Trends in survival and excess risk of death after diagnosis of cancerin 1980–1996 in New South Wales, Australia

Abstract

Survival from almost all cancers has improved during the last 30 years. There is debate over the reasons for the improvement. We examined trends in survival for 28 cancers from 1980 to 1996 in New South Wales (NSW), Australia, with adjustment for disease spread at diagnosis. NSW Central Cancer Registry data were used to estimate 5-year relative survival and relative excess risk of death for patients diagnosed in 1980–84, 1985–88, 1989–92 and 1993–96. Statistical significance of variation in excess deaths between periods of diagnosis was assessed using Poisson regression, with adjustment for age, sex, duration of follow-up, histology and spread of disease at diagnosis. There were statistically significant falls in excess deaths for 20 of the cancers with a 25% fall for all cancers combined. Cancers of the prostate, liver, thyroid, breast, gallbladder, body of uterus, rectum, cervix and ovary had falls of >30%. The falls varied by spread of disease; the largest being in localised and regionally spread tumours. Overall survival, when unadjusted for spread of cancer, generally fell in parallel with that in the specific categories of spread, which implies that stage migration did not contribute importantly to survival trends. While acknowledging the limitations of incomplete data on stage of cancer at diagnosis, we conclude that falls in excess deaths in NSW from 1980 to 1996 are unlikely, for many cancers, to be attributed to earlier diagnosis or stage migration; thus advances in cancer treatment have almost certainly contributed to them. © 2006 Wiley-Liss, Inc.

Survival from almost all cancers has improved, for some markedly, during the last 30 years. Notable successes include childhood leukaemia, testicular cancer and Hodgkin's disease, in which survival improvement has been mainly due to the introduction of more effective treatments.1, 2, 3 Between 1975–79 and 1995–2000 in the USA, 5-year-survival from female breast cancer increased from 75 to 88% and that for colorectal cancer from 50 to 64% (men) and 52 to 63% (women): these improvements were attributed to both earlier detection and more effective treatment of cancer.4 There has been debate, however, over the extent to which improved treatment has contributed to the trend in survival.5 By extrapolating trends in cancer mortality in the USA, Bailar and Gornick argued that newer cancer therapies have produced few real benefits and concluded that recent decreases in cancer mortality were due mainly to falling incidence or earlier detection.6

In this study, we examined time trends in excess risk of death within 5 years of diagnosis from 1980 to 1996 in patients with one of the 28 cancers, in New South Wales (NSW), Australia using data from a population-based cancer registry. To try to exclude impacts of earlier detection, we examined the trends in excess risk of death, with adjustment for a measure of disease spread at diagnosis, along with histological type of cancer, age and sex. In doing so, we also assessed possible effects of stage migration, a shift with time in the stage distribution of a cancer towards apparently higher stage disease because of more complete identification of disease spread. This shift produces an artificial increase in survival in each stage category because of the removal of more advanced disease from earlier stage categories and its transfer as relatively less advanced disease into later stage categories.7 Taking confounding trends in stage and the possibility of stage migration into account allows an interpretation of the trends in excess risk of death in terms of changes in disease management.

Material and methods

Data

Data for patients diagnosed with any of the 28 major cancers during 1980–96 were obtained from the population-based NSW Central Cancer Registry. Notification of cancer has been a statutory requirement for all NSW public and private hospitals, radiotherapy departments and nursing homes since 1972, and for pathology departments since 1985.8 The Central Cancer Registry generally has high standards of data completeness, quality and follow-up; the data are accepted by the International Agency for Research on Cancer for publication in Cancer Incidence in Five Continents.9 Individuals with the first occurrence of a primary cancer between 15 and 89 years of age were included. Cases notified by death certificate only or first identified at postmortem were excluded from the analysis of survival, but included in the calculation of age–sex standardised incidence rates. These cases were 1.8% of the total and were relatively constant over time, except for an increase to 3.3% in 1985–88 caused by lack of Registry resources to investigate them. Data on the population and population mortality used to calculate relative survival and age and sex standardised incidence rates were obtained from the Australian Bureau of Statistics, which conducts Australia's quinquennial population census and collates national death data.

All cases diagnosed from 1980 to 1996 were followed to December 2001 to determine survival status. Identifiers from each were compared with those of all records of deaths in the State Register of Deaths and the National Death Index from their date of diagnosis to 31st December 2001 to find a matching death record, if present. This passive approach to follow-up may fail to ascertain all deaths and may incorrectly link some incidence and death records. A study investigating its completeness and accuracy found loss to follow-up to be uniform from 1980 to 1993 and estimated the resulting overestimation of relative survival to be a maximum of 2%.10 The end of follow-up was the date of death for those who died within five years of diagnosis or five years after diagnosis for those who survived the first 5 years.

All information on primary cancer site and histology was coded according to the International Classification of Diseases for Oncology, second edition (ICDO-2).11 Data on spread of disease at diagnosis were provided by hospital medical record departments and radiotherapy notifiers, and classified into four broad categories: localised, regional (including adjacent organs and regional lymph nodes), distant and unknown to the Registry. This summary classification of stage is used by a number of major cancer registries around the world, including registries in the Surveillance, Epidemiology, and End Results (SEER) program in the USA. While not as detailed as the standard TNM staging system, it can be applied to most cancers occurring in whole populations.12 Degree of spread was not applicable to staging for Hodgkin's disease, non-Hodgkin lymphoma, multiple myeloma, leukaemia and brain cancer, and no other staging data were available for them. Because of the importance of spread of disease in our analysis and the possibility of stage migration, we tabulated changes in the distribution of spread in 1980–96 for the sites for which it was available, with regional and distant spread combined as nonlocalised disease.

Statistical methods

Relative survival and relative excess risk of death.

Cancer patients were followed for five years after diagnosis and relative survival was estimated using the cohort method. Relative survival is the ratio of the observed proportion surviving in a group of patients to the expected proportion that would have survived in a comparable group of people from the general population.13 Observed survival was estimated using the life table method.14 The expected survival from the general population was calculated using all cause mortality for the NSW population by single year of age, sex and calendar year.15

The excess risk of death after diagnosis of a cancer is the risk of death above what would have been observed if the population death rates had been applied to the cancer patients. To analyse trends in excess risk of death, four time periods were defined: 1980–84, 1985–88, 1989–92 and 1993–96. Grouping the dates of diagnosis in periods of a few years increases the likelihood that cancer patients diagnosed within a period followed similar treatment protocols and had similar access to screening. The period 1980–84 preceded compulsory reporting of cancer to the Cancer Registry by pathology laboratories, which was introduced in 1985.

Statistical modelling of excess risk of death.

To determine the change in survival over time after adjustment for possible confounders, we fitted a Poisson regression model for excess deaths from each type of cancer.16 The model included time period of diagnosis, age group at diagnosis (15–44, 45–59, 60–74 and 75–89 years), year of follow-up since diagnosis, sex (where applicable), histological type (based on ICDO-2 and with the less common histological types grouped together) and spread of disease at diagnosis (where applicable) as independent variables. We then fitted another model by adding the interaction of spread of cancer by period of diagnosis to the main effects model for cancers to which spread of disease at diagnosis was applicable. Finally, we compared the difference between the deviance from this model with that from the main effects model to determine whether addition of the interaction produced a statistically significant difference in model deviance (p < 0.05), using the chi-square test. If the difference was statistically significant, we then carried out further analyses to examine the differences in trends across categories of spread of cancer.

The modelling methods we used are described in detail by Dickman et al.17 Briefly, data from individual records were aggregated to yield a count of deaths for each combination of the variables included in the model, and then a generalised linear model with a Poisson error structure based on aggregated data using exact survival time (person-years) was fitted for each cancer. This model quantifies the extent to which the excess risk of death in a given period differs from the excess risk of death in the reference period (1980–84) after controlling for the factors included in the model. The relative excess risk of death (RER) in the period 1980–84 was set to a value of 1. A RER of less than 1 in another period indicates that the excess risk of death in that period was less than that in the reference period, and vice versa. Ninety-five per cent confidence intervals (CIs) for the RERs were calculated using the estimated coefficients and standard errors from the Poisson models. The statistical significance of each variable in the model was determined by the log-likelihood ratio test with a p-value of <0.01 taken to indicate statistical significance. All analyses were done using SAS version 8.2 and the procedure GENMOD was used to fit the models and assess the effects of the variables on excess risk.

Trends in incidence rates.

To estimate trends in incidence, which we reported to give context to the survival trends, we calculated annual age-sex standardised incidence rates for the resident population in NSW during the period 1980–96 for each of the 28 cancers. These rates were expressed per 100,000 of the population, and age and sex were adjusted by the direct method to the Australian estimated residential population of 2001. Trends in incidence were summarized by calculating the annual percent change in age-standardised incidence rates over the 17 years for each cancer. The annual percent change was estimated by fitting a Poisson regression model to the natural logarithm of the rates, with calendar year as a continuous independent variable. The assumption of a linear trend was reasonable for all cancers except for melanoma and prostate cancer in which there have been sizeable short-term perturbations in the long-term trend.8

Results

A total of 343,034 newly diagnosed cancers were included in this analysis, with the commonest cancers being breast (41,476), lung (39,769) and prostate (37,374) (Table I). Age-standardised incidence of most cancers increased during the period 1980–96. The largest annual percent increases were for cancers of the prostate, liver and thyroid, and mesothelioma, and the largest falls were for cancers of the stomach, cervix and bladder. The incidence of prostate cancer showed a dramatic rise between 1990 and 1994, followed by a fall after 1994.8 The Registry's report on cancer incidence in 20038 is available at http://www.cancerinstitute.org.au/cancer_inst/statistics/pdfs/IncidenceMortalityReport2005.pdf.

Table I. Age and Sex Standardised Incidence Rates (per 100,000) and Average Annual Percent Change, with the Corresponding 95% CI, for 28 Cancers, in NSW, Australia, 1980–96

Cancer type	Number of new cases	Age-standardised incidence rates1	Average annual percent change 1980–96	95% CI
1 Age and sex adjusted to the Australian estimated residential population of 2001.
Lip	3,562	9.0	2.51	0.85, 4.19
Head and neck	12,553	28.3	0.00	−0.58, 0.58
Oesophagus	4,167	10.5	1.21	0.54, 1.88
Stomach	10,354	22.5	−2.66	−3.11, −2.21
Colon	32,414	80.6	0.56	0.19, 0.93
Rectum	17,688	44.5	0.94	0.57, 1.32
Liver	1,891	6.7	7.91	6.48, 9.36
Gallbladder	2,709	6.6	0.29	−0.63, 1.23
Pancreas	8,091	19.8	0.04	−0.43, 0.51
Lung	39,769	90.7	−0.56	−0.78, −0.33
Melanoma	32,316	84.4	3.06	2.15, 3.97
Mesothelioma	1,634	5.1	5.11	4.14, 6.08
Connective tissue	2,088	4.6	−0.16	−1.25, 0.93
Breast	41,476	113.9	2.71	2.30, 3.13
Cervix	5,957	11.6	−1.63	−2.22, −1.04
Body of uterus	5,793	14.0	0.93	0.27, 1.60
Ovary	5,375	12.1	−0.30	−0.87, 0.27
Prostate	37,374	146.1	7.09	5.46, 8.75
Testis	2,314	5.4	2.68	1.79, 3.57
Bladder	12,139	25.2	−2.90	−3.63, −2.16
Kidney	9,053	24.3	2.79	2.24, 3.33
Thyroid	3,438	10.2	4.92	3.86, 5.99
Brain	5,404	12.8	1.08	0.62, 1.55
Hodgkin's disease	1,856	3.7	−0.76	−1.49, −0.03
Non-Hodgkin lymphoma	12,688	34.8	3.04	2.68, 3.41
Multiple myeloma	4,229	11.0	1.09	0.38, 1.81
Leukaemia	9,365	23.5	0.48	−0.04, 1.00
Unspecified	17,337	40.2	−0.49	−1.03, 0.05

There were substantial changes in the distribution of the cancers by spread of disease at diagnosis over the period of study (Table II). The proportion with localised disease fell for all but three cancer types—melanoma, breast cancer and testicular cancer—and the proportion with unknown stage increased for all except these same three cancer types. The proportions with regional and distant disease fell for 15 cancer types (including melanoma and breast cancer) and increased for 8 types, the most substantial being in cancers of the colon, rectum and ovary (5 percentage points or more). The increase in the proportion of cancers of unknown stage occurred mainly in 1993–96 and, to a much less extent, in 1989–92 for 13 of the 20 cancer types in which an increase occurred. These increases were probably largely due to a change from paper-based to electronic notification of cancer from some hospitals, introduced from 1992, which meant that some information on stage provided through manual notification was no longer available to the Registry. In the remaining 7 cancer types, the increase was more gradual across the whole period of observation, indicating that additional factors, unknown to us, were influencing completeness of stage information reported to the Registry for these cancers.

Table II. Proportion of Spread of Disease at Diagnosis by Period of Diagnosis for 23 Major Cancers Diagnosed in NSW, Australia, 1980–961

Cancer type	Localised				Non-localised				Unknown
	1980–84	1985–88	1989–92	1993–96	1980–84	1985–88	1989–92	1993–96	1980–84	1985–88	1989–92	1993–96
1 Brain cancers, non-Hodgkin lymphoma, Hodgkin disease, multiple myeloma and leukaemia are not included in this table because classification of stage at diagnosis on the basis of spread of disease at diagnosis does not apply to them.
Lip	81.0	72.4	69.7	62.7	7.2	6.9	6.6	5.6	11.8	20.8	23.7	31.7
Head and neck	47.1	45.3	41.9	30.8	36.8	36.2	36.3	39.5	16.0	18.5	21.8	29.7
Oesophagus	34.8	42.0	38.5	27.4	42.3	36.0	35.7	37.9	22.9	22.0	25.8	34.7
Stomach	24.7	25.2	23.5	17.1	59.2	57.3	60.1	59.9	16.1	17.5	16.4	23.0
Colon	33.7	30.8	30.4	23.7	55.5	57.5	60.9	63.8	10.8	11.7	8.7	12.4
Rectum	42.3	39.4	38.2	30.9	47.9	49.5	51.9	52.8	9.8	11.2	9.9	16.2
Liver	49.6	39.1	37.7	23.6	32.7	23.5	22.5	18.7	17.7	37.3	39.7	57.8
Gallbladder	24.3	25.0	23.6	17.3	64.5	57.5	56.1	49.9	11.3	17.5	20.2	32.8
Pancreas	14.3	16.4	16.5	9.6	68.5	56.1	56.0	52.6	17.2	27.5	27.5	37.8
Lung	23.2	30.3	26.4	17.0	49.8	44.7	44.6	46.8	27.0	25.0	29.0	36.2
Melanoma	82.9	87.3	88.2	89.0	9.6	7.5	8.1	6.8	7.5	5.2	3.7	4.2
Mesothelioma	28.8	42.7	35.9	19.6	49.2	18.6	21.7	21.6	22.0	38.7	42.4	58.8
Connective tissue	53.8	55.6	48.2	34.3	27.0	21.8	18.9	16.0	19.2	22.5	32.9	49.7
Breast	47.3	47.9	49.3	51.0	37.6	37.8	37.8	33.5	15.1	14.3	12.9	15.4
Cervix	57.0	67.3	63.7	50.7	31.3	24.9	28.6	25.9	11.7	7.8	7.7	23.4
Body of uterus	67.9	69.4	65.7	61.5	17.2	17.8	20.8	21.3	14.9	12.8	13.5	17.2
Ovary	28.2	27.4	24.4	16.0	61.1	58.7	65.4	70.0	10.7	13.9	10.2	14.0
Prostate	59.0	52.5	51.1	40.6	21.8	21.4	16.6	10.0	19.3	26.1	32.3	49.4
Testis	59.1	65.5	66.6	60.5	20.4	23.5	23.1	21.2	20.4	11.0	10.4	18.3
Bladder	77.8	71.7	65.9	47.2	11.6	11.6	15.3	15.4	10.6	16.7	18.8	37.4
Kidney	51.2	49.8	50.1	44.3	38.5	38.8	36.0	35.1	10.3	11.4	13.9	20.6
Thyroid	64.9	59.0	62.7	51.7	25.4	28.6	24.3	24.9	9.7	12.4	13.1	23.4
Unspecified	0.7	1.4	1.2	0.8	95.6	88.6	89.4	87.4	3.7	10.0	9.4	11.8

The excess risk of death fell over time for most cancers except cancers of the connective tissue and unspecified site, for which it rose significantly (p < 0.0001), and cancer of the bladder, where the increase was not statistically significant (Table III). For all cancers together, the excess risk of death fell by a relative 25%. The most dramatic falls were in cancers of the liver, prostate and thyroid, with the excess risk of death in the latest period being less than 60% of that in the earliest period. Other notable falls were for cancers of the oesophagus, rectum, gallbladder, breast, cervix, uterus and ovary.

Table III. Relative Excess Risk (RER) during the First 5-Years after Diagnosis, with Adjustment for Age, Sex, Spread of Cancer, Years since Diagnosis and Histological Type for 28 Cancers Diagnosed in 1980–96 and followed to 2001, in NSW, Australia

Cancer type	5-year RSR1 1993–96 (%)	RER2 and 95% CI				p-value
		1980–84	1985–88	1989–92	1993–96
Values in parentheses represent 95% CI. 1 Five-year relative survival for the period 1993–96. 2 RER for the period of 1980–84 as reference (set to 1.0). Each model included age group at diagnosis, sex, year since diagnosis, period of diagnosis, histological type and spread of cancer. 3 p-value for the main effect of period for cancer types with nonsignificant interaction term for period by stage. 4 p-value for the interaction of period by stage for cancer types with significant interaction term for period by stage. 5 Classification of stage by spread of cancer not applicable for these cancers.
Lip	93.1	1.00	0.87 (0.52–1.47)	0.61 (0.36–1.04)	0.72 (0.43–1.19)	0.283
Head and neck	55.6	1.00	0.93 (0.86–1.01)	0.87 (0.81–0.95)	0.78 (0.72–0.85)	<0.00013
Oesophagus	15.4	1.00	0.85 (0.77–0.94)	0.75 (0.68–0.83)	0.70 (0.63–0.77)	<0.00014
Stomach	23.8	1.00	0.93 (0.87–1.00)	0.88 (0.82–0.94)	0.82 (0.77–0.88)	<0.00014
Colon	60.0	1.00	0.84 (0.80–0.89)	0.80 (0.76–0.85)	0.71 (0.68–0.75)	<0.00014
Rectum	59.4	1.00	0.86 (0.80–0.92)	0.75 (0.70–0.81)	0.67 (0.62–0.71)	<0.00014
Liver	11.4	1.00	0.71 (0.60–0.85)	0.73 (0.62–0.87)	0.55 (0.47–0.65)	<0.00013
Gallbladder	19.5	1.00	0.83 (0.73–0.93)	0.78 (0.69–0.88)	0.61 (0.54–0.69)	<0.00014
Pancreas	5.3	1.00	0.85 (0.79–0.91)	0.94 (0.88–1.01)	0.86 (0.80–0.92)	<0.00014
Lung	12.5	1.00	0.94 (0.91–0.97)	0.90 (0.87–0.93)	0.82 (0.79–0.84)	<0.00014
Melanoma	90.9	1.00	0.82 (0.73–0.93)	0.72 (0.64–0.81)	0.72 (0.64–0.81)	<0.00013
Mesothelioma	4.9	1.00	0.88 (0.74–1.04)	0.87 (0.74–1.02)	0.92 (0.79–1.07)	0.353
Connective tissue	63.0	1.00	1.08 (0.86–1.35)	1.11 (0.89–1.39)	1.05 (0.84–1.30)	<0.00014
Breast	85.0	1.00	0.88 (0.82–0.94)	0.76 (0.71–0.81)	0.61 (0.57–0.65)	<0.00014
Cervix	73.1	1.00	1.03 (0.91–1.18)	0.78 (0.68–0.89)	0.68 (0.59–0.78)	<0.00013
Body of uterus	79.2	1.00	0.79 (0.66–0.94)	0.66 (0.55–0.79)	0.61 (0.51–0.72)	<0.00013
Ovary	37.3	1.00	1.03 (0.93–1.14)	0.84 (0.76–0.92)	0.68 (0.62–0.75)	<0.00014
Prostate	86.9	1.00	1.09 (1.01–1.19)	0.95 (0.87–1.03)	0.54 (0.49–0.59)	<0.00014
Testis	95.5	1.00	0.72 (0.42–1.23)	0.51 (0.31–0.84)	0.63 (0.38–1.04)	0.053
Bladder	62.5	1.00	1.06 (0.96–1.18)	1.08 (0.97–1.21)	1.13 (1.02–1.26)	0.143
Kidney	57.4	1.00	0.86 (0.78–0.94)	0.85 (0.77–0.93)	0.73 (0.67–0.80)	<0.00014
Thyroid	93.5	1.00	0.81 (0.57–1.14)	0.63 (0.44–0.89)	0.58 (0.42–0.81)	<0.00013
Brain5	17.9	1.00	0.88 (0.80–0.96)	0.94 (0.86–1.02)	0.86 (0.79–0.94)	0.0033
Hodgkin's disease5	77.3	1.00	1.13 (0.86–1.48)	0.99 (0.74–1.33)	0.85 (0.63–1.15)	0.263
Non-Hodgkin lymphoma5	53.6	1.00	0.94 (0.86–1.02)	0.92 (0.85–0.99)	0.86 (0.80–0.93)	0.0023
Multiple myeloma5	31.9	1.00	0.95 (0.85–1.07)	0.91 (0.82–1.02)	0.86 (0.77–0.97)	0.073
Leukaemia5	38.2	1.00	0.84 (0.78–0.92)	0.74 (0.68–0.80)	0.82 (0.76–0.88)	<0.00013
Unspecified	11.0	1.00	1.06 (1.01–1.11)	1.04 (0.99–1.09)	1.10 (1.05–1.15)	<0.00014
All cancer	59.6	1.00	0.92 (0.91–0.94)	0.86 (0.85–0.88)	0.75 (0.74–0.76)	<0.00014

There were nonsignificant falls in excess risk of death from cancers of the lip and testis, mesothelioma, Hodgkin's disease and multiple myeloma. The CIs about the RERs of less than unity for cancers of the lip and testis diagnosed in 1993–96 were wide and included the point estimate of 0.75 for RER of all cancers together. These were among the least frequent of the cancers studied and were in the 3 least fatal; thus real downtrends in excess risk of death for these cancers are possible, but our data had limited statistical power to detect them with any certainty. Hodgkin's disease and multiple myeloma shared with non-Hodgkin lymphoma a 14–15% reduction in excess risk of death between 1980–84 and 1993–96, which was statistically significant for the latter.

The trend in excess risk of death was significantly heterogeneous among categories of spread of disease for 13 of the 23 cancers for which disease stage was available. Some patterns are shown in terms of trends in 5-year relative survival in Figure 1. For all cancers combined and cancers of the colon, breast and prostate, the main trends in 5-year relative survival were increases in the localised, regional and unknown categories, with no improvement for distant cancers. For cancer of the rectum, the 5-year relative survival increased in all categories though much less so for cancers with distant spread. For ovarian cancer, the increase in survival was evident in cancers that were localised, regional or distant but hardly at all in unknown spread or all degrees of spread together (unadjusted for degree of spread); this strongly suggests that stage migration caused the apparently increased survival in the individual degree of spread categories (Fig. 1).

Figure 1. Trends in 5-year relative survival by spread of cancer for all cancers combined and cancers of the breast, colon, ovary, prostate and rectum (New South Wales, Australia, 1980–96).

There was also evidence of stage migration in the pattern of change in the degree of spread with time for cancer of the ovary: the proportion of localised cancers fell, that of regionally and distant spread cancers increased correspondingly and that of unknown spread cancers changed little. Cancers of the colon, rectum, head and neck, and body of uterus showed similar patterns of change in degree of spread. For colon and rectal cancers, however, overall survival, unadjusted for spread of disease, paralleled the trends in survival in the individual spread of cancer categories, thus ruling out substantial stage migration (Fig. 1). For cancers of the head and neck and body of uterus, the uptrend in overall survival was modest relative to that in individual categories of degree of spread (data not shown), thus suggesting some stage migration.

Discussion

We found that excess risk of death after diagnosis for 20 of the 28 categories of cancer type, adjusted where possible for degree of spread of cancer (cancer stage) at diagnosis, fell significantly, between 1980 and 1996 in New South Wales. These results are consistent with the beneficial effects of newer cancer therapies on cancer survival but incomplete data on degree of spread and its increase with time may limit the confidence with which this conclusion can be drawn.

Our data are population-based and thus represent the experience of a general population of people with cancer, not one that has been selected by referral to a particular hospital or expert centre. We were able to take some account of trends in stage, which many population-based studies cannot because stage data are not collected by the relevant cancer registry. Our observations are also based on large numbers of patients and deaths and thus can give quite precise estimates of trends for many different cancers.

That our data on stage of cancer are incomplete and have become more incomplete with time mean that we cannot rule out entirely the possibility that increased cancer screening or better methods of diagnosis, leading to earlier diagnosis, have contributed to the favourable trends we have observed.18 In addition to the impacts on cancer mortality that screening and improved diagnosis might have had, both lead-time bias (advancing the date of diagnosis without postponing the time of death) and length bias (detection of slower growing tumours that would not otherwise have been diagnosed or have caused death) could have produced apparent falls in excess risk of death.19, 20 The measure we used to adjust for stage, spread of disease at diagnosis, is a very powerful predictor of survival in our data,10 but it was missing for more than 10% of most cancer types; this high prevalence of missing data would reduce our ability to control statistically for effects of trends in earlier diagnosis on trends in survival. These issues notwithstanding, the lack, for most cancers, of an increase in the proportion of cases with localised stage (Table II) suggests that these factors were not generally important contributors to apparently improved survival.

Stage migration could also have produced artefactual falls in observed excess risk of death. As suggested earlier, this might explain the whole of the apparent improvement in stage-adjusted survival for ovarian cancer and might have contributed to the falls in stage-adjusted survival for cancers of the head and neck and body of uterus. It was probably not important for other cancers.

There would, perhaps, be greatest concern with inadequate adjustment for stage at diagnosis in cancers for which the proportion of early stage disease increased with time: melanoma, breast cancer and testicular cancer. Introduction of mammographic screening for breast cancer in the 1980s should have reduced the excess risk of death from breast cancers diagnosed in 1980–96.21 Its effects are probably evidenced in the increasing proportion of localised disease and the somewhat greater downtrend in excess risk for breast cancers than for most other cancers (Table III). However, if screening was the main reason for the reduced excess risk of death, survival improvement should have been most evident in the target age group for screening (50–69 years), whereas it was seen in all age groups (Fig. 2). There is no formal screening program for melanoma in Australia; self and professional skin examination is strongly encouraged and there is evidence of its impact.22 Here the problem of adequate stage adjustment is even greater: as most melanomas are diagnosed when localised to the skin, a measure of local stage, namely thickness of the melanoma, is a much better indicator of early diagnosis than is clinical stage. We did not use it in this study. Thus we cannot infer benefits from trends in melanoma treatment from our results. While testicular self examination has been promoted in Australia, the trend towards an increase in localised cancer was paralleled by an increase in nonlocalised cancer (Table II) and a fall in cancer of unknown stage. Thus there is no certain trend to earlier diagnosis of testicular cancer and the reduction in excess risk of death for this cancer probably reflects the known improvements in treatment in the 1970s and 80s2, with the lowest RER in 1989–92 (Table III) consistent with achievement of a minimum in testicular cancer mortality in NSW in this period.8

Figure 2. Trends in 5-year relative survival by age at diagnosis for breast cancer (New South Wales, Australia, 1980–96).

Although there is no formal screening program for prostate cancer in Australia and any trend towards earlier diagnosis of it has probably been masked by the great increase in proportion of cancers of unknown stage, the near 50% apparent fall in excess risk of death from it has probably been caused by the large increase in screening with prostate specific antigen (PSA) testing in Australia in the 1990s, with the associated large increase in prostate cancer incidence.23 The 5-year relative survival for localised prostate cancers increased from 82.7% in 1989–92 to 97.2% in 1993–96 (Fig. 1). Compared with the model without adjustment for spread of cancer, adjustment increased the RER in 1993–96 from 0.38 to 0.54 (data not shown). Thus, although adjustment for spread of cancer removed more than 40% of the reduction in excess risk for prostate cancer, the recently increased proportion of prostate cancers (49.4%; Table II) with missing spread of disease would probably have prevented full control of the effects of stage shift. The increasing proportion with unknown stage for cancers of the liver and thyroid would similarly reduce our ability to control for stage; these also showed substantial falls in excess risk over the period. The fall in excess risk for liver cancer may be due to its rapidly rising incidence due to chronic viral hepatitis in Australia8, 24 and associated greater detection due to surveillance of infected patients. Thyroid cancer is similarly increasing in incidence8 and at least some of this is due to greater detection, possibly of lesions with limited potential to advance.25

Our results are consistent with those of population-based studies of trends in cancer survival from other countries in which the effects of trends in stage at diagnosis have been considered. The most comprehensive analysis of this type was presented by Dickman et al. using 560,000 cases in 37 categories of cancer type registered by the Finnish cancer registry, diagnosed in 1955–94 and followed up to 1995.26 Stage data were available in the same categories as we have used and 22% of all cancers were of unknown stage. Time trends in relative survival were presented only graphically and without adjustment for possible confounding variables. Five-year relative survival from all cancers increased by about 20 percentage points from 1955–64 to 1985–94. Similar uptrends were observed in each category of stage, with the absolute increase being greater in localised and, to a lesser extent, in disease with regional spread than in disease with distant spread. This is similar to the pattern we observed (Fig. 1). A few population-based studies of cancer survival in other countries, limited to colorectal or breast cancers, have taken account of trends in stage at diagnosis. All except one of colorectal cancer, in Singapore, were done in European countries, and all found reductions in fatality or increases in survival, which were apparently independent of trends in stage.27, 28, 29, 30, 31

How plausible is it that downtrends in excess risk of death that we have observed are due to improvements in cancer management? There are good grounds for believing that such improvements have occurred. Surgical techniques developed during the 1980s and the introduction of adjuvant chemotherapy in the 1990s for colorectal cancer patients may have contributed to their apparently improved outcome.32 These modalities are now in common use in NSW.33 The increased use of tamoxifen and adjuvant chemotherapies since the later 1980s should have contributed to the improved survival for breast cancer patients.18, 34, 35 These therapies too are in common use in Australia.36, 37 For prostate cancer, increased survival could also be due to changes in treatment practice in the late 1980s and early 1990s when hormonal therapy was introduced for patients with advanced disease and older patients.38, 39 More recently, increasing use of radical prostatectomy in early stage prostate cancer may also have improved outcome.40

The approximate 15% fall in excess risk of death from each of Hodgkin's disease, non-Hodgkin lymphoma and multiple myeloma is compatible with 9–16% relative increases in 5-year survival from these cancers between 1980–84 and 1995–97 in data from the US SEER registries.41 The relatively modest changes for these three related cancers are probably due to the most important therapies that changed outcomes for these diseases, having been already well established in 1980.42, 43, 44

The small but significant uptrends in excess risk of death from cancers of the connective tissue and unspecified sites might be explained by their falls in incidence over the period of study (Table I) due, perhaps, to increasing classification of better differentiated or less widely spread cancers to more specific sites with more specific histopathological diagnosis or more effective location of the probable site. The statistically nonsignificant uptrend in excess risk for bladder cancer is probably mainly due to a fall in registration of noninvasive tumours of the bladder after 1985, with availability of pathology reports and reduced reporting of bladder papillomata as cancer.45

These considerations notwithstanding, there remains justifiable concern about the impact of the increase in the proportion of unknown stage cancers on our adjustment for stage of cancer and especially the possibility that the distribution of stages within this unknown stage group may also have changed with time, a possibility we cannot rule out. There was, however, little correlation (Pearson's correlation coefficient of 0.10) between the size of the change in the proportion of unknown stage between 1989–92 and 1993–96 and the effect of stage adjustment on the RER in 1993–96, which suggests little such bias. Moreover, a comparison of Tables II and III shows that for all cancers for which there was little change in the proportion of unknown stage cancer between 1980–84 and 1989–92, there was still an important reduction in excess risk, with an upper confidence limit for the RER of less than 1.0 in this interval. This was so for cancers of the oesophagus, stomach, colon, rectum, lung, cervix, body of uterus and kidney, for which we suspect no other sources of bias, as well as for melanoma, breast cancer and thyroid cancer, for which we do. Thus for the former set of these, at least, the best explanation for reduced fatality in the period of study is improvement in treatment.

Acknowledgements

We thank the three anonymous reviewers for providing constructive comments, which led to significant improvement in the manuscript. We thank the NSW Central Cancer Registry for providing the data. Bruce Armstrong's research is supported by a University of Sydney Medical Foundation Programme Grant.