Regional differences in population‐based cancer survival between six prefectures in Japan: Application of relative survival models with funnel plots

Abstract

We used new methods to examine differences in population‐based cancer survival between six prefectures in Japan, after adjustment for age and stage at diagnosis. We applied regression models for relative survival to data from population‐based cancer registries covering each prefecture for patients diagnosed with stomach, lung, or breast cancer during 1993–1996. Funnel plots were used to display the excess hazard ratio (EHR) for each prefecture, defined as the excess hazard of death from each cancer within 5 years of diagnosis relative to the mean excess hazard (in excess of national background mortality by age and sex) in all six prefectures combined. The contribution of age and stage to the EHR in each prefecture was assessed from differences in deviance‐based R ² between the various models. No significant differences were seen between prefectures in 5‐year survival from breast cancer. For cancers of the stomach and lung, EHR in Osaka prefecture were above the upper 95% control limits. For stomach cancer, the age‐ and stage‐adjusted EHR in Osaka were 1.29 for men and 1.43 for women, compared with Fukui and Yamagata. Differences in the stage at diagnosis of stomach cancer appeared to explain most of this excess hazard (61.3% for men, 56.8% for women), whereas differences in age at diagnosis explained very little (0.8%, 1.3%). This approach offers the potential to quantify the impact of differences in stage at diagnosis on time trends and regional differences in cancer survival. It underlines the utility of population‐based cancer registries for improving cancer control. (Cancer Sci 2009; 100: 1306–1311)

The Japanese Government launched the Fundamental Planning of Cancer Control Promotion based on the Fundamental Bill on Cancer Control in June 2007. One of the mainstays of this new strategy was to ‘narrow the inequalities of cancer medical services’. Monitoring cancer survival among the prefectures of Japan is important, both to evaluate progress toward this goal and as a contribution to the next Cancer Control Plan or regional cancer control planning. Wide regional differences in cancer survival in Japan have been reported, but the findings were only adjusted by age at diagnosis.^{( 1 )}

Multivariable models of relative survival have increasingly been used to quantify the impact of various prognostic factors (e.g. country, hospital, calendar period, age).^{( 2} , ³ , ^{4 )} Funnel plots, mostly used in meta‐analyses, have been used more recently as additional tools for such comparisons.^{( 5} , ⁶ , ^{7 )} In the present study, we combined multivariable relative survival models with the funnel plot approach,^{( 8 )} to investigate differences in population‐based cancer survival between six prefectures in Japan. The role of age and stage at diagnosis was evaluated for cancers of the stomach, lung, and breast (women).

Materials and Methods

Patients.  The collaborative study of cancer survival^{( 9 )} collated data from 11 prefectural cancer registries on some 373 000 cancer patients diagnosed between 1993 and 1996. The national cancer survival figures were estimated on 279 469 records from the seven registries (Yamagata, Miyagi, Niigata, Osaka, Fukui, Tottori, and Nagasaki) that met the quality requirements (death certificate only cases less than 25%; death certificate notification less than 30%; vital status unknown for less than 5% of patients).^{( 9 )}

These data formed the basis of the analyses reported here, but the data from the Tottori registry (4% of the total) were excluded because tumor stage was missing. Overall, 84 350 cases diagnosed with a first, primary, invasive malignant tumor of the stomach (ICD‐10^{( 10 )} code C16), lung (C33‐C34), or breast (C50; only women) between 1993 and 1996 and followed up for at least 5 years were considered as eligible for survival analysis. Of these, we excluded 11 874 patients (14.1% of those eligible) for whom the tumor stage at diagnosis was unknown, and 72 476 patients (85.9%) were included in the survival analyses.

Methods.  We first applied relative survival models to examine differences in cancer survival between the six prefectures. The adjusted excess hazard of death for each prefecture was then compared with the grand mean using the funnel plot approach.

In a second step, focussing on the prefecture with the lowest survival, we assessed the influence of age and stage at diagnosis on survival using the R ² measure to estimate the proportion of variation explained by each variable.

Regional differences in survival up to 5 years since diagnosis: the funnel plots.  The excess hazard ratios (EHR) of death from each cancer within 5 years of diagnosis were estimated for each prefecture with a Poisson regression model for relative survival,^{( 11 )} adjusting first for age, then for age and stage combined. The expected (background) mortality, which is removed from the observed overall mortality, was obtained from complete (single‐year‐of‐age) national life tables.^{( 12 )} The contrasts used in the model were modified such that the excess hazard of each prefecture was compared to the overall mean hazard of death in excess of the national background mortality. This ‘grand mean’ across the six prefectures represents the ‘target’ in the funnel plots,^{( 7} , ^{8 )} that is, the excess hazard of death against which the hazard among cancer patients in each prefecture was compared. Both 95 and 99.8% control limits were estimated according to the ‘precision’, represented by the inverse variance of the grand mean, and displayed on the x‐axis of the funnel plots. An excess hazard outside the 95 (dotted lines) or 99.8% (dashed lines) control limits means that the excess hazard of death from that cancer in that prefecture was considerably higher (if above the limits) or lower (if below) than the risk of death from that cancer in all the prefectures combined.

Evaluation of the role played by prognostic factors on the lowest survival.  We then focused on the prefecture with the lowest survival for each cancer and evaluated the role of age and tumor stage at diagnosis using R ² measures for the Poisson regression model, based on deviance residuals.^{( 13 )} We used four models to quantify the effect of adjusting the excess hazard for age and stage. Model 1 comprised the follow‐up time (0‐, 0.25‐, 0.5‐, 1‐, 2‐, and 3–5 years since diagnosis) and the region. In model 2, age at diagnosis was added to model 1, whereas model 3 consisted of model 1 plus stage at diagnosis. Model 4 included both age and stage. The effect of age adjustment was defined as the difference in R ² between model 4 (adjusted for both age and stage) and model 3 (adjusted for stage). The effect of adjusting the excess hazard for stage was represented by the difference in R ² between model 4 (age and stage) and model 2 (age).

Results

Stomach cancer.  Five‐year relative survival was lower in Osaka than in the other five prefectures for both sexes (data not shown). After adjustment for age at diagnosis, the excess hazard of death in Osaka was above the upper 99.8% control limit (Fig. 1). Additional adjustment for stage at diagnosis reduced the excess hazard in Osaka slightly, but it was still above the upper 95% control limit for both sexes (Fig. 2). Some realignment of the prefectural excess hazards was also observed. The data from Miyagi and Niigata showed a significantly low excess hazard of death from stomach cancer. In Miyagi, this persisted after adjustment for both age and stage (1, 2).

Figure 1.

Funnel plots of the age‐adjusted log excess hazard of death within 5 years of diagnosis, by prefecture: cancers of the stomach, lung, and breast. Precision (x‐axis) is the inverse of the variance of the age‐adjusted log excess hazard of death. The target (‘grand mean’) is the average of the log excess hazard of death across the six prefectures

Figure 2.

Funnel plots of the age‐ and stage‐adjusted log excess hazard of death within 5 years of diagnosis, by prefecture: cancers of the stomach, lung, and breast. Precision (x‐axis) is the inverse of the variance of the age‐ and stage‐adjusted log excess hazard of death. The target (‘grand mean’) is the average of the log excess hazard of death across the six prefectures

We examined further the role of age and stage on the lower survival in Osaka. Cancer patients in Osaka tended to be diagnosed at a younger age and, for stomach cancer, at a more advanced stage (Table 1). We further restricted the analysis to those cancer registries that conducted active follow up of cancer patients, namely Osaka, Yamagata, and Fukui.

Table 1.

Characteristics of cancer patients diagnosed between 1993 and 1996 in six prefectures in Japan: selected cancers

		Prefecture												Total
		Yamagata		Fukui		Osaka		Niigata		Miyagi		Nagasaki
Resident population (1995)		1 256 958		826 996		8 797 268		2 488 364		2 328 739		1 544 934		17 243 259
Stomach
Men	Incidence (per milion) †	111.8		104.3		74.2		113.5		97.7		82.3		87.1
	Mortality (per milion) ‡	48.5		37.5		47.5		49.3		42.8		37.5		42.1
	Age (years)	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%
	15–44	157	4.5	87	4.7	466	4.5	263	4.6	177	5.2	125	5.5	1275	4.7
	45–54	389	11.2	208	11.1	1582	15.4	658	11.5	437	12.9	233	10.2	3507	13.0
	55–64	908	26.0	484	25.9	3212	31.3	1637	28.6	992	29.2	636	27.9	7869	29.1
	65–74	1308	37.5	649	34.7	3170	30.9	2082	36.4	1215	35.8	813	35.6	9237	34.2
	75–99	724	20.8	441	23.6	1830	17.8	1082	18.9	577	17.0	474	20.8	5128	19.0
	Stage
	Localized	2007	57.6	1040	55.6	4830	47.1	3391	59.3	1914	56.3	1193	52.3	14 375	53.2
	Regional	941	27.0	499	26.7	3442	33.5	1619	28.3	919	27.0	716	31.4	8136	30.1
	Distant	538	15.4	330	17.7	1988	19.4	712	12.4	565	16.6	372	16.3	4505	16.7
	Total	3486	100.0	1869	100.0	10260	100.0	5722	100.0	3398	100.0	2281	100.0	27 016	100.0
Women	Incidence (per milion) †	48.5		44.0		28.2		40.8		35.2		34.4		33.7
	Mortality (per milion) ‡	22.0		17.8		17.9		17.5		14.9		14.4		16.4
	Age (years)	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%
	15–44	110	5.9	84	7.9	438	8.7	215	7.5	170	10.4	107	9.1	1124	8.3
	45–54	138	7.4	109	10.3	876	17.4	309	10.8	210	12.8	134	11.4	1776	13.0
	55–64	354	19.1	208	19.6	1118	22.3	610	21.4	330	20.1	243	20.7	2863	21.0
	65–74	690	37.2	312	29.4	1371	27.3	951	33.3	545	33.3	362	30.8	4231	31.1
	75–99	565	30.4	348	32.8	1221	24.3	772	27.0	384	23.4	330	28.1	3620	26.6
	Stage
	Localized	1018	54.8	526	49.6	2198	43.8	1668	58.4	841	51.3	590	50.2	6841	50.2
	Regional	526	28.3	351	33.1	1761	35.1	841	29.4	495	30.2	371	31.5	4345	31.9
	Distant	313	16.9	184	17.3	1065	21.2	348	12.2	303	18.5	215	18.3	2428	17.8
	Total	1857	100.0	1061	100.0	5024	100.0	2857	100.0	1639	100.0	1176	100.0	13 614	100.0
Lung
Men	Incidence (per milion) †	51.8		56.8		65.0		63.4		60.3		68.8		55.9
	Mortality (per milion) ‡	45.7		50.8		57.9		47.8		50.8		55.3		47.3
	Age (years)	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%
	15–44	22	1.8	21	2.2	179	2.7	41	1.7	35	2.7	36	2.8	334	2.4
	45–54	63	5.1	50	5.3	696	10.4	180	7.3	101	7.8	92	7.2	1182	8.5
	55–64	263	21.3	184	19.5	1636	24.4	509	20.5	275	21.4	256	19.9	3123	22.4
	65–74	517	41.8	386	40.9	2514	37.5	1077	43.5	587	45.6	565	44.0	5646	40.5
	75–99	372	30.1	303	32.1	1680	25.1	671	27.1	290	22.5	335	26.1	3651	26.2
	Stage
	Localized	236	19.1	240	25.4	1209	18.0	808	32.6	245	19.0	320	24.9	3058	21.9
	Regional	444	35.9	372	39.4	2826	42.1	986	39.8	485	37.7	507	39.5	5620	40.3
	Distant	557	45.0	332	35.2	2670	39.8	684	27.6	558	43.3	457	35.6	5258	37.7
	Total	1237	100.0	944	100.0	6705	100.0	2478	100.0	1288	100.0	1284	100.0	13 936	100.0
Women	Incidence (per milion) †	15.6		14.9		19.0		17.4		16.0		19.2		16.8
	Mortality (per milion) ‡	12.0		9.4		17.1		10.5		10.8		14.1		12.6
	Age (years)	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%
	15–44	17	3.7	13	3.8	101	3.9	24	2.9	19	4.4	23	4.4	197	3.8
	45–54	44	9.6	26	7.6	309	12.1	77	9.3	48	11.0	43	8.3	547	10.6
	55–64	97	21.1	61	17.9	526	20.5	177	21.3	104	23.9	117	22.5	1082	21.0
	65–74	156	33.9	119	35.0	814	31.8	310	37.3	156	35.8	195	37.4	1750	34.0
	75–99	146	31.7	121	35.6	813	31.7	242	29.2	109	25.0	143	27.4	1574	30.6
	Stage
	Localized	133	28.9	110	32.4	514	20.1	344	41.4	109	25.0	178	34.2	1388	27.0
	Regional	112	24.3	114	33.5	1003	39.1	250	30.1	131	30.0	160	30.7	1770	34.4
	Distant	215	46.7	116	34.1	1046	40.8	236	28.4	196	45.0	183	35.1	1992	38.7
	Total	460	100.0	340	100.0	2563	100.0	830	100.0	436	100.0	521	100.0	5150	100.0
Breast
Women	Incidence (per milion) †	43.5		40.3		41.6		38.8		53.5		43.1		43.6
	Mortality (per milion) ‡	9.2		9.6		12.0		8.3		9.8		9.7		10.4
	Age (years)	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%	No.	%
	15–44	189	19.9	152	20.2	1174	19.7	511	23.9	371	21.5	272	22.1	2669	20.9
	45–54	247	25.9	232	30.9	2121	35.6	647	30.2	536	31.0	352	28.6	4135	32.4
	55–64	201	21.1	156	20.7	1330	22.3	424	19.8	389	22.5	240	19.5	2740	21.5
	65–74	210	22.1	135	18.0	840	14.1	389	18.2	305	17.7	254	20.6	2133	16.7
	75–99	105	11.0	77	10.2	489	8.2	171	8.0	127	7.3	114	9.3	1083	8.5
	Stage
	Localized	563	59.1	439	58.4	3275	55.0	1215	56.7	930	53.8	641	52.0	7063	55.4
	Regional	320	33.6	263	35.0	2324	39.0	827	38.6	686	39.7	521	42.3	4941	38.7
	Distant	69	7.2	50	6.6	355	6.0	100	4.7	112	6.5	70	5.7	756	5.9
	Total	952	100.0	752	100.0	5954	100.0	2142	100.0	1728	100.0	1232	100.0	12 760	100.0

^† The age‐adjusted incidence rates per 100 000 (Standard Population: Japanese 1985 model population) in 1998 (the estimation of the incidence in each prefecture was based on the collaborative study of cancer incidence in Japan^{( 21 )} and the total incidence was estimated using data from the 12 population‐based cancer registries in Japan^{( 22 )}).

^‡ Age‐adjusted mortality rate per 100 000 (Std. Pop.: 1985 Japanese model population) in 1998 (data from vital statistics of Japan^{( 23 )}).

In this restricted analysis, the excess hazard of death for both sexes in Osaka was still significantly higher than in the comparison group of Yamagata and Fukui combined (Table 2: model 1). The EHR barely changed after adjustment for age (model 2). The EHR fell after accounting for stage (models 3 and 4), but it was still significantly high. We estimated that differences in age at diagnosis explained as little as 0.8% in men and 1.3% in women of the difference in cancer survival between Osaka and Yamagata and Fukui combined (Table 3). By contrast, differences in tumor stage appeared to explain 61.3 and 56.8% of the survival differences in men and women respectively (Table 3). This mainly reflects a higher proportion of patients with advanced stage (Table 1), particularly for regional disease (data not shown).

Table 2.

Stomach cancer: excess hazard ratio (EHR) of death within five years since diagnosis in Osaka relative to Yamagata and Fukui combined: patients diagnosed 1993–1996

	No. patients	Model 1		Model 2		Model 3		Model 4
		Follow‐up time and region		Model 1 + age at diagnosis		Model 1 + stage at diagnosis		Model 1 + age and stage
		EHR	95% CI	EHR	95% CI	EHR	95% CI	EHR	95% CI
Men
Region
Yamagata + Fukui		5355		1.00		1.00		1.00
Osaka	10 260	1.53	1.44–1.62	1.59	1.50–1.68	1.26	1.19–1.33	1.29	1.22–1.36
Age group
15–59	4858			1.00				1.00
60–69	5508			1.29	1.21–1.37			1.16	1.09–1.24
70–99	5249			1.77	1.66–1.88			1.42	1.33–1.52
Stage
Localized	7877					1.00		1.00
Regional	4882					12.43	11.07–13.95	11.54	11.07–13.95
Distant	2856					47.02	41.78–52.90	42.77	41.78–52.90
Women
Region
Yamagata + Fukui	2918	1.00		1.00		1.00		1.00
Osaka	5024	1.55	1.44–1.68	1.65	1.53–1.78	1.34	1.25–1.45	1.43	1.32–1.54
Age group
15–59	2522			1.00				1.00
60–69	2057			1.09	0.99–1.20			1.12	1.02–1.24
70–99	3363			1.70	1.57–1.85			1.52	1.40–1.65
Stage
Localized	3742					1.00		1.00
Regional	2638					12.90	10.96–15.17	11.69	10.03–13.63
Distant	1562					49.43	41.85–58.38	44.19	37.75–51.73

CI, confidence interval.

Table 3.

Summary of the excess hazard of death for stomach cancer patients in Osaka compared with Fukui + Yamagata

Model	Variables included in model	Men		Women
		EHR	R 2	EHR	R 2
1	Follow up, region	1.53	0.344	1.55	0.380
2	+ Age	1.59	0.364	1.65	0.403
3	+ Stage	1.26	0.970	1.34	0.958
4	+ Age and stage	1.29	0.978	1.43	0.971

The effect of age (difference in R ² between model 4 and model 3: see text) was 0.8 and 1.3% in men and women respectively. The effect of stage (difference in R ² between model 4 and model 2) was 61.3 and 56.8% in men and women respectively. EHR, excess hazard ratio.

Lung cancer.  Age‐adjusted excess hazards were lower than the 99.8% control limit in both sexes in Niigata and among women in Nagasaki (Fig. 1). These populations had a higher proportion of localized tumors (Table 1) and, after additional adjustment for tumor stage, the excess hazards of death were all within the 95% control limits except for men in Miyagi prefecture (Fig. 2).

Breast cancer.  No outlier was found among the six prefectures for 5‐year relative survival or the excess hazard of death from breast cancer within 5 years of diagnosis (1, 2).

Discussion

Analysis of population‐based cancer data showed wide differences in 5‐year relative survival from stomach cancer between the six prefectures, after adjustment for age and stage. Patients in Osaka prefecture had higher than average excess mortality attributable to stomach cancer, whereas lower excess mortality was seen in Miyagi prefecture. Additional analyses restricted to three prefectures showed that more advanced stage at stomach cancer diagnosis accounted for approximately 60% of the excess hazard of death in Osaka.

Many cancer screening programmes (stomach, lung, breast, cervix, colorectal, even prostate cancer) have been implemented in Japan with public resources, but they have often not been well organized, with deficient management, poor definition of the target population, low participation (e.g. stomach cancer screening uptake 43.2% in Yamagata, 28.8% in Fukui, 17.9% in Osaka),^{( 14 )} and poor quality control. Although such issues have not yet been fully documented, the uptake or quality of screening may have been worse in Osaka, by far the most populous prefecture examined here (Table 1).

The proportion of records excluded from analysis because of missing data on stage varied widely by prefecture. The inclusion of cases with missing stage in unadjusted analyses did not, however, eliminate regional differences in stomach cancer survival. Stage distribution is an indication of early detection of cancer, but it does not explain the lower overall survival in Osaka: stage‐specific survival was also lower. Patients with regional disease and, to a lesser degree, those with localized cancer, had much lower survival in Osaka prefecture than in Yamagata and Fukui.

Regional disparities in health care management could play a major role in the remaining differences in stomach cancer survival. First, differential cancer screening coverage between Osaka and Yamagata‐Fukui was likely to produce lead‐time bias and/or length bias and might explain some of the differences in survival. Second, only 25% of cancer patients in Osaka were treated in the designated cancer care hospitals, whereas this proportion reached 70–80% in Fukui and Yamagata.^{( 15 )} Third, lower 5‐year survival has been reported for cancer patients treated in low‐volume hospitals:^{( 16} , ¹⁷ , ¹⁸ , ^{19 )} in Osaka, a higher proportion of cases was treated in such hospitals.

Significant differences between prefectures in the age‐adjusted EHR for lung cancer disappeared after adjustment for stage. We infer that the differences in lung cancer survival arose mainly from differences in stage at diagnosis. In particular, cancer patients in Niigata and Nagasaki were on average diagnosed at an earlier stage than those in other prefectures. The high proportion of localized cases in Niigata could be explained by the high participation in screening. In Miyagi, the survival of localized cases was much higher than in other prefectures (data not shown), which could be due to lead‐time bias and/or length bias among screen‐detected cases. Niigata and Miyagi are two of the prefectures that have promoted cancer screening the most.

By contrast, no large disparities in survival were observed for breast cancer while all six prefectures achieved high survival on an international scale.^{( 5 )} Such observations show that regional differences in survival are not inevitable and that the overall organization of health care (from early diagnosis and screening through to treatment) can reach a uniformly high standard. It also demonstrates that the differences in survival observed for the other cancers were not simply the result of a complex data artefact.

This is to our knowledge the first report of differences in population‐based cancer survival in Japan using multivariable relative survival models, whereas crude survival (e.g. estimated with Cox proportional hazard models) does not account for the differences in background mortality. We did not control for background mortality by prefecture because it was shown to vary very little.^{( 20 )}

The contrast used here for the funnel plots enabled us to examine the distribution of the excess hazard of death from each cancer in each prefecture in relation to an overall mean excess hazard, after adjustment for age and stage at diagnosis. This approach also enabled us to take into account the differences in precision of the estimates arising from the wide differences in the population of each prefecture.

This cancer survival study was limited to six prefectures. Population‐based cancer registration is present in 35 of the 47 prefectures and one city in Japan, but the quality of registration and follow up is often too poor and the proportion of records with missing information on stage too high for systematic survival analysis. Even in the six prefectures that met the predetermined quality criteria, there were some unresolved data management issues. Furthermore, we limited the additional analysis on three prefectures with similar follow‐up procedure in order to make the results more comparable.

Analysis of secular trends in these regional disparities in survival, using more recent data, will enable us to improve these investigations. Comparable approaches could also be applied to examine differences in cancer survival between smaller administrative geographies within a given prefecture, such as second‐level medical care districts.

High‐quality cancer registries with individual follow‐up information are a key requirement for effective cancer control. The infrastructure of cancer registration in Japan has lagged behind that in European countries, Canada, and the USA. Systematic analysis of the data from a network of cancer registries is indispensable for monitoring improvements in cancer survival, for assessing equity in the outcome of cancer care, and for implementing and evaluating cancer control policies.

We showed that the use of the multivariable relative survival model combined with funnel plot approach was useful for assessing the regional disparities in cancer survival. It enabled us to quantify the impact of differential age and stage distributions on these regional inequalities. Our study illustrates the value of population‐based cancer registries for improving cancer control.