In international comparisons or time trend analyses of cancer patient survival, it is common practice to restrict analyses to patients with a first cancer diagnosis, i.e., to exclude patients with a previous cancer diagnosis. For example, in various rounds of the EUROCARE study, a comparative analysis of cancer survival in Europe, patients with a previous diagnosis of cancer were generally excluded.1, 2 However, the proportion of patients excluded because of this criterion varied between cancer registries [e.g., it ranged from 0 to 18.1% in the EURCOCARE-III study2], and it was higher for cancer registries initiated many years ago, such as the registries from the Northern European countries, than in “younger” cancer registries. This pattern reflects the fact that the proportion of cancer patients with a known previous cancer diagnosis depends on the time a cancer registry has been running, as a previous diagnosis has a higher chance to be found in the records of an “older” registry. If prognosis of patients with a previous cancer diagnosis differs from prognosis of patients with a first cancer diagnosis, varying degrees of exclusions of patients with a previous cancer diagnosis may bias survival comparisons between registries with different running times. However, to our knowledge, no previous studies have addressed potential occurrence and size of such bias. The aim of this study was to empirically evaluate the dependence of the proportions of cancer patients with known previous diagnosis on the time since initiation of cancer registration and to evaluate the impact of excluding such patients on cancer survival estimates.
In international comparisons or time trend analyses of cancer survival, it is common practice to restrict analyses to patients with a first cancer, i.e., to exclude patients with previous cancer diagnoses. However, the proportion of cancer patients with known previous cancer depends on the time cancer registries have been running, which results in varying proportions of excluded patients across registries. If prognosis of patients with second cancers differs from prognosis of patients with first cancers, varying exclusions may bias survival comparisons. We empirically evaluate the dependence of proportions of patients recorded as having a first cancer on time since initiation of cancer registration and the impact of excluding patients with known previous cancer on cancer survival estimates using the data of the nationwide Finnish Cancer Registry. Among 20 common cancer sites investigated, the proportion of “first cancers” varied between 97.4 and 99.7% in 1953–1957, the first 5-years of cancer registration, and decreased continuously to levels between 83.9 and 92.7% in 1993–1997. Excluding patients with a previous cancer diagnosis had little impact on estimates of survival of cancer patients diagnosed in 1953–1957, but increased 5-year relative survival estimates among patients diagnosed in 1993–1997 for each of the 20 cancers. The extent of the increase varied by cancer site and age. The increase ranged up to 2.9% points for crude and up to 1.7% points for age adjusted 5-year relative survival. These results caution against exclusion of patients with previous cancer diagnosis in comparative analyses of cancer survival. © 2007 Wiley-Liss, Inc.
Our analysis is based on data from the Finnish Cancer Registry which is among the oldest population-based cancer registries in the world, and which is well known for its high levels of data quality and completeness.3 Cancer registration is mandatory by law in Finland, and essentially complete population coverage of cancer registration has been achieved since 1953, 1-year after its initiation. Analyses are shown for patients diagnosed with 1 of 20 common forms of cancer in 1953–1997. Patients aged 15 or older at diagnosis were included. We excluded patients notified by death certificate or by autopsy only, and those with unknown month of diagnosis.
In a first step, we determined the total numbers of patients registered and the proportion of them registered as having a first cancer diagnosis for consecutive 5-year calendar periods 10-years apart, starting with 1953–1957, i.e., for calendar periods 1953–57, 1963–67, 1973–77, 1983–87 and 1993–97. Next, for each of the 20 forms of cancer, we assessed 5-year relative survival of all patients diagnosed within each calendar period and its change by exclusion of patients with known previous cancer diagnosis. Relative rather than absolute survival was estimated, as this is the approach most commonly used in population-based cancer survival analyses. Relative survival reflects the probability of surviving the cancer of interest rather than the total survival probability,4, 5 taking expected deaths in the absence of cancer into account. For this analysis, the expected numbers of deaths were derived from age, gender and calendar period specific mortality figures of the general population of Finland according to Hakulinen's method.6
Because the probability of having a previous cancer diagnosis is expected to increase with age, we carried out additional analyses, in which the proportions of patients with a first cancer diagnosis and the change in survival estimates by restricting the analysis to patients with a first cancer diagnosis were assessed by age among patients registered in 1993–1997. Five major age groups, which are commonly used in age specific analyses of cancer survival were assessed: 15–44, 45–54, 55–64, 65–74 and 75+ years. In addition to age specific analyses, we also calculated age adjusted 5-year relative survival, using the standard cancer populations proposed by Corazziari et al.7 and the method proposed by Brenner et al.8 for age adjustment.
Overall, 577,924 patients aged 15 or older were reported to the Finnish Cancer Registry with a first diagnosis of cancer between 1953 and 1997. Of these, we excluded 2.6% notified by death certificate only, another 2.6% notified by autopsy only, and 0.1% because of missing information on month of diagnosis. The 20 forms of cancer specifically addressed in this article include about 89.6% of the remaining cancer cases (n = 490,279).
The numbers of cancer patients diagnosed in various 5-year periods are shown in Table I. In 1953–1957, stomach cancer was the by far the most common form of cancer in Finland, followed by lung cancer and breast cancer. For all but 3 forms of cancer (stomach, esophagus and cervical cancer), case numbers per 5-year period strongly increased over time. The increase was particularly pronounced for breast and prostate cancer, which were the most common forms of cancer in 1993–1997. The minimum number of cancer cases per 5-year period was 299 (thyroid cancer in 1953–1957), but for most combinations of cancer type and calendar period, case numbers were well above 1,000.
|Cancer site||Total number of cases||Proportion of first cancers (%)|
In 1953–1957, the proportion of 1st cancers ranged from 97.4% (urinary bladder cancer) to 99.7% (lymphomas), with a median level of 98.8%. The proportion of cancers registered as 1st cancers steadily decreased over time for all forms of cancers, reaching ranges (median levels) of 95.5–99.2% (97.0%) in 1963–1967, 92.9–97.5% (94.6%) in 1973–1977, 88.1–96.4% (91.0%) in 1983–1987 and 83.9–92.7% (88.0%) in 1993–1997.
Table II shows 5-year relative survival by cancer site and period of diagnosis. Five-year relative survival varied strongly between cancer sites and substantially increased over time for most, but not all cancer sites. Excluding patients with a previous cancer diagnosis had little impact on survival estimates of cancer patients diagnosed in 1953–1957, but increased 5-year relative survival estimates among patients diagnosed in 1993–1997 for each of the 20 cancers. Although this increase was below 1% point for 11 cancers, larger increases up to 2.9% points were seen for other common cancers.
|Cancer site||5-year relative survival of all cases||Change in 5-year relative survival (in % units) when the analysis is restricted to 1st cancers|
Table III shows the numbers of cases and the proportion of first cancers registered in 1993–1997 by 5 major age groups. With few exceptions (which might be due to random variation or due to small age-specific numbers, especially in the younger age groups), the proportion of first cancers generally decreases by age, reaching ranges of 83.9–91.1% in age group 65–74 and 72.6–88.0% in age group 75+.
|Cancer site||Total number of cases by age||Proportion of first cancers (%) by age|
With few exceptions (especially breast cancer and prostate cancer), 5 year relative survival of patients diagnosed in 1993–1997 generally decreased with age (Table IV). Restriction of analysis to1st cancers increased estimates of 5-year relative survival in most cases even within age groups. With values of up to 3.5 (leukaemia), 2.2 (oral cavity) and 2.6 percentage points (leukaemia), increases were strongest in age groups 75+, 65–74 and 55–64 years, respectively, where the largest proportions of patients with a previous cancer diagnosis were excluded. Exclusion of patients with a previous cancer diagnosis also increased age adjusted survival estimates for most cancers, albeit to a lesser extent than the crude survival estimates. With 1.7% points, the largest increase was again seen for patients with leukaemia. For all other cancers, the increase was below 1% point.
|Cancer site||5-year relative survival of all cases, by age||Change in 5-year relative survival (in % units) when the analysis is restricted to 1st cancers, by age|
To our knowledge, this is the first study that empirically addressed the impact of the “age” of a population-based cancer registry on the proportion of cancer patients registered as having a first cancer diagnosis and its implications on assessing survival after a first cancer. We found a steady decrease of the proportion of first cancers from levels close to 100% during the first 5-years of registration to median levels around 97, 95, 91 and 88% for 20 common forms of cancer 10, 20, 30 and 40 years later. Excluding patients with a previous cancer diagnosis increased estimated 5-year relative survival in most cases. This increase was generally modest in the first decades after initiation of cancer registration, but became larger with increasing “running time” of the registry. The increase was partly because of higher exclusion rates of olderpatients, who have lower survival for most cancers, and it was reduced to some extent, but not entirely overcome by age adjustment.
The rationale for excluding patients with a previous cancer diagnosis in comparative analyses of cancer survival is to remove the influence of a cancer other than the one of primary interest, which might, in theory, lead to a more homogeneous definition of the disease. However, such exclusion does not guarantee derivation of a measure of “net survival” for the cancer of interest, as it does not remove the influence of other chronic diseases, which may have an even larger impact on patient survival. The concept of relative survival per se appears to be suited to achieve this goal. One may even argue that exclusion of patients with a previous cancer may not be indicated at all in analyses of relative survival. In relative survival analysis, people with other diseases, including previous cancers, are generally included in estimates of expected survival (the denominator of relative survival estimates), and excluding them in estimating observed survival (the numerator) may actually bias estimates of relative survival.
The proportions of patients with no prior cancer diagnosed in 1993–1997, more than 40 years after initiation of cancer registration in Finland, probably come closer to the true proportions of such patients than the proportions observed in the earlier calendar periods. The latter may have missed cancers diagnosed prior to 1953. Although this possibility also exists for cancer patients diagnosed in 1993–1997, it is likely to play a minor role, as it would mostly arise from nonfatal cancers having occurred at relatively young ages: the median age at diagnosis of cancer patients in 1993–1997 ranged from 51 to 73 years for the cancers studied, i.e., cancers diagnosed more than 40 years earlier would mostly have occurred below 30 years of age. The number of patients diagnosed with cancer below age 30 before 1953 and surviving for more than 40 years is likely to be very small. Among 33,917 patients diagnosed with any cancer at any age in Finland in 1953–1957, only 1,317 (3.9%) were below 30 years of age at diagnosis, and of these, only 15 (1.1%) were still alive after 40 years. Other missed previous cancers may result from immigration or imperfect incompleteness of registration, both of which are likely to play only a relatively minor role in Finland.
Since most cancer registries have much shorter “running times” than the Finnish Cancer Registry, exclusion of cancer patients with a known previous cancer diagnosis is likely to be far from complete in analyses aiming for such exclusions. Our analysis suggests that exclusion of patients with a prior cancer diagnosis, who tend to have a somewhat worse prognosis, typically leads to higher estimates of survival. Even though the differences are quite limited in most cases, various degrees of completeness of such exclusions may provide a source of bias in comparative analyses of cancer survival based on cancer registries with different running times. Comparisons of cancer survival over time may likewise be biased even if based on a single registry, as registry running time necessarily also increases over time.
The patterns observed in our study, along with the variation of the proportion of second cancers with age of the registries observed in international comparative survival studies,1, 2 strongly suggest, that classification of a cancer as second cancer in population-based cancer registries is mostly based on existence of a previous registration with another cancer. Other potential sources of identification of second cancers, such as explicit notification of previous cancers by clinicians on registry records, which would be expected to be independent of registry running time, may also play a role in some registries. The importance of such potential sources may again vary between registries. It may be low in many registries, particularly those that are heavily relying on pathology reports as a main source of notification. Nevertheless, variation in the importance of such sources between registries may provide yet another source of heterogeneity in international comparisons of survival focusing on first cancers only. Furthermore, criteria for the definition of multiple tumours may not be homogeneous across registries. To evaluate potential implications of such inter-registry variation, it might be useful to carry out similar analyses in other population-based cancer registries.
One might argue that the impact of excluding patients with previous cancers was overall small in our empirical evaluations and played a major role only after a very long running time of the cancer registry and in the older age groups. Furthermore, given the decrease in 5-year survival with age for most forms of cancer, the lower survival of patients with previous cancers is partly because of the fact that these patients are on average somewhat older. Exclusion of these patients therefore leads to some shift of the age distribution towards younger ages. To the extent that the change of survival estimates by excluding patients with first cancers is due to this age shift, potential bias in comparison of survival between cancer registries with different running time can be overcome by age adjustment, which is common practice in international comparisons of cancer survival. However, as shown in our analysis, excluding patients with a first cancer diagnosis may lead to nonnegligible changes in survival estimates even within age groups commonly used for age adjustment, particularly the older age groups. As a result, the impact of excluding varying proportions of patients with previous cancer is not entirely overcome by age adjustment. But even if changes in survival estimates were entirely negligible after age adjustment, exclusion of patients with a previous cancer diagnosis was unnecessary at best, and it might even be potentially harmful by unnecessarily reducing precision of survival estimates.
In conclusion, our analyses suggest that patients with a prior cancer diagnosis should usually not be excluded from comparative analyses of cancer survival between cancer registries or over time. If exclusion does not or only slightly change survival estimates, it is unnecessary and only reduces precision. To the extent that there are relevant changes in survival estimates even after age adjustment, variation in proportions of excluded patients between registries for reasons such as different registry running times may introduce or increase bias in international comparisons of cancer survival.
The work of Dr. Hermann Brenner was partly supported by a grant from the German Cancer Foundation (Deutsche Krebshilfe, Project No. 70-3166-Br 5). Dr. Timo Hakulinen's work was supported by grants from the Academy of Finland and the Cancer Society of Finland.