in men and 49·0% in women, slightly higher than the pooled estimate for the UK, 43·5% in men and 44·4% in women (table 2).
The national estimate of 5-year survival for colon cancer in Australia was 57·8% in men and 57·7% in women. Survival ranged from 50-62% in the eight states and territories: it was highest in New South Wales, the Australian Capital Territory, and Queensland, and lowest in Tasmania, Northern Territory, and Western Australia (9·6% of the population; table 1).
5-year relative survival from rectal cancer, agestandardised to the ICSS weights, ranged from about 60% to around 20% in both sexes, with Japan, Canada, the USA, France, the Netherlands, Sweden, and Australia at the upper end of the range, and Algeria, Estonia, Poland, and Slovakia at the lower end (table 2 and fi gure 3).
5-year relative survival in Goiânia, Brazil, was 49·3% in men and 38·4% in women. No data were available for Campinas. 5-year survival in Cuba was 59·2% in men and 62·8% in women, based on analysis of about 700 patients in each sex, although 782 (36%) patients had been excluded as DCO (available online 34 ). In Canada, the pooled estimate of 5-year survival was 58·7% for women and 53·1% for men, slightly lower in the global range than for cancers of the breast, colon, or prostate in Canada. Survival in men ranged from 51·1% (Ontario) to 64·6% (British Columbia), and from 57·8% to 62·8% in women. In the USA, 5-year relative survival for all races combined was 56·9% in men and 59·8% in women, with a range from 46-67% in men and 52-66% in women (table 2 and fi gure 2). Again, most of the estimates were within a narrow range, from 55-60% in men and 57-62% in women. By contrast with colon cancer, survival from rectal cancer was slightly higher in women than in men.
5-year survival for rectal cancer in the USA was generally lower for blacks than for whites, in both sexes (webfi gure 4). The overall estimate of 5-year survival in men was 47·4% for blacks and 57·3% for whites; for women, the estimates were 49·4% for blacks and 60·4% for whites (table 3). When survival for blacks was above 60%, or higher than for whites in the same population (Colorado, Connecticut, Nebraska, Rhode Island), the estimates for blacks were based on around 50 or fewer patients, with wide CIs, and were usually not age-standardised (table 3 and webfi gure 4). Even where blacks comprised 25% or more of the population (Atlanta, GA, Detroit, MI, New York City, NY, Louisiana), 5-year survival was 6-16% lower than for whites in the same area. The pooled estimate of 5-year survival in areas covered by NPCR was 56·3% for men and 58·8% for women, some 2-3% lower than for SEER areas (58·5% men, 61·8% women; table 3 and fi gure 2).
5-year survival ranged from 46·2% to 67·9% in whites; in blacks, the range of age-standardised survival was from 42·8% to 63·5% (table 3 and webfi gure 4). Survival in blacks was generally lower than the mean value for all included US populations, although more often within the control limits. Survival for whites was also generally within the control limits, with the exception of New York City, NY, where survival was below the control limits (webfi gure 3).
In Europe, the geographical pattern for age-standardised 5-year survival was similar to that for colon cancer. For men, the range was from 28-30% in Poland, West Bohemia (Czech Republic), and Slovakia to 53-55% in France, Sweden, and the Netherlands; whereas for women, the range was from 30-32% in Poland, Estonia, and Slovakia up to 63·9% in France, where three of the four contributing registries ranked the highest in Europe (table 2 and fi gure 3). The pooled estimates for the 51 contributing registries in 23 Euro pean countries were 43·2% in men and 47·4% in women. Data on rectal cancer were not available for Isère (France), northern Netherlands, or the fi ve Swiss regis tries. 5-year survival in Ireland was 41·1% in men and 52·5% in women. The estimate for women is not age-standardised, but it is based on over 200 patients (table 1), and similarity between the raw and standardised estimates for cancers of the colon and colon and rectum combined (less than 1%, data not shown) suggests that an age-standardised estimate for rectal cancer would not have been very diff erent. In Northern Ireland, the estimates were 48·2% in men and 43·8% in women (pooled UK estimates were 40·6% in men and 45·3% in women; table 2).
The national estimate of age-standardised 5-year survival for rectal cancer in Australia was 54·8% in men and 59·2% in women. Survival ranged from 45-57% in men and from 55-61% in women.
5-year relative survival, age-standardised to the ICSS weights, ranged from 80% or higher in the USA (92%), Canada and Austria to less than 40% in Denmark, Poland, and Algeria (table 2 and fi gure 1).
In Brazil, 5-year survival was 34·4% in Campinas and 55·7% in Goiânia. Some 30% of tumour records in Campinas were excluded with major errors. Notably, 20 (13·4%) men in Campinas and 71 (21·8%) men in Goiânia died within 1 month of diagnosis, which are the highest proportions of any of the participating registries (available online 34 ). 5-year survival in Cuba was 69·7% (table 3). This estimate was based on 4300 patients, but 54% of the original data set of 9500 patients had been excluded as DCO (table 1 and data available online 34 ).
The pooled estimate of 5-year survival for prostate cancer in Canada was 85·1%, ranging from 77·5% in Saskatchewan to 89·3% in British Columbia (table 2 and fi gure 1). In the USA, 5-year relative survival from prostate cancer was 91·9% for all races combined, with a range from 81·6% in New York City, NY up to 95·0% in Seattle, WA (table 2 and fi gure 2), but most of the estimates were within a fairly narrow range, from 88·6% (Louisiana) to 94·0% (Atlanta, GA). The relative survival estimate for the state of Michigan was 100%, although in the city of Detroit, MI, with 42% of the state population (webtable), survival from prostate cancer in the same period was 93·8%.
Age-standardised 5-year relative survival for prostate cancer in blacks was lower than for whites in all populations for which this could be separately assessed with race-specifi c life tables (webfi gure 2). The overall estimate of 5-year survival was 85·8% for blacks and 92·4% for whites, with an overall diff erence of 6·6% (table 3). The diff erence in survival between blacks and whites ranged from 5·0% (Florida) to 14-16% (Nebraska, Rhode Island), and although the largest diff erences arise where the black populations are smallest, each survival estimate was based on at least 70 patients (webtable). Survival in whites ranged from 83·3% (New York City, NY) to 96·1% (Atlanta, GA), and in blacks from 74·0% (New York City, NY) to 89·6% (Seattle, WA). The pooled estimate of 5-year survival was 89·5% in areas covered by NPCR, and 93·1% in SEER areas (table 3).
Survival in blacks was usually lower than the pooled US estimate, and often more than three standard deviations below it, after controlling for precision (webfi gure 3). Survival in whites was generally within the control limits. 5-year survival for whites was above the upper 99·8% control limit in three SEER populations (Atlanta, GA, Seattle, WA, and Detroit, MI). Survival in whites was below the lower control limit in four NPCR populations, but for California and Florida the diff erence was small (2-3%). In New York State, including New York City, NY, survival estimates are precise, but 6-9% below the pooled US estimate of 92·3% and well below the lower control limit (webfi gure 3).
The 5-year relative survival estimates for Michigan State (100% in both blacks and whites) were too high, and they are not shown in webfi gure 3, although the data were included in the pooled estimate. Information about the death had not been linked to the tumour record for some of the apparent 5-year survivors from prostate cancer in Michigan State, leading to an infl ated estimate. This error did not aff ect the survival estimates for prostate cancer in Detroit, MI or those for other cancers in Michigan State.
The pooled estimate of 5-year survival in Japan was 50·4%, much lower on the global scale than for cancers of the breast, colon, or rectum in Japan. Survival estimates were similar in all three prefectures (table 2 and fi gure 1).
The range of 5-year survival in Europe was especially wide for prostate cancer, from less than 40% in Poland and Denmark to more than 80% in Austria (table 2 and fi gure 1). The pooled estimate for the 49 contributing registries in 23 European countries was 57·1%. Data were not available for nine registries: Switzerland (fi ve registries), Isère and Côte d'Or (France), Granada (Spain), and northern Netherlands. 5-year survival in the Ireland was 62·8%. In Northern Ireland, the estimate was 54·0%, slightly higher than the pooled estimate of 51·1% for the UK (table 2).
The national estimate of age-standardised 5-year survival for prostate cancer in Australia was 77·4%. Survival was closely similar in the six largest states, in the range 76-80%, but notably lower in the two smallest regions: 63·7% in Northern Territory (based on 78 patients, estimate not age-standardised) and 70·2% in Tasmania (1321 patients).

Discussion
To our knowledge, the CONCORD study is the fi rst attempt to provide directly comparable data on cancer survival from many countries around the world by use of central qualitycontrol procedures, standard analytic methods, and a single, centralised analysis of individual tumour records from population-based cancer registries. The fi ndings should eventually complement the inter national data series on cancer incidence 63,64 and mortality 65-67 that have been available for several decades. Cancer-mortality statistics have often been used for international comparisons of progress against cancer, [68][69][70][71][72] but they are also aff ected by wellknown problems of comparability, both between countries and between successive revisions of the ICD. [72][73][74][75] The fi ndings presented here should help joint consideration of trends in incidence, survival, and mortality as indicators of cancer control. None of these indicators is perfect, but none is adequate on its own. [76][77][78] Around 2800 life tables were created to enable the estimation of relative survival by age, sex, country, and race. 46 The life tables are available on the CONCORD website. 34 5-year relative survival for breast, colorectal, and prostate cancers was generally higher in North America, Australia, Japan, and northern, western, and southern Europe, and lower in Algeria, Brazil, and eastern Europe.
Exclusions for a previous cancer (5-9%) were not unexpected. Population-based cancer survival analyses are usually restricted to patients with a fi rst, primary invasive cancer, therefore, to the extent that patients with a previous cancer have been completely excluded in this study, this improves the comparability of the fi ndings with other studies. Participating registries began operation between 1950 and 1990. The data from newer registries are more likely to include unrecognised second and subsequent cancers, because any previous cancer(s) in a given patient might have been diagnosed before the registry began operation.
The main indices of data quality for cancer survival are the proportions of registered patients known to the registry by DCO, or lost to follow-up, and histologically verifi ed. Data quality varied between registries (available online 34 ), but was high overall: very few records were excluded with a major error. Exclusions for major errors were high in Campinas, Brazil (26-47%). The overall proportion of patients who died within 1 month of diagnosis was low for breast cancer (2·3%) and prostate cancer (3·2%), but higher for colon (10·9%) and rectal cancers (7·8%). These values varied between registries, but the overall pattern is plausible; up to a third of colorectal cancers present as an emergency with bowel obstruction. Fewer than 1% of patients were censored from the analysis within 5 years of diagnosis.
Three registries were excluded after quality control, because of high losses to follow-up or ineffi cient regional or national linkage of information on the deaths of patients with cancer. The data for three other registries, Cuba, Campinas (Brazil) and, for prostate cancer, Michigan State (USA), are less reliable than those from other registries, although for diff erent reasons, discussed below. As with the fi rst global compilation of cancer incidence data, in the 1960s, retention of the two registries from Central and South America was partly prompted by the paucity of information on cancer survival from that region: "in this situation, even incomplete data have value". 64 Overall, the exclusion of DCO registrations accounted for only 1% (9215) of the eligible records for breast cancer, 3% (13 102) for colon cancer, 1% (3213) for rectal cancer, and 2% (11 163) for prostate cancer (available online 34 ). The percentage was less than 1% in Algeria, USA, Canada, and Australia, and in the range 0-5% in most European countries and in Brazil, but higher in Osaka (Japan; 5-22%), 79 south Thames (UK); 10-16%), and Cuba (28-60%). 80 The proportion of DCOs is not particularly useful as a comparative index of data quality, 81 but a high proportion of DCO records does suggest that routine data-collection systems might not be complete. The relevance of this index also varies with the system of data collection. Sweden does not use DCOs because the registration of patients with cancer at the time of diagnosis is close to 100%; by contrast, hospitals in Cuba are not allowed to retain the clinical records of deceased persons for more than 5 years. 80 The diff erent proportions of DCO records are unlikely to explain the diff erences in survival between Europe and North America, however. The survival of patients whose tumour is registered as a DCO is generally lower than the mean for all registered cancer patients, 82 so if they could have been included, the transatlantic diff erences in survival would have been slightly greater. Furthermore, most DCO records in the European data are for patients aged 75 years or over, 43 where the survival diff erences are in any case more marked. 9 By contrast, if a high proportion of DCO records is taken to suggest under-registration of long-term survivors, this might produce lower survival esti mates. Adjustment for both DCOs and incompleteness of registration in Thames (UK) and Finland had surprisingly little eff ect on survival, however, even when 10-20% of registrations were DCOs, because the two corrections tended to off -set one another. 83 Under-reporting of incident tumours by up to 5% has been shown not to aff ect international comparisons of survival greatly. 84 A plateau was imposed on the relative survival curve at some point during the fi rst 5 years after diagnosis in about 7% of the 6500 age-specifi c survival estimates by registry, cancer, sex, and race (data not shown). The eff ect on the age-standardised survival estimates at national level was almost always less than 1%.
Diagnostic variability between pathologists might contribute to international diff erences in cancer survival. Thus, survival from colorectal cancer in Japan is among the highest reported here, especially for men. In western countries, invasive colorectal carcinoma is diagnosed when neoplastic tissue invades beyond the submucosa of the bowel. Severe cytological or architectural changes confi ned to the mucosa (in situ or intramucosal carcinoma) have no metastatic potential, and are often labelled high-grade dysplastic adenoma. Japanese pathologists rely more on cytological features, however, and do not consider evidence of invasion into the submucosal layer as a mandatory requirement for the diagnosis of colorectal carcinoma. 85 Pathological practice on this issue might vary substantially between western pathologists. Islands of dysplastic tissue might also be displaced or herniated beyond the muscularis mucosae without implying invasive potential (pseudo-invasion), and diff erential diagnosis can be very diffi cult. 86,87 Assessment of the extent to which international survival diff erences might be attributable to diff erences in the pathological defi nition of disease would need blinded review of pathological diagnoses of a sample of patients by an international panel of expert pathologists. Such reviews are invaluable, but rare. 88 Survival in Sétif (Algeria) was the lowest of all the populations in the CONCORD study for each cancer. Even though the dataset was small, and covers only 4% of the national population, there is little doubt that survival in Algeria is very low. The age distribution of patients was younger than in most populations (available online 34 ) and it cannot explain the low overall survival. Survival in Sétif was similar to or even lower than survival in blacks diag nosed during 1993-97 in Harare, the Zimbabwean capital, where the very low survival was attributed to inade quate access to facilities for early diagnosis, clinical investigation, and treatment. 89 Survival in the two Brazilian registries was generally low, although rectal-cancer survival in Goiânia was close to the European mean. Data quality issues prevented the inclusion of data from three of the 20 population-based registries in state capitals: these registries should be used to provide a broader picture of cancer survival in Brazil. Relative survival reported here for patients with cancer diagnosed in Cuba during 1990-94 was about 20% higher than estimates for those diagnosed during 1988-89, just a few years earlier. 80 Cancer survival for children diagnosed in Cuba during 1988-89 was lower than in more developed countries. 90 The high proportion of DCOs in the 1988-89 data was considered less likely to be biased with respect to survival than in other registries, because of the way data were collected, 80 but the survival estimates for Cuba reported here are still likely to be considerably infl ated, and should be interpreted accordingly.
National estimates of survival for patients with cancer diagnosed in Japan during 1993-96 were slightly higher than the estimates for 1990-94 reported here. 79 They were based on data from seven prefectures, including the three reported here (Yamagata, Fukui, and Osaka). As in the CONCORD data, survival in Osaka was generally lower than the mean survival for Japan.
Variation in survival between the provinces of Canada and the states and territories of Australia was generally small, and overall survival was high: this suggests health care of a high standard in most areas. Variation between the countries of Europe was much wider.
The substantial diff erences in survival between Australia and the UK have been noted before. 91 They are unlikely to be because of diff erences in data quality. For breast cancer, survival from both localised and regional disease was higher in Australia, but survival from metastatic disease was similar. Elderly women in England had especially poor survival. More eff ective treatment in Australia is a plausible explanation. 92 Comparisons of cancer survival between Europe and the USA since 2000 have identifi ed wide diff erences, with survival usually higher in the USA. 9 Closer assessment of these diff erences with more detailed data, not routinely collected by all registries, has enabled the explanatory eff ect of clinical variables such as stage at diagnosis, investigative approach, anatomical site, and morphology to be quantifi ed for colorectal cancer 12,93 and breast cancer, 10 and for a range of cancers in children. 11 In those studies, the USA has always been represented by data from the SEER Program registries, representing some 10% of the US population at that time, because no other data have been available. The availability of data from a large number of state-wide population-based cancer registries that began oper ation around 1990, and meet data quality standards comparable with those of the SEER registries, now enables a larger proportion of the US population to be included in national and international comparisons of cancer survival. The CONCORD study provided the fi rst opportunity for the cancer registries of 11 US states in the NPCR to follow up all their patients for vital status and to undertake analyses of cancer survival, and 42% of the US population is included in these analyses.
The survival diff erences between US and European patients with cancer, especially in the oldest patients, seem unlikely to be attributable to artefacts of cancer registration. The CONCORD study has nonetheless identifi ed two methodological issues that probably do explain some of the well-known diff erences in survival between Europe and the USA, from which only SEER data have been available until now.
First, relative survival was about 2-4% higher in SEER-9 areas than in participating NPCR areas of the USA. Consequently, cancer survival in the 42% of the US population covered by the CONCORD study was 1-3% lower than survival in the SEER areas alone (10% of the US population). Direct estimation of cancer survival for other areas of the USA would be desirable.
Second, census-derived US national life tables give higher estimates of all-cause mortality than are noted in the SEER areas, especially with the gradual decline of mortality in the decade after a census. 46 Use of censusderived national life tables to estimate relative survival (the SEER approach) therefore produces estimates that are almost always higher than those obtained with statespecifi c life tables for each calendar year in the decade (CONCORD approach), which we believe to be more appropriate because it provides tighter control for changes over time in background mortality. With the CONCORD approach, age-standardised 5-year survival in the 22 participating areas of the USA was up to 3% lower than with the SEER approach for breast cancer in women, up to 4% lower for colorectal cancer, and up to 5% lower for prostate cancer (available online 34 ).
The diff erences in cancer survival between blacks and whites of both sexes in the USA are large, and remarkably consistent in 16 states and six metropolitan areas-more populations than it has been possible to study in the past. 94 The diff erences were adjusted for age and for diff erences in background mortality between blacks and whites within each state or metropolitan area. It would be interesting to know if the diff erences were attributable to artefact, or diff erences between blacks and whites in tumour biology, in stage at diagnosis, in access to health care, or in compliance with treatment. The survival diff erences seen in this study are consistent with those in other studies. [47][48][49][50][51][52][53] Data-collection systems were identical for all races. The black-white diff erences in relative survival that we report would have been even larger if we had used race-specifi c national life tables instead of race-state life tables, because back ground mortality is higher (and expected survival lower) in blacks than in whites in all the populations studied. 23,95 Survival from cancers of the breast, colorectum, and prostate varied with the type of health insurance in a population-based study: 96 survival was highest in patients who had private insurance, intermediate with federal insurance, and lowest with no insurance. Another study 97 suggested that prostate cancer is not more biologically aggressive in blacks than whites. Late stage, 98 less treatment, and higher mortality seem to be associated with black race, low socioeconomic status, and poor survival in the USA. [99][100][101] Extensive reviews have led to the conclusion that racial disparities in cancer treatment, which are not explained by clinical factors, lead to worse outcomes in blacks. 102,103 Analysis of SEER data suggested that some racial diff erences in treatment and cause-specifi c survival persist after adjustment for poverty. 104 By contrast, the racial diff erence in survival from colorectal cancer was almost absent in patients managed under the equal-access, integrated health-care Veterans' Aff airs system. 105 Finally, overviews of race, socioeconomic status, and cancer outcomes strongly suggest that equal treatment yields equal outcome, irrespective of race. 53,106 The data presented here extend the evidence that cancer survival in the USA is lower in blacks than in whites.
Simple ranking of countries by overall survival can be misleading. Survival is very similar in many European countries, at the centre of the global range, and a small shift in the survival estimate in either direction can entail a large change in the rank. Thus, even the national survival estimates for Iceland and Malta have wide confi dence intervals and unstable rankings because they are based on populations of around 250 000 (fi gures 1 and 3). The detailed data by country and region are tabulated by continent, country, and region, rather than ranked: some estimates, based on sparse data, could not be agestandardised (table 2).
The numbers of patients included in the analysis varied widely, as did the proportion of the national population covered by the data. These proportions aff ect the extent to which the survival estimates can be deemed representative of the country concerned. For example, in Algeria, Brazil, and Germany, only 1-4% of the national populations were covered by the data. Population coverage of participating registries in Italy was about 15%, but they were concentrated in the wealthier north of the country. 22,30 The same point also applies to the USA, however, because the data presented here confi rm suggestions 107 that cancer survival in the SEER Program areas (10% population coverage during the 1990s) was higher than in other parts of the country. By contrast, regional variation in survival in Australia and Canada was much less marked than in the USA. Similarly, survival for 1990-94 in the four French départements reported here (6% national coverage) was high on the European scale for most cancers, 27 and a much larger study for the same period in 14 départements (20% coverage) showed similar patterns of survival. 108,109 For countries with more than one contributing registry but less than 100% population coverage, we have presented estimates of survival derived from the pooled data, not weighted means of the regional estimates of survival. The question of whether pooled survival estimates derived from regional registries with less than 100% national coverage can properly be considered representative of cancer survival in the whole country has been discussed else where. 22 If population-based estimates of cancer survival are deemed reliable, however, they do suggest a poten tially achievable level of survival, irrespective of whether the estimate is for a whole country or only one region in that country. Regional variation in survival within a country tends to prompt eff orts to improve survival in regions where it is low. The same argument should apply on an inter national scale. This has already happened in Europe. 110 No overall worldwide estimate for cancer survival has been presented. The proportionate contributions from each continent to the CONCORD study are very diff erent from the worldwide distribution of cancers of the breast, colon, rectum, and prostate. The national data for Australia alone represented 63% of the population of Oceania in 1995, 111 and the survival estimates for North America included 44% of the combined populations of USA and Canada around 1992, but for Africa, Asia, and South America, the population coverage of these data was much lower. The survival data for Europe were based on 25% of the continental population of 512 million in 1992, 112 but the EUROCARE study (ongoing since 1989) is the largest and most widely cited international study of cancer survival, and all 57 cancer registries in that study, and six others, contributed to CONCORD. To provide an international summary measure of cancer survival for visual comparison, we therefore used the overall estimates for 23 countries in EUROCARE-3, but age-standardised to the ICSS weights used in CONCORD, instead of the weights used in EUROCARE-3. 113 We have presented pooled estimates of survival for Europe and North America (table 3), but not for other continents.
The size of the population covered by the data aff ects the statistical precision of the survival estimates. This is shown by 95% CIs, but ranked graphics do not provide visual appreciation of the extent to which the survival estimate for a given country or region falls outside the distribution of survival estimates that might be expected, under the hypothesis that survival should be the same in all areas. In that situation, regional variation in relative survival should arise only from random variation around some underlying average. We used funnel plots 62 to provide that visual eff ect for geographical and racial variation in survival in the USA, with the target value as the pooled estimate for the USA, age-standardised to the ICSS weights. 60 These plots display striking geographical and racial variation in survival.
Clinical practice has continued to evolve in the 15 years or so since the patients included in this study were diagnosed. Changes in diagnosis, screening, and treatment have undoubtedly improved the prognosis for cancer patients, at least in wealthier countries.
Survival has increased substantially for cancers of the breast (women), colon, rectum, and prostate in the 17 areas of the USA covered by the SEER Program during 1996-2003, 114 and in Canada (1996-98) 115 and Australia (1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004). 116,117 Smaller increases have been reported in 11 of the 47 Japanese prefectures. 118 These estimates of relative survival, published for national purposes, cannot be compared with the data reported here, however, because of diff erences in the quality control of incidence data or completeness of follow-up, and in methods of analysis. Some estimates were not age-standardised for international comparison, whereas others were standard ised to countryspecifi c age weights, rather than the ICSS age weights that we used.
The only recent international study of cancer survival is EUROCARE-4, which included patients diagnosed in 23 countries during all or part of 1995-2002 and followed up to 2003. 29,30 Survival increased substantially in eastern European countries, where it was much lower than in other parts of Europe during 1990-94. This narrowing of the east-west gap suggests substantial improvements in cancer care. The rise in breast cancer survival in several countries was associated with a fall in mortality, possibly because of improved care and screening programmes; the rise in prostate-cancer survival (and incidence) might be a result of more widespread PSA testing. In western Europe, survival in the UK and Denmark was still low for several cancers in the late 1990s.
CONCORD is, by chance, reasonably well-timed to provide a baseline for international comparisons of cancer survival to assess the eff ect of several major public health initiatives for the control of breast, colorectal, and prostate cancers. In 1990, mass population screening for breast cancer with mammography was beginning in many (but not all) participant countries. At that time, intense early diagnostic activity with prostate-specifi c antigen (PSA) had recently become widespread in the USA, but was little used elsewhere. In Denmark, for example, the 50year increase in prostate cancer incidence is considered real, 119 but the Danish Society for Urology assessed the evidence in 1990 120 and decided not to advocate PSA testing in asymptomatic men, because the therapeutic benefi t was very small; 121 only sympto matic patients were off ered treatment. Mass screening for colorectal cancer with faecal occult blood (FOB) testing or endoscopy had not begun during 1990-94 in any contributing country as far as we are aware. Opportunistic screening with the FOB test began in Japan in 1992; by 2004, about 20% of people aged 40 years or over had been tested within the previous year. 122 Opportunistic endoscopy had already become widespread in some parts of the US population by 1990.
The CONCORD study was planned in three phases. The study reported here (phase I, low resolution) was designed to quantify international diff erences in population-based relative survival by age, sex, country, or region for patients diagnosed during 1990-94 with a cancer of the breast, colon, rectum, or prostate. Phase II (high resolution) was designed to help interpret those international diff erences in survival, by use of a subset of registries that could reabstract detailed clinical data, including stage at diagnosis and treatment, from the original medical records for large random samples of patients diagnosed with one of the same cancers during 1996-98. Findings will be reported in due course. Phase III was designed as a blinded, expert review of the pathological diagnosis for a subset of patients from the phase II study, to assess the extent to which international survival diff erences might be attributable to diff erences in the pathological defi nition of disease in participating countries.
The range of survival estimates for each cancer is very wide. Population-based cancer registries are increasingly important in the comparative assessment of cancer outcomes, 123 and even allowing for diff erences in data quality or statistical robustness, there is little doubt that the chances of survival after a cancer diagnosis vary hugely on a global scale.
The comparability of cancer survival estimates between countries is criticised far more often than the comparability of cancer incidence data from the same registries. There is no statistical basis for this distinction. National sensitivities about cancer survival seem to be much greater than sensitivities about cancer incidence. Cancer survival is a measure of the overall eff ectiveness of cancer treatment services, whereas cancer incidence is a measure of the long-term eff ect of prevention policies, which are less visible on a day-to-day basis and can, incorrectly, be seen as less important.
Cancer survival is a valuable indicator for international comparison of progress in cancer control, 76,124,125 despite the fact that part of the variation in cancer survival identifi ed in this study could be attributable to diff erences in the intensity of diagnostic activity (case-fi nding) in participating populations. Notably, the very same point applies to international comparisons of cancer incidence. If overdiagnosis-which depends on diagnostic intensity-is more marked in one country than another, then it will certainly be harder for researchers to compare incidence, mortality, and survival in those countries. But over-diagnosis has diff erent connotations for health-care systems and patients. In each country, the health-care system will have to be funded, staff ed, and equipped to cope with the diagnostic and therapeutic burden of all patients with cancer, however they are diagnosed. The health-care system must make provision accordingly, and monitor the outcome of that provision; cancer survival is one such overall indicator.
Furthermore, a patient with cancer is still a patient with cancer, whether or not they represent over-diagnosis. If it were possible to distinguish the one from the other reliably, it would be done routinely. As it is, a cancer diagnosis represents the best that medicine has to off er in a given country at a given time, and that best is variable. PSA testing for prostate cancer is an example. No matter how a patient with cancer is diagnosed, they have to cope with the consequences, both psychological and physical, and will usually want to be treated. Such patients cannot be excluded from either incidence or survival analyses. We do not know who they are. In this sense, cancer incidence and survival estimates describe as accurately as possible the occurrence and the outcome, respectively, of cancer as it is diagnosed and treated at a given time in a given population.
Most of the wide global range in survival is probably attributable to diff erences in access to diagnostic and treatment services. 3,82,89,91,126 International variation in survival in Europe has been associated with national levels of economic development, as measured by total national expenditure on health. 29 Survival is positively associated with gross domestic product and the amount of investment in health technology such as CT scanners. 124 Part of the international variation in survival is thus probably attributable to under-investment in health resources. 127,128 The variation in survival might be considered intuitively obvious, given wide global variation in expenditure on health care, whether that is expressed in absolute terms or as a proportion of national resources. A parallel could be drawn with diff erences in survival between rich and poor patients with cancer in a given country, which have frequently been reported. 129,130 Until now, however, direct international comparisons of cancer survival between high-income and low-income countries have not generally been available. The information provided here might therefore be a useful stimulus for change.