Impact of Including Second and Later Cancers in Cause-Specific Survival Estimates Using Population-based Registry Data

Abstract

Background:

Second or later primary cancers account for about 20% of incident U.S. cases. Currently, cause-specific survival (CSS) analyses exclude these cancers because the cause of death (COD) classification algorithm was available only for first cancers. We added rules for later cancers to the SEER cause-specific death classification algorithm and evaluated CSS to include people with prior tumors.

Methods:

We constructed two cohorts: 1) First ever primary includes patients whose first cancer was diagnosed in 2000–2016; 2) Earliest matching primary includes patients with any cancer matching the selection criteria irrespective of whether it was the first or a later cancer diagnosed in 2000–2016. We compared the cohorts’ CSS estimates using follow-up through December 31, 2017. The new rules were used in the second cohort for patients whose first cancers in 2000–2016 are their second or later cancer.

Results:

Overall, there were no statistically significant differences in CSS estimates between the two cohorts. Estimates were similar by age, stage, race, and time since diagnosis, except for patients with leukemia and aged 65–74 years (3.4 percentage point absolute difference).

Conclusion:

The absolute difference in CSS estimates for first cancer ever vs. earliest of any cancers in the period was small for most cancer types. As the number of newly diagnosed patients with prior cancers increases, the algorithm will make CSS more inclusive and enable estimating survival for a group of cancer patients for whom life tables are not available or life tables are available but do not capture other-cause mortality appropriately.

Precis:

Our improved algorithm will allow researchers to estimate cause-specific survival in patients with multiple cancers, and it will allow this framework to be more inclusive and more reflective of cancer survival for all recently diagnosed patients. The use of the cause-specific survival framework may be particularly useful in studies of oncology outcomes, or when life tables do not adequately represent the patients' background mortality, or in special populations for which life tables are not available.

Introduction

To ensure consistency over time as well as to remove the effect a prior diagnosis of cancer may have on survival, the National Cancer Institute’s (NCI) Surveillance, Epidemiology, and End Results (SEER) Program has been limiting its reporting of cancer survival statistics to the first cancer ever diagnosed in a patient’s lifetime. However, as the proportion of people diagnosed with a prior cancer has increased due to the aging of populations, improvements in survival, and longer running time of cancer registries,^{1, 2} there is an interest in including people diagnosed with a second or later primary cancer in surveillance of cancer survival. For example, early well-known studies for cancer survival surveillance, such as EUROCARE and CONCORD, restricted analyses to patients with a first cancer diagnosis only.^{3, 4} However, currently these as well as reports published by the North American Association of Central Cancer Registries⁵ and Statistics Canada⁶ have been including patients diagnosed with prior cancers. Depending on the cancer site and age at diagnosis, later primaries may now range between 4% and 37% of all cancers diagnosed in the United States (U.S.).⁷

Cancer survival in the absence of other causes of death, also known as net cancer survival, is the preferred method for reporting cancer survival in population-based studies.⁸ It has been used in the previously cited studies and reports. It can be estimated using either the relative survival (RS) or the cause-specific survival (CSS) frameworks. In brief, RS is defined as the ratio of the observed survival (i.e., the likelihood of surviving all causes of death) in a cohort of cancer patients to the expected survival in a comparable population considered to be free of cancer and usually matched for age, sex, race, and calendar year.^{9, 10} In contrast, the cause-specific framework uses cause of death information and standard survival methods to estimate net survival.¹¹ Here, deaths due to the disease being studied are treated as events, and deaths from other causes are treated as censored observations. Because not all registries collect cause of death, RS has been the preferred framework internationally.

While the methods available in the SEER Program allow estimation of RS in patients with a previous cancer diagnosis, the same does not apply to CSS calculated using the SEER cause-specific death classification variable, an algorithm developed by Howlader et al to accurately capture all causes of death related to a defined cancer.¹² This constraint of the latter approach is because the algorithm was developed for 1) cancer patients diagnosed with one primary only; or for 2) cancer patients with more than one cancer diagnosis, but with analysis limited to the first cancer diagnosed. Rules to classify cause of death differ depending on whether the patient had only one tumor or it was the first of multiple (see Methods).

Several studies were published since SEER implemented its cause-of-death classification variable,^13-16 which approximately reflects the extent to which this variable is being used. When the cause of death is accurately captured, the CSS approach is an important alternative to RS—namely when the cohort under study is not a random sample of the general population,¹⁷ as in clinical studies, or when life tables do not reflect the life expectancy of the patient group. The latter may be the case for patients diagnosed with screen-detectable cancers and patients who have been heavily exposed to specific risk factors, such as smoking or certain types of infectious agents.¹⁸

Prior analyses have focused on estimating the risk and causes of developing second or later primaries relative to the general population risk,^19-23 on the number of persons alive (prevalence) diagnosed with multiple cancers,^{7, 24} or on the survival of these patients estimated through a relative survival framework.^{1, 6, 25-27} Our study aims to comprehensively analyze the impact of including patients with a previous cancer diagnosis in CSS calculations and to make this framework more reflective of cancer survival for all recently diagnosed patients. To do this, we improve the SEER cause-specific death classification variable by extending cause of death classification rules to second and later cancers. We use SEER data to evaluate our improved algorithm and compare CSS and RS using two cohorts of cancer patients.

Methods

Cause of death classification for second and later cancers

We improved the SEER cause-specific death classification algorithm to include rules for second and higher-order primaries. The rules are the same as those used to classify the cause of death in patients with first of multiple cancers. For this group of patients, a death is classified as being related to the specific cancer if it is a cancer attributed to the same cancer site, a cancer death that is attributed to the organ system of the study site, or a cancer death that is attributed to multiple cancers with unknown primary.¹² Deaths from cancer at a selected site that are coded as being attributed to noncancer diseases that are related to the site of first cancer diagnosis are also classified as deaths associated with the first cancer. Finally, because certain cancer types, such as Kaposi sarcoma and non-Hodgkin lymphoma, are more likely to occur in people who are infected with HIV, deaths attributed to AIDS and cancer or to HIV-related causes are also considered in the SEER cause-specific death classification variable. Deaths from all other malignant tumors are not classified as site-specific cancer deaths and, therefore, are considered to be censored events in cause-specific survival calculations. The improved SEER cause-specific death classification variable is now available in the November 2019 and later submissions of SEER data and can be used to generate cause-specific survival estimates in cancer patients diagnosed with one or more primary cancers.

Application to SEER data

We used data from the November 2019 Submission SEER 18 Registries Database for patients diagnosed with a malignant tumor from January 1, 2000, through December 31, 2016, with follow-up through December 31, 2017.²⁸ These registries cover approximately 28% of the U.S. population. Cancer sites selected were based on SEER Site Recode²⁹ and represent the most common malignancies diagnosed in men and women: esophagus, stomach, colon and rectum, liver, pancreas, lung, bone, melanoma, female breast, cervix uteri, ovary, prostate, brain, thyroid, lymphoma, and leukemia. We stratified cancer patients by age group (15–44, 45–54, 55–64, 65–74, 75+, and 15+), summary stage (localized, regional, distant), race/ethnicity [Non-Hispanic white (NHW), non-Hispanic Black (NHB), non-Hispanic American Indian/Alaska Native (NHAIAN), non-Hispanic Asian or Pacific Islander (NHAPI), Hispanic], and time since diagnosis (1-, 2-, 3-, 4-, 5-, 10-year survival). We excluded cases first diagnosed at autopsy, cases for which the death certificate was the only source of the cancer diagnosis information, cases with unknown stage, and cases that were alive but with no survival time. The latter represent cases with less than one month of follow-up time and lacking a date of death.³⁰ These cases are excluded from SEER survival calculations by default. We also excluded cases with missing or unknown cause of death.

We constructed two cohorts of patients: a cohort including cancer patients diagnosed in 2000–2016 with cancers with sequence number 0 or 1 (“first ever primary in the time period”); and a similar cohort but with cancers with any sequence number (“earliest matching primary in the time period”). The sequence number indicates the sequence of all reportable neoplasms over a person’s lifetime.³¹ Sequence number 0 indicates that the person has had only one in situ or one malignant neoplasm (as defined by the Federal reportable list); sequence number 1 indicates the first of two or more reportable neoplasms; sequence number 2 indicates the second of two or more reportable neoplasms; and so on.

For the second cohort, we allowed multiple primary cancers to be included for each patient, but we limited our analysis to one record per patient per statistic. For example, in a survival analysis covering the period 2000–2016 and where statistics are stratified by cancer site (Fig. 1A), a patient with a breast cancer primary in 2005 and a lung cancer primary in 2008 would contribute with the former to the “All Sites” and “Breast” statistics, and with the latter to the “Lung” statistics only. Likewise, if the statistics are stratified by summary stage (Fig. 1B), a patient diagnosed with a localized colon cancer primary in 2009 and a regional colon cancer primary in 2012 would contribute with the former to the “Localized” statistics (and to the “All Stages” statistics if this category is used), and with the latter to the “Regional” statistics only.

Figure 1.

Two hypothetical scenarios for a survival analysis covering the period 2000-2016.

We calculated cause-specific survival estimates by the actuarial method and compared them for the two cohorts mentioned above. The new rules of the improved SEER cause-specific death classification were used in the second cohort for patients whose first cancers are sequence number 2 or later in 2000–2016. We interpreted survival differences between cohorts of greater than a three percentage point absolute difference and no overlap between confidence intervals as significant. To account for differences in the age distribution of each cancer site, we adjusted survival estimates for age (except when stratified by age group) using the International Cancer Survival Standards (ICSS) and age groups 15–44, 45–54, 55–64, 65–74, and 75+.³² ICSS age standard 1 (sites that have increasing incidence with age) was used for esophagus, stomach, colon and rectum, liver, pancreas, lung, bone, breast, ovary, lymphoma, leukemia, and all sites combined (ICSS age standard 1 was also used for prostate cancer but with age groups 15–54, 55–64, 65–74, 75–84, and 85+); ICSS age standard 2 (sites that have relatively consistent incidence by age) was used for melanoma, cervix uteri, brain, and thyroid. We performed a sensitivity analysis using three different periods: 2000–2005 with follow-up through December 2006; 2006–2011 with follow-up through December 2012; 2012–2016 with follow-up through December 2017 (Suppl. Tables 6-9).

We additionally compared CSS estimates with RS estimates. We calculated RS estimates for each of the two cohorts mentioned above by the actuarial method—as the ratio of observed (all-cause) survival to expected survival.⁹ Expected survival can be calculated using different methods which vary with respect to the definition of the matching group. We used the Ederer II method based on life expectancy tables that match the cohort of cancer patients by age, year, sex, race, ethnicity, and county-level SES index.^{33, 34} This method has been shown to align well with the concept of net cancer survival, provided estimates are age-specific or age standardized.³⁵

We also calculated CSS estimates using two different multiple primary rules to understand the impact of different definitions of multiple primaries on survival estimates. SEER rules consider the timing of diagnosis and allow for the counting of new cases at different subsites of the same organ (e.g., colon) or on the opposite side (for paired organs such as the breast);³⁶ rules jointly developed by the International Association of Cancer Registries (IACR) and International Agency for Research on Cancer (IARC) are commonly utilized in international comparison of cancer survival and permit only one cancer per body site in a patient’s lifetime (unless of different histological types).³⁷

All statistical analyses were performed using NCI’s SEER*Stat software version 8.3.7.³⁸

Results

The selected cancer sites accounted for more than 80% of all cancers diagnosed in 2000 to 2016 in the SEER 18 registries catchment area. Overall, the proportion of multiple tumors (i.e., second and higher-order primaries) ranged from 7.2% for cervix uteri to 23.8% for melanoma (Table 1), and this proportion increased with advancing age at diagnosis, irrespective of the cancer site (Suppl. Table 1). Overall, there were no statistically significant differences between including vs. excluding patients with a previous cancer diagnosis in cause-specific survival estimates (Table 2). With very few exceptions, absolute differences were negligible across most age groups (Suppl. Table 2), stages (Table 3), races/ethnicities (Suppl. Table 3), and times since diagnosis (Suppl. Table 4). They were also negligible when using IARC/IACR rules for the definition of multiple primaries (Table 4).

Table 1.

Percentage distribution of new cases by cohort and cancer site. SEER-18, 2000–2016^a

	First Ever Primaryb		Second and later cancersc		Earliest Matching Primaryd
	N	%	N	%	N	%
Esophagus	51,493	79.4	13,352	20.6	64,845	100.0
Stomach	89,209	81.3	20,485	18.7	109,694	100.0
Colon and Rectum	525,717	80.5	127,177	19.5	652,894	100.0
Liver	92,632	87.7	13,010	12.3	105,642	100.0
Pancreas	144,866	81.4	33,054	18.6	177,920	100.0
Lung	668,016	76.9	200,409	23.1	868,425	100.0
Bone	11,855	87.8	1,650	12.2	13,505	100.0
Melanoma	238,418	76.2	74,405	23.8	312,823	100.0
Breast	819,687	82.3	176,351	17.7	996,038	100.0
Cervix Uteri	55,420	92.8	4,281	7.2	59,701	100.0
Ovary	84,523	85.1	14,742	14.9	99,265	100.0
Prostate	862,835	90.9	86,341	9.1	949,176	100.0
Brain	77,732	88.0	10,603	12.0	88,335	100.0
Thyroid	152,722	87.4	21,978	12.6	174,700	100.0
Lymphoma	266,485	82.1	58,027	17.9	324,512	100.0
Leukemia	153,615	80.1	38,253	19.9	191,868	100.0
All Sites	5,518,634	82.0	1,208,012	18.0	6,726,646	100.0

^aAll values are for both sexes except those for breast (women only) and sex-specific cancers

^bCorresponds to cohort 1 in our study and sequence # 0 and 1 in SEER*Stat (current default)

^cCorresponds to first matching primary (if second or later)

^dCorresponds to cohort 2 in our study and sequence # 0 to 59 in SEER*Stat

Table 2.

Five-year, age-standardized, cause-specific survival (%) and relative survival (%) for first ever primary vs. earliest matching primary by cancer site. SEER-18, 2000–2016^a

Cancer Site	Cause-specific Survival					Relative Survival
	First Ever Primaryb		Earliest Matching Primaryc		Abs. Dif. (%)	First Ever Primaryb		Earliest Matching Primaryc		Abs. Dif. (%)
	N	Survival (95% CI)	N	Survival (95% CI)		N	Survival (95% CI)	N	Survival (95% CI)
Esophagus	50,279	20.1 (19.7 to 20.5)	63,156	20.7 (20.3 to 21.0)	−0.6	50,694	18.0 (17.6 to 18.4)	63,638	17.7 (17.4 to 18.1)	0.3
Stomach	86,541	31.4 (31.1 to 31.8)	105,812	32.2 (31.8 to 32.5)	−0.8	87,802	29.4 (29.0 to 29.7)	107,240	28.9 (28.6 to 29.3)	0.5
Colon and Rectum	515,015	65.0 (64.9 to 65.2)	614,816	65.3 (65.2 to 65.5)	−0.3	518,910	64.7 (64.5 to 64.9)	619,356	63.8 (63.7 to 64.0)	0.9
Liver	86,350	19.7 (19.4 to 20.0)	98,638	20.5 (20.2 to 20.8)	−0.8	87,723	16.4 (16.1 to 16.7)	100,144	16.6 (16.3 to 16.9)	−0.2
Pancreas	138,406	9.6 (9.4 to 9.8)	170,448	10.2 (10.0 to 10.4)	−0.6	139,684	9.0 (8.9 to 9.2)	171,936	9.3 (9.1 to 9.5)	−0.3
Lung	645,286	20.6 (20.4 to 20.7)	819,066	22.4 (22.3 to 22.5)	−1.8	650,391	18.4 (18.3 to 18.6)	825,127	19.3 (19.2 to 19.4)	−0.9
Melanoma	234,701	90.4 (90.3 to 90.6)	289,016	90.5 (90.4 to 90.6)	−0.1	235,592	92.0 (91.8 to 92.2)	290,146	91.2 (91.0 to 91.3)	0.8
Breast	810,202	87.1 (87.0 to 87.1)	938,510	87.1 (87.0 to 87.2)	0.0	814,107	89.3 (89.1 to 89.4)	943,078	88.6 (88.4 to 88.7)	0.7
Cervix Uteri	54,407	65.4 (64.9 to 65.9)	58,422	65.4 (65.0 to 65.9)	0.0	54,877	62.5 (62.0 to 63.0)	58,928	61.9 (61.4 to 62.4)	0.6
Ovary	82,001	40.9 (40.5 to 41.2)	96,245	41.7 (41.4 to 42.0)	−0.8	82,579	40.9 (40.5 to 41.3)	96,899	41.1 (40.7 to 41.5)	−0.2
Prostate	846,398	92.2 (92.2 to 92.3)	931,249	92.2 (92.2 to 92.3)	0.0	852,351	97.4 (97.3 to 97.5)	937,752	96.2 (96.1 to 96.3)	1.2
Brain	66,839	27.8 (27.5 to 28.1)	76,691	27.6 (27.3 to 27.9)	0.2	67,525	26.7 (26.4 to 27.0)	77,439	26.4 (26.1 to 26.7)	0.3
Thyroid	150,593	94.5 (94.4 to 94.7)	171,238	94.7 (94.6 to 94.8)	−0.2	150,995	96.2 (96.0 to 96.5)	171,722	94.9 (94.6 to 95.1)	1.3
Lymphoma	257,439	69.7 (69.5 to 69.9)	308,639	69.8 (69.6 to 70.0)	−0.1	259,222	68.6 (68.3 to 68.8)	310,735	67.6 (67.3 to 67.8)	1.0
Leukemia	135,165	59.5 (59.2 to 59.8)	171,063	57.1 (56.9 to 57.4)	2.4	136,153	56.9 (56.6 to 57.2)	172,269	53.7 (53.4 to 54.0)	3.2d
All Sites	5,350,148	65.0 (65.0 to 65.1)	5,928,693	65.0 (65.0 to 65.1)	0.0	5,387,992	64.6 (64.6 to 64.7)	5,970,181	64.0 (63.9 to 64.0)	0.6

^aAll estimates are for both sexes except those for breast (women only) and sex-specific cancers

^bCorresponds to cohort 1 in our study and sequence # 0 and 1 in SEER*Stat (current default)

^cCorresponds to cohort 2 in our study and sequence # 0 to 59 in SEER*Stat

^dSurvival differences of greater than 3 percentage points and no overlap between confidence intervals

CI, Confidence interval

Table 3.

Five-year, age-standardized, cause-specific survival (%) and relative survival (%) for first ever primary vs. earliest matching primary by cancer site and summary stage. SEER-18, 2000–2016^a

Cancer Site	Cause-specific Survival					Relative Survival
	First Ever Primaryb		Earliest Matching Primaryc		Abs. Dif. (%)	First Ever Primaryb		Earliest Matching Primaryc		Abs. Dif. (%)
	N	Survival (95% CI)	N	Survival (95% CI)		N	Survival (95% CI)	N	Survival (95% CI)
Esophagus
Localized	10,303	45.8 (44.7 to 46.9)	13,755	45.6 (44.6 to 46.6)	0.2	10,385	41.3 (40.1 to 42.4)	13,853	39.7 (38.7 to 40.7)	1.6
Regional	15,810	23.5 (22.7 to 24.3)	19,500	23.7 (23.0 to 24.4)	−0.2	15,931	21.3 (20.6 to 22.1)	19,639	21.1 (20.4 to 21.8)	0.2
Distant	18,373	4.6 (4.3 to 5.0)	22,106	4.8 (4.5 to 5.2)	−0.2	18,526	4.1 (3.7 to 4.5)	22,277	4.0 (3.7 to 4.4)	0.1
Stomach
Localized	22,435	70.5 (69.8 to 71.1)	28,664	70.4 (69.8 to 70.9)	0.1	22,670	67.1 (66.3 to 67.8)	28,962	65.0 (64.3 to 65.6)	2.1
Regional	24,899	30.5 (29.9 to 31.1)	29,813	30.6 (30.0 to 31.1)	−0.1	25,302	29.0 (28.3 to 29.6)	30,260	28.3 (27.8 to 28.9)	0.7
Distant	30,196	4.9 (4.6 to 5.2)	35,852	5.1 (4.9 to 5.4)	−0.2	30,700	4.5 (4.2 to 4.8)	36,408	4.4 (4.2 to 4.7)	0.1
Colon and Rectum
Localized	199,937	89.7 (89.5 to 89.8)	247,853	88.8 (88.7 to 88.9)	0.9	201,169	89.9 (89.7 to 90.1)	249,377	87.6 (87.4 to 87.8)	2.3
Regional	185,591	69.8 (69.6 to 70v)	223,163	69.5 (69.2 to 69.7)	0.3	186,886	70.2 (70.0 to 70.5)	224,708	68.9 (68.7 to 69.2)	1.3
Distant	105,315	12.9 (12.7 to 13.1)	124,065	13.2 (13.0 to 13.4)	−0.3	106,411	12.5 (12.3 to 12.8)	125,285	12.5 (12.2 to 12.7)	0.0
Liver
Localized	38,337	33.8 (33.2 to 34.4)	44,543	34.6 (34.1 to 35.2)	−0.8	38,915	28.8 (28.2 to 29.3)	45,191	28.7 (28.1 to 29.2)	0.1
Regional	22,701	11.5 (11.0 to 12.0)	25,634	11.9 (11.4 to 12.4)	−0.4	23,135	9.5 (9.0 to 10.0)	26,105	9.4 (9.0 to 9.9)	0.1
Distant	14,051	2.6 (2.2 to 2.9)	15,864	2.9 (2.5 to 3.2)	−0.3	14,296	2.0 (1.8 to 2.4)	16,125	2.2 (1.9 to 2.5)	−0.2
Pancreas
Localized	13,155	37.6 (36.7 to 38.5)	17,258	38.7 (37.9 to 39.5)	−1.1	13,265	36.1 (35.2 to 37.1)	17,395	35.9 (35.0 to 36.7)	0.2
Regional	39,006	12.5 (12.1 to 12.9)	48,560	13.0 (12.6 to 13.4)	−0.5	39,423	12.0 (11.6 to 12.4)	49,046	12.2 (11.8 to 12.5)	−0.2
Distant	72,611	3.1 (2.9 to 3.2)	87,807	3.3 (3.1 to 3.4)	−0.2	73,287	2.9 (2.7 to 3.0)	88,578	2.9 (2.8 to 3.1)	0.0
Lung
Localized	102,491	62.5 (62.2 to 62.8)	148,436	62.8 (62.5 to 63.1)	−0.3	103,269	57.5 (57.2 to 57.9)	149,478	56.2 (55.9 to 56.5)	1.3
Regional	144,220	30.7 (30.4 to 31.0)	188,618	31.7 (31.5 to 32.0)	−1.0	145,287	28.0 (27.7 to 28.3)	189,931	28.2 (27.9 to 28.4)	−0.2
Distant	362,631	5.4 (5.3 to 5.5)	449,215	5.9 (5.8 to 6.0)	−0.5	365,634	4.7 (4.6 to 4.8)	452,677	4.9 (4.8 to 5.0)	−0.2
Melanoma
Localized	196,019	96.1 (96.0 to 96.2)	242,386	95.9 (95.8 to 96.0)	0.2	196,664	98.3 (98.2 to 98.5)	243,221	97.2 (97.1 to 97.4)	1.1
Regional	20,280	66.1 (65.3 to 66.8)	25,813	66.1 (65.4 to 66.8)	0.0	20,404	64.9 (64.1 to 65.7)	25,963	64.1 (63.4 to 64.9)	0.8
Distant	9,078	23.6 (22.6 to 24.7)	12,266	24.5 (23.5 to 25.6)	−0.9	9,141	22.4 (21.3 to 23.5)	12,353	22.5 (21.5 to 23.5)	−0.1
Breast
Localized	497,146	95.5 (95.4 to 95.6)	594,899	95.0 (94.9 to 95.1)	0.5	499,048	98.8 (98.7 to 98.9)	597,259	97.8 (97.6 to 97.9)	1.0
Regional	254,291	82.0 (81.8 to 82.2)	291,198	81.7 (81.5 to 81.9)	0.3	255,728	82.7 (82.4 to 83.0)	292,863	81.7 (81.4 to 81.9)	1.0
Distant	43,802	25.3 (24.8 to 25.8)	51,497	25.8 (25.3 to 26.2)	−0.5	44,198	24.4 (23.9 to 24.9)	51,946	24.5 (24.0 to 24.9)	−0.1
Cervix Uteri
Localized	25,480	88.8 (88.2 to 89.4)	27,145	88.5 (87.9 to 89.1)	0.3	25,613	87.1 (86.2 to 87.9)	27,288	86.1 (85.2 to 86.8)	1.0
Regional	19,336	59.4 (58.6 to 60.1)	20,852	59.0 (58.3 to 59.8)	0.4	19,533	55.8 (55.0 to 56.6)	21,063	55.0 (54.2 to 55.8)	0.8
Distant	7,153	18.0 (17.0 to 19.0)	7,773	18.1 (17.2 to 19.1)	−0.1	7,253	16.3 (15.4 to 17.3)	7,880	16.3 (15.3 to 17.2)	0.0
Ovary
Localized	12,124	88.2 (87.2 to 89.1)	14,296	87.9 (87.0 to 88.7)	0.3	12,166	90.1 (88.6 to 91.4)	14,348	88.4 (87.1 to 89.6)	1.7
Regional	15,421	64.5 (63.6 to 65.4)	18,319	64.9 (64.1 to 65.8)	−0.4	15,502	66.3 (65.2 to 67.3)	18,416	65.8 (64.8 to 66.8)	0.5
Distant	49,177	26.9 (26.5 to 27.4)	57,468	27.4 (27.0 to 27.8)	−0.5	49,586	26.9 (26.5 to 27.3)	57,925	26.9 (26.5 to 27.3)	0.0
Prostate
Localized	665,913	96.6 (96.6 to 96.7)	733,972	96.5 (96.5 to 96.6)	0.1	670,324	100 (99.9 to 100)	738,814	99.9 (99.8 to 100)	0.1
Regional	103,559	91.1 (90.8 to 91.4)	111,519	91.0 (90.7 to 91.3)	0.1	104,150	96.6 (96.0 to 97.1)	112,154	95.3 (94.8 to 95.8)	1.3
Distant	41,713	32.9 (32.3 to 33.5)	46,506	32.9 (32.4 to 33.5)	0.0	42,260	31.2 (30.6 to 31.8)	47,094	30.7 (30.1 to 31.3)	0.5
Brain
Localized	50,625	29.5 (29.1 to 29.9)	58,225	29.3 (29.0 to 29.7)	0.2	51,144	28.4 (28.1 to 28.8)	58,796	28.1 (27.8 to 28.4)	0.3
Regional	10,800	18.5 (17.8 to 19.3)	12,360	18.2 (17.5 to 19.0)	0.3	10,906	17.7 (16.9 to 18.4)	12,475	17.2 (16.5 to 18.0)	0.5
Distant	936	21.0 (18.6 to 23.6)	1106	20.4 (18.0 to 22.8)	0.6	950	19.5 (17.2 to 22.0)	1,121	18.5 (16.3 to 20.9)	1.0
Thyroid
Localized	101,468	99.1 (99.0 to 99.2)	115,165	98.9 (98.8 to 99.0)	0.2	101,650	99.8 (99.7 to 99.8)	115,386	99.4 (99.3 to 99.5)	0.4
Regional	39,780	94.2 (93.8 to 94.5)	45,115	94.2 (93.9 to 94.5)	0.0	39,891	95.5 (94.9 to 96.0)	45,253	93.9 (93.4 to 94.4)	1.6
Distant	6,231	61.7 (60.6 to 62.9)	7,501	62.4 (61.3 to 63.5)	−0.7	6,320	60.1 (58.8 to 61.3)	7,601	59.6 (58.4 to 60.7)	0.5
Lymphoma
Localized	68,180	80.3 (80.0 to 80.7)	84,249	80.2 (79.9 to 80.4)	0.1	68,608	80.7 (80.2 to 81.1)	84,779	79.2 (78.7 to 79.6)	1.5
Regional	47,031	72.3 (71.8 to 72.8)	55,196	72.3 (71.8 to 72.7)	0.0	47,265	71.3 (70.7 to 71.9)	55,474	70.2 (69.6 to 70.7)	1.1
Distant	122,204	62.7 (62.4 to 63.0)	147,418	62.5 (62.2 to 62.8)	0.2	123,120	60.9 (60.6 to 61.3)	148,483	59.9 (59.6 to 60.2)	1.0
All Sites
Localized	2,431,233	88.4 (88.3 to 88.4)	2,792,017	87.7 (87.6 to 87.7)	0.7	2,444,753	90.5 (90.4 to 90.5)	2,807,627	88.4 (88.3 to 88.5)	2.1
Regional	1,092,910	62.2 (62.1 to 62.3)	1,283,419	61.7 (61.7 to 61.8)	0.5	1,100,619	61.7 (61.6 to 61.8)	1,292,313	60.3 (60.2 to 60.4)	1.4
Distant	1,043,971	21.0 (21.0 to 21.1)	1,252,598	21.5 (21.5 to 21.6)	−0.5	1,053,677	19.7 (19.6 to 19.8)	1,263,557	19.6 (19.6 to 19.7)	0.1

^aAll estimates are for both sexes except those for breast (women only) and sex-specific cancers

^bCorresponds to cohort 1 in our study and sequence # 0 and 1 in SEER*Stat (current default)

^cCorresponds to cohort 2 in our study and sequence # 0 to 59 in SEER*Stat

CI, Confidence interval

Table 4.

Five-year, age-standardized, cause-specific survival (%) for earliest matching primary^a by cancer site and set of multiple primary (MP) rules. SEER-18, 2000–2016^b

	SEER MP Rules			IACR/IARC MP Rules			Abs. Dif. (%)
Cancer Site	N	Survival	MPs (%)	N	Survival	MPs (%)
Esophagus	63,156	20.7	20.4	63,123	20.7	20.3	0.0
Stomach	105,812	32.2	18.2	105,691	32.1	18.1	−0.1
Colon and Rectum	614,816	65.3	16.2	611,272	65.3	15.7	0.0
Liver	98,638	20.5	12.5	98,621	20.5	12.4	0.0
Pancreas	170,448	10.2	18.8	170,437	10.2	18.8	0.0
Lung	819,066	22.4	21.2	815,935	22.3	20.9	−0.1
Bone	10,892	59.6	13.5	10,851	59.6	13.2	0.0
Melanoma	289,016	90.5	18.8	283,849	90.5	17.3	0.0
Breast	938,510	87.1	13.7	911,291	87.2	11.1	0.1
Cervix Uteri	58,422	65.4	6.9	58,384	65.5	6.8	0.1
Ovary	96,245	41.7	14.8	96,218	41.7	14.8	0.0
Prostate	931,249	92.2	9.1	931,238	92.2	9.1	0.0
Brain	76,691	27.6	12.8	76,450	27.7	12.6	0.1
Thyroid	171,238	94.7	12.1	171,062	94.7	12.0	0.0
Lymphoma	308,639	69.8	16.6	304,681	70	15.5	0.2
Leukemia	171,063	57.1	21.0	165,918	58.1	18.5	1.0
All Sites	5,928,693	65	9.8	5,884,541	64.9	9.1	−0.1

^aCorresponds to cohort 2 in our study and sequence # 0 to 59 in SEER*Stat.

^bAll estimates are for both sexes except those for breast (women only) and sex-specific cancers.

SEER, Surveillance, Epidemiology, and End Results

IACR, International Association of Cancer Registries

IARC, International Agency for Research on Cancer

By age, differences in the CSS setting ranged from −4.4% (bone cancer in patients aged 75 years and older) to 2.8% (leukemia in patients aged 55–64 years) (Suppl. Table 2). However, most ranged between −1.0% and 1.0%, and some presented no differences at all (e.g., breast, cervical, or prostate cancer in patients overall). Only patients diagnosed with leukemia and aged 65–74 years presented statistically significant differences between cohorts (3.4%). In the RS setting, differences ranged between −1.0% (lung cancer in patients aged 65–74 years and 75 and older) and 4.5% (leukemia in patients aged 65–74 years) (Suppl. Table 2). Only patients diagnosed with leukemia and aged 55–64, 65–74, and 15+ years presented statistically significant differences between cohorts (3.5%, 4.5%, 3.2%, respectively).

By stage, differences in the CSS setting ranged from −1.1% (patients with localized pancreas cancer) to 0.9% (patients with localized colorectal cancer) (Table 3). In the RS setting, differences ranged between −0.2% (patients with distant liver and lung cancer, and regional pancreas and lung cancer) and 2.3% (patients with localized colorectal cancer) (Table 3). There were no statistically significant differences by stage.

By race/ethnicity, differences in the CSS setting ranged from −2.6% (bone cancer in Hispanics) to 2.4% (leukemia in NHW and NHB) (Suppl. Table 3), although most ranged between −1.0% and 1.0%, and some presented no differences at all (e.g., prostate cancer in NHW and NHB). There were no statistically significant differences in the CSS setting. In the RS setting, differences ranged between −1.6% (cervical cancer in NHAIAN) and 4.8% (bone cancer in NHB), but most ranged between −1.0% and 1.0% (Suppl. Table 3). Only leukemia in NHW and NHB presented statistically significant differences between cohorts (3.3% and 3.1%, respectively).

By time since diagnosis, differences in the CSS setting ranged from −2.4% (patients with lung cancer at 1-year survival) to 2.5% (patients with leukemia at 2-year survival) (Suppl. Table 4). However, most ranged between −1.0% and 1.0%, and some presented no differences at all (e.g., patients with cervical and prostate cancer at 5- and 10-year survival). There were no statistically significant differences in the CSS setting. In the RS setting, differences ranged between −1.7% (patients with lung cancer at 1-year survival) and 3.3% (patients with leukemia at 4-year survival), but most ranged between −1.0% and 1.0% (Suppl. Table 4). Only patients with leukemia at 2-, 3-, 4-, and 5-year survival presented statistically significant differences between cohorts (3.1%, 3.2%, 3.3%, and 3.2%, respectively).

Discussion

In this study, we improved the SEER cause-specific death classification algorithm by extending rules classifying cause of death to second or later cancers. We compared CSS estimates for two cohorts of patients, the “first ever primary” cohort, which included only the first cancer diagnosed, and the “earliest matching primary” cohort, which included the first cancer matching the selection criteria irrespective of whether it was the first or a later cancer of the patient diagnosed in the SEER 18 registries catchment areas during 2000–2016. Our results show that, in general, the impact of including first cancer ever versus earliest matching cancer in cause-specific survival analyses is very small, with most differences not exceeding one percentage point, and results were no different when a more stringent set of rules for coding multiple primaries (IARC/IACR rules) was used. The largest differences were observed for leukemia, where CSS for first ever versus earliest matching was 59.5% and 57.1%, respectively, representing a 2.4% difference in survival. Except for higher CSS estimates in more recent periods, our sensitivity analyses did not change these results.

We also compared RS for the two cohorts, which showed that, for most cancers, the absolute difference in RS of including the first cancer only (current default in SEER*Stat) compared to earliest matching was also very small. Again the largest differences were observed for leukemia, where RS for first ever versus earliest matching was 56.9% and 53.7%, respectively, representing a 3.2% difference in survival. Larger differences were observed for analysis stratified by age (e.g., 3.5 and 4.5 percentage point absolute difference in age groups 55–64 and 65–74 years, respectively; Suppl. Table 2). In a supplementary analysis, we show that differences are attenuated when stratifying by leukemia subtype, except for other myeloid/monocytic leukemia (Suppl. Table 5).

Our results agree with studies that found a systematic decrease in the age-standardized RS when all individuals were included in the analysis (the earliest matching cohort).^{1, 26, 27} The lower survival for the earliest matching cohort may be explained by the higher likelihood of frailty among patients with prior cancers and prior treatments.^{39, 40} However, we did not find this systematic difference for CSS. In the CSS setting, more cancer sites presented higher survival estimates when the earliest matching tumor was included. Reasons for this discrepancy may be related to challenges in the classification of causes of death. It is possible that the cause of death is more commonly attributed to the first or prior cancer than to later cancers. In this case, the SEER cause-specific death classification for the second or later cancer would be a censoring event, and survival would be better.

Our rules to improve the classification of causes of death among people with multiple tumors have some limitations related to the challenges in identifying a single cause of death for this group. For patients diagnosed with two primary cancers located at the same site, the SEER cause-specific death classification codes both cancers as dead due to the cancer (i.e., we would not know which one actually contributed to the death). Another example is patients diagnosed with a cancer whose prognosis is generally good (e.g., breast) and, later on, with a poor prognosis cancer (e.g., lung), and whose death certificate states only the good prognosis cancer as the cause of death; this may be one reason for CSS estimates being greater when including all multiple primaries in the analysis, as found in our study. A third example is when a patient has three primaries (e.g., breast, lung, and colon) and the death certificate states a miscellaneous cancer as the cause of death. In this case, the cause of death in the SEER cause-specific death classification for each of those three cancer sites would be “Dead (attributable to this cancer diagnosis),” but only one of the three cancers was probably the cause of death. In all of these scenarios, only a thorough review of the clinical information may elucidate which cancer contributed to the patient's death.

Although the RS approach also has some limitations when including second and later tumors, as when life tables may not be representative of the other causes of mortality for patients with a prior cancer, international studies have been using the earliest matching approach to report and compare RS. The main reasons for using this approach are: 1) the varying ability of registries to identify first ever primary, such that this approach enables fairer comparisons; and 2) inclusion of people with multiple tumors, such that this approach better reflects the cancer survival. SEER still uses the first ever approach to report survival statistics.

When the goal is to compare survival between registries, or within the same registry over time, another source of variation is the capacity to identify cases diagnosed before the start of registration (prevalent cases) or cases that are extensions, recurrences, or metastases of a previous cancer.^{1, 26, 36} This effort is particularly challenging in recently or less-well established registries, which may not be able to distinguish between recurrences/disease progression from true subsequent primary cancers. When the registry cannot make that distinction, a non-negligible proportion of second and higher-order cancers may end up being included in survival calculations as if they were first cancers. Again, this would bias the survival downward, as second and higher-order cancers typically present lower survival than first cancers, not least when the RS framework is used to estimate net survival. On the other hand, if multiple cancers are to be systematically included in the analyses, this may increase the comparability of survival estimates between longstanding and recently established cancer registries.⁴¹

Strengths of this study include the quality of the registry data, the availability of numerous cases from a representative U.S. population, and the fact that it is the first to assess the impact of including patients with a known previous cancer in cause-specific survival estimates. Our study has limitations, chiefly the challenges in assigning a cause of death in patients diagnosed with multiple primary cancers. Despite these challenges, our algorithm was shown to be robust, as the differences in CSS for first ever primary versus earliest matching primary were small and similar to RS. We also showed that the impact of differences when stratified by stage decreased more for CSS than for RS. A second limitation is that our results are not directly generalizable to registries where the cause of death is not readily available or is inaccurate. Third, we did not provide a comparison between registries with different running times (e.g., Connecticut [1935] vs. Los Angeles [1988]). Longer running time increases the proportion of patients with multiple primary cancers.²⁶ For newer registries, the differences in survival estimates between first ever primary versus earliest matching primary would have been smaller. Finally, a bootstrap analysis might be applied to generate a single confidence interval to all absolute differences in CSS, for each cancer site. This would bring more assurance when interpreting differences between cohorts.

As the number of newly diagnosed patients who have had prior cancers increases, the improved algorithm will allow CSS to be more inclusive and more reflective of cancer survival for all recently diagnosed patients, and it will allow researchers to estimate CSS in patients with multiple cancers. The improved algorithm will also make CSS analyses more comparable with RS analyses that include all primaries and for registries that differ in their ability to identify tumors prior to their establishment. The CSS framework may be particularly useful in studies of oncology outcomes, or when life tables do not adequately represent the patients' background mortality,¹⁸ or in special populations for which life tables are not available (e.g., Chinese Asians). In line with other studies, we recommend not excluding patients with a known prior cancer, given that some level of bias may exist when restricting the analysis to first cancers only. Including subsequent primaries will improve the comparability of data from different registries and over time, and it will improve precision in the analyses being conducted.