Adjusting survival estimates for differential loss to follow-up by race and ethnicity: a SEER analysis

Abstract

Background

Apparent survival advantages for Hispanic and non-Hispanic Asian and Pacific Islander individuals, often labeled the “Hispanic paradox” or “healthy immigrant effect,” have been reported for multiple cancers compared with non-Hispanic White and non-Hispanic Black persons. Higher proportions of loss to follow-up among Hispanic and non-Hispanic Asian and Pacific Islander persons overestimate survival because of informative censoring. We developed a sensitivity analysis–based method to address potential impacts of differential loss to follow-up.

Methods

We estimated 5-year age-standardized cancer-specific survival for non-Hispanic White (reference), non-Hispanic Black, Hispanic, non-Hispanic Asian and Pacific Islander, Hispanic, and non-Hispanic American Indian or Alaska Native persons diagnosed between 2005 and 2019 in 22 Surveillance, Epidemiology, and End Results Program registries. Persons alive with less than 5 years or maximum possible follow-up were classified as lost to follow-up. For each cancer stage–race and ethnicity combination, for cases lost to follow-up exceeding levels for non-Hispanic White persons, we modeled 2 plausible scenarios relative to survival of non-Hispanic White patients: (1) “lower” survival and (2) “higher” survival. Across 1000 simulations, we constructed a sensitivity interval spanning the average lower 95% confidence bound of “lower” survival to the average upper 95% confidence bound of “higher” survival.

Results

Loss to follow-up was higher for Hispanic and non-Hispanic Asian and Pacific Islander persons and Hispanic individuals across cancer sites (range = 2.2%-20.7%) than for non-Hispanic White persons (range = 0.4%-5.4%). Sensitivity intervals were consistently wider than standard 95% CIs. After accounting for differential loss to follow-up, previously observed equivalent or higher survival disappeared for Hispanic and non-Hispanic Asian and Pacific Islander persons (eg, prostate cancer) and Hispanic persons (eg, lung cancer).

Conclusions

Ignoring differential loss to follow-up overstates survival advantages among racial and ethnic minorities. Sensitivity analysis revealed that reported survival advantages are artifacts of higher loss to follow-up, underscoring the need to address this censoring bias in population-based survival studies.

Introduction

Cancer survival differences across racial and ethnic groups have become a major focus of modern epidemiologic research, reflecting broader inequalities in healthcare access, socioeconomic context, and cultural perceptions of health. These differences are shaped by cancer type, stage at diagnosis, screening practices, and demographic factors. In surveillance systems based on hospital and population registries, additional artifacts such as loss to follow-up influence measured survival. Loss to follow-up disproportionately affects non-White and foreign-born populations because vital status ascertainment for these groups is often hindered by multiple factors. Discrepancies in personal identifiers in record linkages (eg, Social Security numbers, variability in surname structures) can lead to difficulties in ascertaining deaths and obtaining dates of last contact among living patients. The “Salmon bias,”¹ whereby patients return to their home countries near the end of life, can result in uncaptured deaths in US mortality files.

Although racial and ethnic groups differ in loss to follow-up, these differences alone do not bias survival estimates. Survival analysis remains valid under unequal censoring rates as long as the censoring mechanism is unrelated to prognosis. The critical issue, first demonstrated by Pinheiro et al.² in an analysis of survival data from the National Cancer Institute’s (NCI’s) Surveillance, Epidemiology, and End Results (SEER) Program, is that loss to follow-up among Hispanic and Asian patients is prognostically informative. Patients in these groups who become lost to follow-up are disproportionately diagnosed at later stages and have higher short-term mortality risk. Therefore, it is the association between loss to follow-up and advanced disease stage that violates the assumption of noninformative censoring and leads to inflated survival estimates under the “reported alive” framework.

In cancer registry data, reported alive refers to patients for whom there is documented evidence of being alive after diagnosis, based on a valid last-known-alive contact date. This information is obtained through multiple follow-up sources, including death certificate linkages, Social Security Administration data linkages, hospital and physician records, and other administrative databases. Importantly, SEER follow-up standards require that approximately 95% of all patients with cancer have complete follow-up, defined as either a recorded date of death or a verified last-alive contact date, which ensures that survival estimates are based on actively ascertained follow-up rather than assumptions of continued survival.

Both Pinheiro et al.² and Qiao et al.³ documented substantially higher loss to follow-up among Hispanic, Asian, and foreign-born patients compared with White and US-born patients, respectively. Pinheiro et al.² further demonstrated that loss to follow-up was concentrated among patients diagnosed at later disease stages, providing evidence that censoring is related to prognosis. Despite these findings, current survival estimates in SEER*Stat do not incorporate any adjustment for informative censoring. Thus, widely cited survival advantages among Hispanic and Asian populations, often attributed to the “healthy immigrant effect” or the “Hispanic paradox,” may partly reflect inflation caused by loss to follow-up rather than true prognostic differences. Because SEER survival metrics can shape federal health policy priorities, unadjusted loss to follow-up has substantial implications for public health.

Our study addressed this problem by developing a sensitivity analysis framework that generates alternative survival scenarios by stage at diagnosis to more accurately quantify racial and ethnic differences. Our approach assumes similar proportions lost to follow-up across racial and ethnic groups and considers stage distribution in the prediction of outcomes among patients lost to follow-up. By removing the source of informative censoring, the framework produces improved survival estimates and clarifies the extent to which differential loss to follow-up contributes to observed survival differences in an increasingly diverse US population.

Methods

Our analysis considered non-Hispanic White, non-Hispanic Black, non-Hispanic Asian and Pacific Islander, non-Hispanic American Indian or Alaska Native, and Hispanic individuals of all races. Individuals with unknown or another race were excluded from the analysis.

Survival data

Cause-specific survival data from 22 SEER registries (excluding Arizona Indian and Cherokee Nation persons) for patients diagnosed between 2005 and 2019 with follow-up through December 2020 were extracted using SEER*Stat, version 8.4.3.^4,5 As the focus of this study was methodological, use of SEER data with more recent follow-up was not a main priority. The major cancer sites examined were lung, colon and rectum, pancreas, prostate (male only), and breast (female only). Additional cancer sites considered included anus, urinary bladder, brain and other nervous system, Hodgkin lymphoma, kidney and renal pelvis, leukemia, liver and intrahepatic bile duct, melanoma of the skin, myeloma, non-Hodgkin lymphoma, oral cavity and pharynx, stomach, and thyroid.⁶ Sex-specific cancers were examined for male (testis) and female (triple-negative breast, cervix uteri, corpus and uterus, and ovary) patients. All cancer sites were defined using the site recode International Classification of Diseases for Oncology, 3rd Edition/World Health Organization 2008 variable.⁷ Triple-negative breast cancer was identified using hormone receptor and ERBB2 (formerly HER2) status information (available for cases diagnosed from 2010 onward).

Five-year cause-specific survival was calculated for each cancer site by combined summary stage (localized, regional, distant, unknown/unstaged, all stages; available for 2004 onward) as well as race and ethnicity for both sexes combined, unless the site was sex specific. Actuarial survival was calculated using monthly intervals and age standardized using the site-appropriate International Cancer Survival Standard.⁸ Standard 95% CIs were constructed.⁹ For any group with fewer than 25 cases across all ages, survival was suppressed for statistical precision.

Defining loss to follow-up

Cases lost to follow-up were identified using a definition based on Pinheiro et al.² Individuals diagnosed between 2005 and 2015 who survived at least 5 years after diagnosis per last-alive contact or had died at any point had complete follow-up. Individuals diagnosed between 2016 and 2019 who had not died had to have achieved the maximum possible follow-up through December 2020 to be considered complete. A living patient diagnosed in 2016 had to survive at least 4 years after diagnosis, a living patient diagnosed in 2017 had to survive at least 3 years after diagnosis, and so on. Cases that did not meet the definition of complete follow-up were considered lost to follow-up. The proportion lost to follow-up was calculated by cancer site, stage, and racial and ethnic group.

Sensitivity analysis to examine the impact of differential loss to follow-up

Non-Hispanic White was chosen as the reference group because these individuals had the lowest proportion lost to follow-up and made up the largest racial and ethnic group in the United States. For any group with a percentage lost to follow-up greater than that of non-Hispanic White, we focused on the excess cases lost to follow-up in that group. For example, if the non-Hispanic White group had 2% lost to follow-up and the non-Hispanic Black group had 4% lost to follow-up, we considered the additional 2% that were excess lost to follow-up among the non-Hispanic Black group, which comprises 50% (2%/4%) of all non-Hispanic Black patients lost to follow-up. Any subgroup with a percentage lost to follow-up less than or equal to that of the non-Hispanic White group was not included in the sensitivity analysis, and we assumed that the SEER*Stat survival was accurate, thereby assuming that loss to follow-up was noninformative for that subgroup.

We reassigned the excess cases lost to follow-up to plausible best-case and worst-case survival scenarios based on stage-specific non-Hispanic White survival to construct upper and lower bounds of cause-specific survival: (1) all excess cases lost to follow-up assumed to have lower survival (“lower bound”) and (2) all excess cases lost to follow-up assumed to have higher survival (“upper bound”). In both scenarios, the excess cases lost to follow-up would now be considered to have complete follow-up. A summary of the lower bound and upper bound definitions for each stage at diagnosis is provided in Table 1. As shown in Table 1, to construct the sensitivity interval bounds, typically 1 SE is added or subtracted to the stage-specific 5-year cause-specific survival estimate for non-Hispanic White patients. The choice of adding or subtracting 1 SE is arbitrary, but it represents a simple way to capture a large percentage of the distribution of potential survival estimates in either direction (ie, higher or lower survival relative to non-Hispanic White). As a sensitivity analysis, we also explored adding or subtracting 2 SEs in each scenario described in Table 1 for the major cancer sites (see supplementary material).

Table 1.

Summary of sensitivity analysis bounds, by stage at diagnosis, where non-Hispanic White group survival is used as reference survival for each stage.

		Stagea
Reference group	Bound for sensitivity analysis	Localized	Regional	Distant	Unknown	All stages combined
Non-Hispanic White	Lower bound	(5-y survival ‒ 1 SE) among excess loss to follow-up	(5-y survival ‒ 1 SE) among excess loss to follow-up	All excess lost to follow-up dead 1 y after date of last follow-up	(5-y survival ‒ 1 SE) among excess loss to follow-up	${\sum }_k{p}_k{S}_{k, \mathit{lower}}\left(5\right)$ $p_k=$ % in stage $k$ $S_{k,\mathit{lower}}\left(5\right)=$ 5-y age-standardized survival lower bound for stage$k$
	Upper bound	All excess loss to follow-up alive 5 y after date of last follow-up	(5-y survival + 1 SE) among excess loss to follow-up	(5-y survival + 1 SE) among excess loss to follow-up	(5-y survival + 1 SE) among excess loss to follow-up	${\sum }_k{p}_k{S}_{k, \mathit{upper}}\left(5\right)$ $p_k=$ % in stage$k$ $S_{k,\mathit{upper}}\left(5\right)=$ 5-y age-standardized survival upper bound for stage$k$

^a Five-year survival refers to the 5-year age-standardized cause-specific survival probability for the non-Hispanic White group for a particular disease stage.

Excess cases lost to follow-up were randomly selected among all cases lost to follow-up according to the probability $({\%\mathit{LFU}}_{\mathit{comp}}-{\%\mathit{LFU}}_{\mathit{NHW}})/{\%\mathit{LFU}}_{\mathit{comp}}$, where %LFU refers to the percentage lost to follow-up, the subscript $\mathit{comp}$ refers to the racial or ethnic group under comparison, and the subscript $\mathit{NHW}$ refers to the reference non-Hispanic White group. Random selection was performed to ensure that excess cases lost to follow-up were similar to nonexcess cases lost to follow-up in terms of patient characteristics. For each site–stage combination, 5-year age-standardized cause-specific survival estimates and 95% CIs under each scenario were recalculated 1000 times for the comparison group. Sensitivity intervals were constructed as lower 95% confidence bound for the “lower bound” scenario survival estimate and upper 95% confidence bound for the “upper bound” scenario survival estimate averaged across 1000 simulations. In the supplementary materials, a step-by-step calculation of the sensitivity interval is outlined using regional lung cancer data as an example.

The sensitivity intervals we constructed serve as an alternative to the CIs SEER*Stat normally reports. Due to the variability introduced by random sampling in the construction of our sensitivity intervals as well as the additional uncertainty arising from assumptions about excess cases lost to follow-up, we expected that they will be wider than the corresponding CIs. If the point estimate for the non-Hispanic White group is not included in either the confidence or sensitivity intervals for a particular minority group, this may indicate that the observed survival difference is in fact a true difference, even after adjusting for differential loss to follow-up. All analyses were conducted using R, version 4.2.2, software (R Foundation for Statistical Computing).

Results

We analyzed 2 495 730 cases across the 5 major cancer sites of lung (n = 529 991), colon and rectum (n = 425 431), pancreas (n = 127 979), prostate (n = 698 100), and female breast (n = 714 229) cancer. Cases were representative of all major racial and ethnic groups in the United States: non-Hispanic White (n = 1 732 383 [69.4%]), non-Hispanic Black (n = 300 786 [12.1%]), Hispanic and non-Hispanic Asian and Pacific Islander (n = 187 464 [7.5%]), Hispanic (n = 260 982 [10.5%]), and non-Hispanic American Indian or Alaska Native (n = 14 115 [0.6%]). Across all major cancer sites and stages, non-Hispanic White individuals, especially Hispanic and non-Hispanic Asian and Pacific Islander persons and Hispanic individuals, were more likely to be lost to follow-up at 5 years than non-Hispanic White patients (range across sites = 0.4%-2.6%): non-Hispanic Black (range = 0.4%-3.8%), Hispanic and non-Hispanic Asian and Pacific Islander (range = 2.2%-10.6%), Hispanic (range = 2.4%-13.8%), and non-Hispanic American Indian or Alaska Native (range = 0.0%-3.9%). Distributions of loss to follow-up by registry (Table S1) and by race and ethnicity within registry (Table S2) are provided in the supplementary materials.

Plots comparing 95% CIs and sensitivity intervals for lung, colon and rectum, pancreas, prostate, and female breast cancer by race and ethnicity as well as stage are provided in Figures S1-S5, respectively. Unknown/unstaged observations were excluded from plots due to small numbers. SEER*Stat survival estimates, with respective 95% CIs, and constructed sensitivity intervals by stage, race, and ethnicity are provided in Tables 2-6. Results for additional cancer sites are contained in Figures S6-S23 and Tables S3-S20. From Figures S1-S5, sensitivity intervals tended to be wider than the corresponding 95% CIs. Survival advantages suggested by 95% CIs often disappeared when sensitivity intervals were applied.

Table 2.

Comparison of 5-year age-adjusted cause-specific survival for lung cancer, by racial group and stage at diagnosis, for both sexes combined.

		5-y age-standardized cause-specific survival
Site	Statistic	Non-Hispanic White	Non-Hispanic Black	Hispanic and non-Hispanic Asian and Pacific Islander	Hispanic	Non-Hispanic American Indian or Alaska Native
Lung, all stages	No.a	399 700	56 491	36 730	34 101	2969
	Lost to follow-up, %	0.7	0.8	3.7	3.6	0.8
	SEER*Stat survival (95% CI)	24.6 (24.4 to 24.8)	20.6 (20.2 to 21.0)	26.9 (26.4 to 27.5)	24.0 (23.5 to 24.6)	21.3 (19.4 to 23.1)
	Sensitivity intervalb	N/Ac	19.7-21.1	25.2-27.4	22.2-24.4	17.8-24.2
Lung, localized	No.a	73 642	8384	5720	5353	466
	Lost to follow-up, %	1.4	1.6	4.6	4.6	1.7
	SEER*Stat survival (95% CI)	67.6 (67.2 to 68.0)	61.1 (59.8 to 62.3)	73.1 (71.8 to 74.4)	69.8 (68.4 to 71.2)	61.4 (56.1 to 66.8)
	Sensitivity intervalb	N/Ac	59.8-62.4	71.3-75.2	68.0-72.0	56.1-66.9
Lung, regional	No.a	88 400	12 468	6831	6618	631
	Lost to follow-up, %	0.7	0.9	4.3	3.8	0.2
	SEER*Stat Survival (95% CI)	35.4 (35.0 to 35.8)	31.1 (30.2 to 32.1)	42.4 (41.0 to 43.8)	37.5 (36.1 to 38.8)	31.6 (27.4 to 35.9)
	Sensitivity intervalb	N/Ac	30.1-32.1	40.1-42.9	35.4-38.1	—d
Lung, distant	No.a	218 370	33 450	22 461	20 008	1713
	Lost to follow-up, %	0.4	0.6	3.2	3.4	0.7
	SEER*Stat Survival (95% CI)	6.7 (6.6 to 6.9)	6.6 (6.2 to 6.9)	11.0 (10.4 to 11.5)	7.8 (7.3 to 8.3)	6.3 (4.5 to 8.1)
	Sensitivity intervalb	N/Ac	6.2-6.9	9.6-10.8	6.5-7.7	4.5-8.0
Lung, unknown	No.a	19 288	2189	1718	2122	159
	Lost to follow-up, %	0.7	0.9	4.5	2.9	1.3
	SEER*Stat Survival (95% CI)	17.0 (16.1 to 17.9)	15.8 (13.8 to 17.9)	21.5 (18.7 to 24.3)	17.1 (14.8 to 19.3)	20.1 (12.5 to 27.8)
	Sensitivity intervalb	N/Ac	13.7-17.8	17.3-22.7	14.1-18.5	11.5-27.3

Abbreviations: N/A = not applicable; SEER = Surveillance, Epidemiology, and End Results.

^aNumber diagnosed between 2005 and 2019 in SEER-22 registry areas.

^bAverage lower 95% CI bound for “lower bound” scenario survival estimate, average upper 95% CI bound for “upper bound” scenario survival estimate.

^cLost to follow-up adjustment not conducted for the reference group of US-born non-Hispanic White individuals.

^dLost to follow-up adjustment not conducted for subgroups with a percentage lost to follow-up less than or equal to that of the reference group (non-Hispanic White).

Table 3.

Comparison of 5-year age-adjusted cause-specific survival for colon and rectal cancer, by racial group and stage at diagnosis, for both sexes combined.

		5-y age-standardized cause-specific survival
Site	Statistic	Non-Hispanic White	Non-Hispanic Black	Hispanic and non-Hispanic Asian and Pacific Islander	Hispanic	Non-Hispanic American Indian or Alaska Native
Colon and rectum, all stages	No.a	282 127	49 184	38 213	52 600	3307
	Lost to follow-up, %	1.7	2.0	5.2	7.3	2.7
	SEER*Stat Survival (95% CI)	66.2 (66.0 to 66.4)	58.9 (58.4 to 59.4)	68.2 (67.6 to 68.7)	64.5 (64.0 to 64.9)	62.3 (60.4 to 64.2)
	Sensitivity intervalb	N/Ac	58.0-59.5	66.6-68.4	62.7-64.6	59.2-65.3
Colon and rectum, localized	No.a	109 123	18 266	14 427	19 280	1192
	Lost to follow-up, %	2.1	2.5	5.4	8.1	3.9
	SEER*Stat survival (95% CI)	90.3 (90.1 to 90.5)	87.0 (86.4 to 87.6)	91.4 (90.9 to 91.9)	89.3 (88.7 to 89.8)	89.6 (87.4 to 91.9)
	Sensitivity intervalb	N/Ac	86.4-87.6	90.7-92.1	88.6-90.2	87.3-92.0
Colon and rectum, regional	No.a	101 195	15 966	14 302	19 165	1182
	Lost to follow-up, %	1.7	2.2	5.7	7.7	2.6
	SEER*Stat survival (95% CI)	72.3 (72.0 to 72.6)	66.0 (65.1 to 66.9)	73.7 (72.9 to 74.5)	69.9 (69.1 to 70.7)	68.0 (64.9 to 71.2)
	Sensitivity intervalb	N/Ac	65.1-66.8	72.5-74.1	68.8-70.3	64.8-71.1
Colon and rectum, distant	No.a	59 170	12 431	7589	11 477	784
	Lost to follow-up, %	0.8	1.0	3.3	5.3	1.3
	SEER*Stat survival (95% CI)	14.8 (14.5 to 15.1)	10.2 (9.6 to 10.9)	15.9 (14.9 to 16.9)	15.0 (14.3 to 15.8)	13.7 (10.8 to 16.6)
	Sensitivity intervalb	N/Ac	9.5-10.8	14.1-16.4	12.9-15.2	10.7-16.5
Colon and rectum, unknown	No.a	12 639	2521	1895	2678	149
	Lost to follow-up, %	1.7	2.3	6.5	7.5	0.7
	SEER*Stat Survival (95% CI)	47.7 (46.6 to 48.7)	48.5 (46.3 to 50.7)	52.0 (49.4 to 54.6)	49.4 (47.2 to 51.6)	54.5 (46.1 to 62.9)
	Sensitivity intervalb	N/Ac	46.2-50.6	48.4-53.5	46.2-50.5	—d

Abbreviations: N/A = not applicable; SEER = Surveillance, Epidemiology, and End Results.

^aNumber diagnosed between 2005 and 2019 in SEER-22 registry areas.

^bAverage lower 95% CI bound for “lower bound” scenario survival estimate, average upper 95% CI bound for “upper bound” scenario survival estimate.

^cLost to follow-up adjustment not conducted for reference group of US-born non-Hispanic White individuals.

^dLost to follow-up adjustment not conducted for subgroups with a percentage lost to follow-up less than or equal to that of the reference group (non-Hispanic White).

Table 4.

Comparison of 5-year age-adjusted cause-specific survival for pancreatic cancer, by racial group and stage at diagnosis, for both sexes combined.

		5-y age-standardized cause-specific survival
Site	Statistic	Non-Hispanic White	Non-Hispanic Black	Hispanic and non-Hispanic Asian and Pacific Islander	Hispanic	Non-Hispanic American Indian or Alaska Native
Pancreas, all stages	No.a	87 078	14 683	10 277	15 154	787
	Lost to follow-up, %	0.6	1.2	2.8	3.9	0.8
	SEER*Stat Survival (95% CI)	11.8 (11.6 to 12.1)	10.7 (10.1 to 11.3)	14.1 (13.3 to 14.9)	12.2 (11.6 to 12.8)	9.2 (6.9 to 11.4)
	Sensitivity intervalb	N/Ac	9.3-11.5	11.9-15.1	10.5-13.0	4.3-12.8
Pancreas, localized	No.	9550	1636	1214	1717	89
	Lost to follow-up, %	1.4	2.3	4.1	5.9	3.4
	SEER*Stat Survival (95% CI)	43.2 (42.1 to 44.3)	35.6 (33.1 to 38.1)	52.2 (49.3 to 55.2)	44.7 (42.1 to 47.2)	32.8 (23.2 to 42.3)
	Sensitivity intervalb	N/Ac	33.0-38.6	48.6-56.0	41.4-48.9	23.3-47.2
Pancreas, regional	No.a	25 487	3974	3112	4195	195
	Lost to follow-up, %	0.7	1.2	3.1	4.4	0.0
	SEER*Stat Survival (95% CI)	14.7 (14.2 to 15.3)	14.4 (13.1 to 15.6)	15.9 (14.3 to 17.4)	14.4 (13.2 to 15.6)	6.8 (2.6 to 11.0)
	Sensitivity intervalb	N/Ac	13.0-15.5	13.9-17.0	12.6-15.0	—d
Pancreas, distant	No.a	44 904	8061	5141	7950	434
	Lost to follow-up, %	0.4	0.9	2.4	3.4	0.5
	SEER*Stat Survival (95% CI)	3.7 (3.5 to 4.0)	3.1 (2.6 to 3.6)	4.2 (3.5 to 4.9)	3.6 (3.1 to 4.1)	3.3 (1.4 to 5.2)
	Sensitivity intervalb	N/Ac	2.5-3.5	2.9-4.4	2.7-3.8	1.4-5.2
Pancreas, unknown	No.a	7137	1012	810	1292	69
	Lost to follow-up, %	0.6	1.3	2.2	2.4	1.4
	SEER*Stat Survival (95% CI)	13.2 (11.9 to 14.4)	14.3 (11.7 to 16.9)	10.9 (7.2 to 14.7)	13.2 (10.8 to 15.6)	11.9 (2.5 to 21.3)
	Sensitivity intervalb	N/Ac	11.1-16.4	6.9-14.3	10.6-15.3	2.5-21.2

Abbreviations: N/A = not applicable; SEER = Surveillance, Epidemiology, and End Results.

^aNumber diagnosed between 2005 and 2019 in SEER-22 registry areas.

^bAverage lower 95% CI bound for the “lower bound” scenario survival estimate, average upper 95% CI bound for the “upper bound” scenario survival estimate.

^cLost to follow-up adjustment not conducted for the reference group of US-born non-Hispanic White individuals.

^dLost to follow-up adjustment not conducted for subgroups with a percentage lost to follow-up less than or equal to that of the reference group (non-Hispanic White).

Table 5.

Comparison of 5-year age-adjusted cause-specific survival for prostate cancer, by racial group and stage at diagnosis, for men.

		5-y age-standardized cause-specific survival
Site	Statistic	Non-Hispanic White	Non-Hispanic Black	Hispanic and non-Hispanic Asian and Pacific Islander	Hispanic	Non-Hispanic American Indian or Alaska Native
Prostate, all stages	No.a	481 409	105 580	36 879	71 427	2805
	Lost to follow-up, %	1.6	2.7	4.4	6.5	2.4
	SEER*Stat Survival (95% CI)	91.9 (91.8 to 92.0)	89.3 (89.0 to 89.6)	92.8 (92.5 to 93.1)	90.8 (90.5 to 91.0)	86.2 (84.6 to 87.8)
	Sensitivity intervalb	N/Ac	89.6-90.6	91.8-93.3	90.0-91.4	83.2-89.0
Prostate, localized	No.a	370 745	82 730	26 848	51 285	1981
	Lost to follow-up, %	1.6	2.7	4.2	6.3	2.5
	SEER*Stat Survival (95% CI)	96.8 (96.8 to 96.9)	95.7 (95.5 to 96.0)	97.6 (97.4 to 97.9)	96.8 (96.5 to 97.0)	94.9 (93.5 to 96.4)
	Sensitivity intervalb	N/Ac	95.5-96.0	97.4-97.9	96.5-97.1	93.5-96.4
Prostate, regional	No.a	63 991	10 902	5340	9527	370
	Lost to follow-up, %	1.6	2.7	4.3	6.5	1.9
	SEER*Stat Survival (95% CI)	91.6 (91.2 to 92.0)	88.9 (87.4 to 90.3)	93.8 (92.8 to 94.9)	91.7 (90.6 to 92.8)	85.6 (81.5 to 89.7)
	Sensitivity intervalb	N/Ac	87.4-90.3	92.6-94.7	90.5-92.7	81.6-89.7
Prostate, distant	No.	27 072	7120	2633	5209	294
	Lost to follow-up, %	1.1	2.2	5.7	7.8	2.4
	SEER*Stat Survival (95% CI)	34.1 (33.2 to 34.9)	32.5 (31.1 to 33.9)	44.9 (42.2 to 47.7)	36.9 (35.1 to 38.6)	35.1 (28.2 to 42.0)
	Sensitivity intervalb	N/Ac	30.5-33.7	39.0-46.6	31.0-37.3	27.2-41.5
Prostate, unknown	No.a	19 601	4828	2058	5406	160
	Lost to follow-up, %	2.1	2.7	5.6	7.3	3.1
	SEER*Stat Survival (95% CI)	85.8 (85.2 to 86.3)	81.3 (80.0 to 82.6)	86.4 (84.0 to 88.8)	86.3 (85.2 to 87.4)	72.5 (62.2 to 82.9)
	Sensitivity intervalb	N/Ac	80.0-82.6	83.8-88.6	84.8-87.0	62.2-82.9

Abbreviations: N/A = not applicable; SEER = Surveillance, Epidemiology, and End Results.

^aNumber diagnosed between 2005 and 2019 in SEER-22 registry areas.

^bAverage lower 95% CI bound for the “lower bound” scenario survival estimate, average upper 95% CI bound for the “upper bound” scenario survival estimate.

^cLost to follow-up adjustment not conducted for the reference group of US-born non-Hispanic White individuals.

Table 6.

Comparison of 5-year age-adjusted cause-specific survival for breast cancer, by racial group and stage at diagnosis, for women.

		5-y age-standardized cause-specific survival
Site	Statistic	Non-Hispanic White	Non-Hispanic Black	Hispanic and non-Hispanic Asian and Pacific Islander	Hispanic	Non-Hispanic American Indian or Alaska Native
Female breast, all stages	No.a	482 069	74 848	65 365	87 700	4247
	Lost to follow-up, %	2.1	2.4	5.7	9.6	2.6
	SEER*Stat Survival (95% CI)	88.5 (88.4 to 88.6)	80.5 (80.1 to 80.9)	90.2 (89.8 to 90.5)	87.0 (86.6 to 87.3)	85.4 (83.8 to 87.0)
	Sensitivity intervalb	N/Ac	79.3-80.6	89.0-90.3	85.7-87.1	82.5-87.9
Female breast, localized	No.a	312 006	40 706	41 682	50 034	2500
	Lost to follow-up, %	2.2	2.5	5.3	9.2	2.8
	SEER*Stat Survival (95% CI)	96.0 (95.9 to 96.1)	92.7 (92.4 to 93.1)	97.1 (96.8 to 97.3)	95.4 (95.2 to 95.7)	94.4 (93.1 to 95.8)
	Sensitivity intervalb	N/Ac	92.4-93.1	96.8-97.4	95.1-95.9	93.1-95.8
Female breast, regional	No.a	136 158	26 339	19 484	30 944	1380
	Lost to follow-up, %	2.0	2.4	6.1	10.5	2.3
	SEER*Stat Survival (95% CI)	84.1 (83.8 to 84.3)	75.9 (75.2 to 76.7)	84.8 (84.0 to 85.7)	82.8 (82.0 to 83.5)	80.6 (77.1 to 84.1)
	Sensitivity intervalb	N/Ac	75.2-76.7	83.8-85.5	81.8-83.2	77.1-84.1
Female breast, distant	No.a	26 295	6348	3131	4927	274
	Lost to follow-up, %	1.3	1.6	6.1	6.2	2.6
	SEER*Stat Survival (95% CI)	29.3 (28.7 to 30.0)	19.5 (18.2 to 20.8)	33.5 (31.3 to 35.8)	28.6 (26.8 to 30.3)	33.9 (26.0 to 41.9)
	Sensitivity intervalb	N/Ac	18.1-20.8	28.9-34.7	24.4-29.3	25.4-41.3
Female breast, unknown	No.a	7610	1455	1068	1795	93
	Lost to follow-up, %	2.6	3.8	10.6	13.8	3.2
	SEER*Stat Survival (95% CI)	62.1 (60.9 to 63.3)	59.0 (56.0 to 62.1)	62.8 (58.9 to 66.6)	64.0 (61.1 to 66.9)	57.5 (45.0 to 69.9)
	Sensitivity intervalb	N/Ac	55.8-61.9	57.9-65.5	59.7-65.5	45.0-69.8

Abbreviations: N/A = not applicable; SEER = Surveillance, Epidemiology, and End Results.

^aNumber diagnosed between 2005 and 2019 in SEER-22 registry areas.

^bAverage lower 95% CI bound for the “lower bound” scenario survival estimate, average upper 95% CI bound for the “upper bound” scenario survival estimate.

^cLost to follow-up adjustment not conducted for the reference group of US-born non-Hispanic White individuals.

Compared with non-Hispanic White patients, non-Hispanic Black and non-Hispanic American Indian or Alaska Native patients generally had substantially lower 5-year cause-specific survival for all stages combined based on 95% CIs (Tables 2-6; Figures S1-S5). After adjusting for differential loss to follow-up, differences in survival for all stages combined persisted for all major cancer sites except pancreas, where the 5-year survival for non-Hispanic American Indian or Alaska Native patients (9.2 [sensitivity interval = 4.3-12.8]) was no longer substantially lower than that of non-Hispanic White patients (11.8). For non-Hispanic Black patients, survival differences identified based on 95% CIs for localized, regional, or distant disease remained statistically significant after applying our sensitivity analysis to each major cancer site. For non-Hispanic American Indian or Alaska Native patients, survival differences remained statistically significant for localized lung, prostate, and female breast and regional colon and rectum and prostate cancer.

Survival advantages for Hispanic and non-Hispanic Asian and Pacific Islander patients over non-Hispanic White patients persisted, even after adjusting for differential loss to follow-up for the following cancer sites and stages: lung (all stages combined); colon and rectum (localized, regional, all stages combined); pancreas (localized, all stages combined); prostate (localized, regional, distant); and female breast (all stages combined) (Tables 2-6; Figures S1-S5). For distant-stage colon and rectum cancer, all stages combined prostate cancer, and distant-stage female breast cancer, Hispanic and non-Hispanic Asian and Pacific Islander patients had substantially higher survival than non-Hispanic White patients based on the 95% CIs. The sensitivity intervals, however, were wider than the corresponding 95% CIs and contained the 5-year survival probabilities for non-Hispanic White patients. Specifically, for distant-stage colon and rectal cancer, the 95% CI for Hispanic and non-Hispanic Asian and Pacific Islander patients was 14.9 to 16.9 compared with non-Hispanic White patients (14.8 [95% CI = 14.5 to 15.1]), while the sensitivity interval was 14.1-16.4 (Table 3; Figure S2). For all stages combined prostate cancer, Hispanic and non-Hispanic Asian and Pacific Islander patients had a 95% CI of 92.5 to 93.1 compared with non-Hispanic White patients (91.9 [95% CI = 91.8 to 92.0]), but the sensitivity interval was 91.8-93.3 (Table 5; Figure S4). For distant-stage breast cancer, the 95% CI for Hispanic and non-Hispanic Asian and Pacific Islander patients was 31.3 to 35.8 compared with non-Hispanic White patients (29.3 [95% CI = 28.7 to 30.0]), whereas the sensitivity interval was 28.9-34.7 (Table 6; Figure S5).

For Hispanic patients, substantially lower survival than for non-Hispanic White patients persisted, even after adjustment for differential loss to follow-up for the following cancer sites and stages: colon and rectum (localized, regional, all staged combined) and female breast (localized, regional, all stages combined) (Tables 3 and 6; Figures S2 and S5). For regional and all stages combined lung cancer and distant-stage prostate cancer, Hispanic patients had substantially higher survival than non-Hispanic White patients based on 95% CIs. The sensitivity intervals, however, contained the 5-year survival probabilities for non-Hispanic White patients. Specifically, for regional lung cancer, the 95% CI for Hispanic patients was 36.1 to 38.8 compared with non-Hispanic White patients (35.4 [95% CI = 35.0 to 35.8]), while the sensitivity interval was 35.4-38.1 (Table 2; Figure S1). For all stages combined lung cancer, the 95% CI for Hispanic patients was 23.5 to 24.6 compared with non-Hispanic White patients (24.6 [95% CI = 24.4 to 24.8]), whereas the sensitivity interval was 22.2-24.4. For distant-stage prostate cancer, Hispanic patients had a 95% CI of 35.1 to 38.6 compared with non-Hispanic White patients (34.1 [95% CI = 33.2 to 34.9]), but the sensitivity interval was 31.0-37.3 (Table 5; Figure S4).

From our sensitivity analysis using ±2 SEs as opposed to ±1 SE, as outlined in Table 1 (Tables S21-S25), our conclusions remained unchanged, highlighting the robustness of our proposed methodology.

Discussion

We have demonstrated that non-Hispanic Black, Hispanic and non-Hispanic Asian and Pacific Islander, non-Hispanic American Indian or Alaska Native, and Hispanic patients invariably have higher proportions of loss to follow-up than do non-Hispanic White patients. Furthermore, differential loss to follow-up creates spurious survival advantages in minority populations in SEER. These findings indicate that differential loss to follow-up can introduce informative censoring that may affect interpretation of survival differences across racial and ethnic groups. Loss to follow-up is specific to the “reported alive” framework. In “presumed alive” approaches, these individuals are implicitly assumed to be alive, which may result in a higher degree of survival inflation when deaths are not captured.

For Hispanic and non-Hispanic Asian and Pacific Islander patients and Hispanic patients in particular, higher loss to follow-up may be the result of higher proportions of foreign-born individuals and difficulties in death ascertainment.² Greater loss to follow-up can bias estimates, thereby inflating survival and undermining the accuracy of population-based survival studies. Our proposed sensitivity analysis highlights how the reported SEER*Stat CIs are influenced by the degree of loss to follow-up. Thus, conclusions made based on the CIs could be misrepresentations of true heterogeneity between groups.

The cancer survival literature includes numerous studies suggesting survival advantages for Asian, Hispanic, and foreign-born individuals, but differences between racial and ethnic groups are nuanced and cannot be simplistically attributed to epidemiological phenomena such as the “Hispanic paradox” or the “healthy immigrant effect.” Our findings suggest that such conclusions may be inaccurate. For Hispanic and non-Hispanic Asian and Pacific Islander individuals, after adjusting for differential loss to follow-up, survival advantages for all stages combined disappear for prostate, bladder, and triple-negative breast cancer. For Hispanic patients, lung cancer survival became worse following adjustment. Furthermore, magnitudes of persistent survival advantages were often attenuated after applying our sensitivity analysis method.

As noted earlier, loss to follow-up is often higher for Hispanic and non-Hispanic Asian and Pacific Islander and Hispanic populations because discrepancies in identifiers used in registry linkages may reduce the accurate capture of deaths, along with phenomena such as the Salmon bias. Loss to follow-up likely affects both Hispanic and non-Hispanic Asian and Pacific Islander and Hispanic populations, potentially in different ways that cannot be fully characterized with the available data. Therefore, differences in the impact of loss to follow-up across groups should not be interpreted as evidence that follow-up completeness is equivalent or that bias is absent in either group. Rather, our findings suggest that even after applying our proposed method to account for differential loss to follow-up, Hispanic and non-Hispanic Asian and Pacific Islander populations seem to retain a survival advantage relative to non-Hispanic White individuals. The corresponding pattern among Hispanic persons is less consistent across analyses but generally shows a survival disadvantage relative to non-Hispanic White patients. The reasons for this difference in trend between Hispanic and non-Hispanic Asian and Pacific Islander individuals and Hispanic individuals may be related to differences in median household income, access to high-quality health care, or educational attainment between these 2 racial and ethnic groups. Both the Hispanic and non-Hispanic Asian and Pacific Islander and Hispanic categories, however, comprise heterogeneous subgroups with well-documented variation in health outcomes, which may further contribute to and complicate the observed patterns. Others have examined the potential impact of loss to follow-up through simulation. Okuyama et al.¹⁰ simulated scenarios using hospital-based cancer registry data in Japan to characterize the impact of loss to follow-up on survival for breast, colon, liver, lung, and stomach cancer and all cancers combined. The authors found that higher loss to follow-up resulted in inflated estimates of observed survival, with greater bias for cancer sites with poor prognosis, a finding that is consistent with Pinheiro et al.² Gil et al.¹¹ observed that censoring cases lost to follow-up led to an overestimate of net and overall survival, especially for cancers with a poor prognosis. Our results mirror these prior findings. Loss to follow-up had a greater impact on pancreatic cancer, a site with a generally poor prognosis, as indicated by substantially wider sensitivity intervals for Hispanic and non-Hispanic Asian and Pacific Islander individuals and Hispanic persons than for prostate cancer, a site with better overall prognosis.

Our proposed method is not without its limitations. We assumed stage-specific scenarios and applied these uniformly across cancer sites. Future work could tailor the scenarios also to be site specific. In addition, stage-specific estimates alone cannot fully resolve bias arising from prognostically informative loss to follow-up because the proportion of unknown-stage cases varies across racial and ethnic groups. Unknown-stage cancers typically have a poorer prognosis, and groups with higher loss to follow-up also tend to have more unknown-stage cases, reinforcing the potential for upwardly biased survival. By applying our sensitivity analysis to all excess cases lost to follow-up, including those with unknown stage, our method more effectively mitigates this limitation compared with standard stage-stratified survival analyses. We also used non-Hispanic White as the reference group for survival, which may lead to optimistic sensitivity interval bounds for certain groups known to have worse survival (ie, non-Hispanic Black and non-Hispanic American Indian or Alaska Native individual) given the potential for residual loss to follow-up. Furthermore, identifying differences through nonoverlapping CIs is more conservative but may represent stronger evidence than conventional statistical tests.

The major strength of this work is the use of population-based data from 22 SEER registries representing nearly 50% of the total US population. We also applied our methodology to a comprehensive list of cancer sites for a thorough illustration of the impact of loss to follow-up on survival. The sensitivity analysis methodology is general and can be applied to registry-based survival modeling, including Cox proportional hazards analysis, where many survival advantages originate. Furthermore, based on a sensitivity analysis, our methodology is robust to the specification of the number of SEs to add or subtract for the construction of alternative survival scenarios.

Our sensitivity analysis represents an essential advance in addressing the distorting effects of loss to follow-up in population-based cancer data. Although it cannot eliminate all potential sources of informative censoring, it substantially reduces their influence by balancing differences in loss to follow-up proportions across groups and incorporating stage distribution into the prediction of outcomes among patients who are lost to follow-up. This approach provides a more realistic characterization of the possible impact of loss to follow-up on survival patterns across diverse populations and helps distinguish true prognostic differences from artifacts created by differential loss to follow-up. It also prevents overinterpretation of inflated survival among groups with a high proportion of foreign-born individuals and guards against unwarranted pessimism for groups with higher proportions of US-born individuals, such as non-Hispanic White, non-Hispanic Black, and non-Hispanic American Indian or Alaska Native individuals. Given the inherent complexity of population-based cancer surveillance, meaningful reductions in loss to follow-up or major gains in the completeness of death ascertainment are unlikely soon. Consequently, the framework we propose offers the most practical and informative strategy currently available for evaluating heterogeneous cancer survival across racial and ethnic groups and provides a necessary foundation for more accurate interpretation of survival differences.

Even when applying our sensitivity analysis, non-Hispanic Black and non-Hispanic American Indian or Alaska Native individuals continue to have markedly lower survival. This result is consistent with well-known racial discrepancies that exist in cancer survival^12-16 and underscores the importance of not only identifying but also eliminating such discrepancies to improve survival for all patients with cancer. In addition, although it has been shown that Hispanic and non-Hispanic Asian and Pacific Islander individuals and Hispanic individuals have higher proportions lost to follow-up, nativity may also affect the proportion lost to follow-up, as reported in prior studies.^2,10,11 In future work, we plan to apply our proposed sensitivity analysis method by nativity to examine the impact birthplace has within each race or ethnicity and to further elucidate important discrepancies in survival that exist among diverse groups in the United States.

This study ultimately emphasizes the need for updated SEER guidelines regarding follow-up. The current SEER requirement of approximately 95% complete follow-up is not sufficient to address differential loss to follow-up across racial and ethnic groups, and this has direct implications for federal cancer statistics, the interpretation of health disparities, and how resources are prioritized and allocated. To mitigate the impact of differential loss to follow-up by race and ethnicity, we recommend that the 95% rule be adapted separately to major racial and ethnic groups. Such adjustments would support more accurate measurement of cancer survival across diverse populations, which can ultimately have implications for future patients with cancer. The NCI is working to incorporate this methodology into SEER*Stat–reported survival.

Conclusion

We have developed a novel sensitivity analysis method to examine the impact of the proportion lost to follow-up on cancer survival. Applying this method to survival data stratified by race and ethnicity, we have found that across multiple cancer sites, the currently reported 95% CIs for survival of non-White individuals may be too narrow and should be interpreted with caution. Our sensitivity analysis represents a new way of accounting for loss to follow-up to more accurately identify potential survival differences in the increasingly diverse population of patients with cancer in the United States.

Author contributions

Theresa Devasia (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Visualization, Writing—original draft, Writing—review & editing), Paulo S. Pinheiro (Conceptualization, Data curation, Methodology, Supervision, Writing—original draft, Writing—review & editing), and Mandi Yu (Conceptualization, Data curation, Methodology, Supervision, Writing—original draft, Writing—review & editing)

Supplementary material

Supplementary material is available at JNCI: Journal of the National Cancer Institute online.

Funding

T.D. and M.Y. performed this work as part of their official duties as employees of the NCI. P.S.P. is supported by the NCI through a consulting agreement with Westat Inc.

Conflicts of interest

The authors declare no potential conflicts of interest.

Data availability

SEER survival data can be accessed at https://seer.cancer.gov/. Because access to SEER data requires a formal request, the data used in this study cannot be made publicly available. For the specific data files used in the analysis, please contact the corresponding author.