Temporal changes in liver cancer incidence rates in Japan: Accounting for death certificate inaccuracies and improving diagnostic techniques

Primary liver cancer (PLC) rates have risen dramatically during the past few decades in some regions, particularly in Japan, where PLC is now the third major cause of cancer death. PLC is one of the most difficult tumors to diagnose correctly, because (i) the liver is a frequent site of cancer metastasis and (ii) death from PLC is often attributed to cirrhosis or chronic hepatitis. Also, because the disease is often rapidly fatal, a large proportion of liver cancer cases are identified based on death certificates alone without confirmation by clinical records. Thus, worldwide differences in published incidence rates for this disease reflect regional or national differences in both the accuracy of death certificates and the sensitivity of diagnostic methods. By comparing death certificate causes of death with those based on pathology review, we were able to adjust 1958–1994 incidence rates for a large Japanese cohort for these errors. Although the death certificate false‐positive error rate declined, the false‐negative error rate remained high throughout the study. The introduction of improved liver cancer diagnostic methods in Japan in the early 1980s was associated with a sharp increase in PLC incidence. We conclude that errors in death certificate causes of death and changes in liver cancer diagnostic techniques have had an important impact on the reported incidence of this disease. Taking these factors into account, rates of hepatocellular carcinoma rose between 2.4‐ and 4.3‐fold in our Japanese cohort from 1960 to 1985, peaked about 1993 and declined thereafter. Incidence rates of cholangiocarcinoma remained stable through 1987. © 2001 Wiley‐Liss, Inc.

Over the last three decades, incidence and mortality rates for primary liver cancer (PLC) have risen in some areas, including the United States, 1 France, 2 the United Kingdom 3-5 and particularly Japan. 6,7 In comparison, however, PLC rates have declined in Singapore, Spain and through 1980 in Sweden. 7 There are several difficulties in determining the true incidence of PLC and in assessing changes over time. The liver is a frequent site of cancer metastasis, with as many as 40 -50% of liver cancers reported on death certificates originating in other organs. 8,9 In addition, death from liver cancer is often attributed to one or more of the two conditions that frequently accompany it: cirrhosis and chronic hepatitis. 10,11 PLC diagnostic methods have improved over time, especially with the introduction of ultrasound and computed tomography (CT) scanning to diagnose PLC in Japan and other developed countries in the early 1980s, 12 and many PLC cases are now being diagnosed that might once have gone undetected. Thus, in order to describe changes in Japanese PLC rates over time, it is important to determine both how many deaths due to metastatic liver cancer or other diseases were misdiagnosed as PLC and how many PLC cases were missed because of less sensitive diagnostic methods or errors.
Another problem with assessing changes in incidence and mortality rates of PLC in Japan and worldwide is that "primary liver cancer" consists mainly of two types of cancers with significantly different etiologies: hepatocellular carcinoma (HCC) and cholangiocarcinoma (CC). 13 Because there are geographical differences in the distribution of PLC cases by histological subtype, changes in PLC incidence will reflect both these differences and underlying changes in HCC and CC rates. In Japan in recent years, PLC has been primarily HCC, with HCC comprising 94% of PLC cases with known histology reported to the Osaka registry from 1988 to 1992 and CC comprising just 5%. 14 In contrast, in the United States, 74% of PLC cases in a recent study were determined to be HCC. 1 Previous studies have shown increases in recent years in combined HCC and CC incidence rates 2,3 as well as in HCCspecific rates in Scotland, 4 the United States 1 and Japan. 15 Few studies have focused specifically on CC, and little is known about temporal trends in CC morbidity and mortality rates in recent decades. The goal of our study was to investigate the overall and subtype-specific incidence of primary liver cancer over time in our Japanese cohort, correcting for diagnostic errors by classifying liver cancer cases as detected true-positives, estimated true-positives, detected false-negatives and estimated false-negatives.

Study population
We conducted population-based follow-up of the extended Life Span Study (LSS) cohort of atomic-bomb survivors in Hiroshima and Nagasaki at the Radiation Effects Research Foundation (RERF). This cohort consists of 120,321 radiation-exposed and unexposed persons who were official residents of either city at the time of the bombings in 1945 and who were alive and residing there at the time of censuses conducted in 1950, 1951 and 1952, as described in more detail elsewhere. 16 The results were adjusted to remove estimated radiation effects.
Deaths were ascertained using the Japanese koseki, a family registry system that permits virtually complete tracking of deaths and emigration status of cohort members throughout Japan. Cancer incidence data were retrieved from the Hiroshima and Nagasaki city tumor registries established in 1957 and 1958; the Nagasaki registry was expanded to cover Nagasaki prefecture in 1985. All persons who were alive and not previously diagnosed with cancer were followed from 1958 to 1994 or until the earliest of first cancer diagnosis, death or loss to follow-up. Because cancer data other than death certificates were rarely available for cohort members who resided outside these two catchment areas, such cases were excluded from incidence analyses.
Methods used to ascertain LSS liver cancer incidence cases are described in more detail elsewhere. 17 Basically, a review panel of three pathologists compared clinical and histologic records with causes of death listed on death certificates for persons diagnosed through 1987. Subjects whose death certificates listed any of the following as underlying causes of death (or as other physical conditions at time of death) were identified and reviewed: primary, secondary or unspecified cancers of the liver and biliary systems, including cancer of the intra-or extrahepatic bile ducts, or of overlapping or multiple sites (ICD-7,-8 and -9 codes 155.0 and 155.1; ICD-7 codes 155.8 and 156; ICD-8 and -9 codes 155.2, 156.1, 156.9 and 197.7; ICD-8 code 197.8 and ICD-9 code 156.8). The pathologists also searched for missed PLC cases by examining clinical and histologic records for subjects whose death certificates listed them as having died from diseases frequently confused with PLC: chronic liver disease, including cirrhosis (ICD-7 code 581; ICD-8 and -9 code 571), other liver diseases (ICD-7 codes 580, 582 and 583; ICD-8 and -9 codes 570, 572 and 573), pancreatic cancer (ICD-7, -8 and -9 code 157) and gallbladder or biliary disease (ICD-7 code 586, ICD-8 code 576 and ICD-9 codes 575 and 576).
The pathologists independently reviewed case materials, meeting jointly to reach consensus when they disagreed. Tissue slides were prepared by hematoxylin and eosin stains. Histological classification followed that proposed by the World Health Organization. 18 About 28% of reviewed subjects for whom PLC was listed as cause of death received autopsies at RERF after death certificates were completed; about half of reviewed subjects with other official causes of death received such autopsies.

Estimating death-certificate accuracy
By comparing pathological and clinical tissue samples and records with death certificate causes of death, the pathology panel was able to ascertain the true cause of death for many study subjects. However, in many instances death occurred before sufficient pathological or clinical examinations could be performed, resulting in the source of diagnosis being limited to the death certificate only (DCO). This made pathology review impossible. We used logistic regression to estimate the probability of falsepositive PLC death certificate diagnoses adjusted by sex, age at death and year of death. Included in logistic models were all deceased subjects who had death certificates listing their causes of death as liver cancer who also had clinical or histologic records that could be reviewed, regardless of whether subjects died outside the catchment areas of the study or had first primary tumors. The outcome variables were the proportions of liver cancer deaths that the Hiroshima-Nagasaki Pathology Review Panel determined to be PLC or not, the latter subjects either having metastatic liver cancer or not having liver cancer at all.
We used a similar procedure to correct DCO cause of death determinations for false-negative errors. We limited analysis to subjects who had death certificates with causes of death including chronic liver disease, cirrhosis or pancreatic cancer who also had clinical or pathology records that could be examined by the pathology team. We compared these two sets of records and calcu-lated the proportion of deaths attributed to one of these causes that were found by the review panel to be PLC. Thus, we were able to estimate the probability of false-negative DCO diagnoses for grouped years of the study.

Estimation of primary liver cancer incidence
Although we included all subjects who were reviewed in determining the accuracy of death certificate diagnoses, incidence calculations were limited to first PLC cases diagnosed within our catchment areas. Restricting analysis to such cases, some of whom were still alive, we calculated PLC incidence rates over time, adding together the pathologist-confirmed PLC cases and the DCO cases. We adjusted the latter numbers for false-positive diagnoses by multiplying each year's DCO cases by the appropriate sex-and age-specific probabilities that DCO cases were true-positives for PLC. To calculate incidence rates we used Poisson regression based on person-year tables 19,20 that were constructed using gender and city categories and by stratifying age and calendar period into 5-year intervals. Because there was no active PLC ascertainment outside of the Hiroshima and Nagasaki catchment areas, we adjusted person-years of follow-up by using previously estimated gender-, age-and city-specific migration probabilities in a 15% sample of the cohort for whom migration status was known. 21 We modeled age and calendar-period trends as log-linear functions of calendar year and the logarithm of age. Calendar year (y), log age (ln[a]), city (c) and sex (s) variables were centered so that the values of all fitted model parameters represented the average background incidence in the cohort: calendar year by subtracting 1971, log age by dividing age by the person-year weighted mean age during the follow-up period (48.4 years for males and 51.4 years for females) and city and sex by subtracting the relevant proportion in each city or of each sex. The fitted model was c,s ͑a, y͒ ϭ exp͕␣ 0 ϩ ␣ sex s ϩ ␣ city c ϩ ␣ ln͑age͒ ln͑a͒ ϩ ␣ ln͑age͒ 2ln͑a͒ 2 ϩ ␣ sex,ln͑age͒ s ln͑a͒ ϩ ␣ year y ϩ ␣ city,year cy} Radiation risk was adjusted using a linear excess relative risk model with adjustment for age at exposure. 17 To take advantage of all available cases, unexposed members of the LSS cohort who were not present at the time of the bombings and persons with unknown radiation dose were included in the analysis by defining special categories for them.
To adjust incidence rates for missed PLC cases, we first calculated by grouped years the number of DCO subjects whose causes of death were cirrhosis, chronic hepatitis or pancreatic cancer. Then we multiplied these numbers by the proportions of such subjects dying during these time frames that pathology review had shown to be PLC cases. We added these cases to the numbers of confirmed PLC cases and the downwardly adjusted numbers of DCO subjects who were likely to be true-positive PLC cases. The PLC incidence trends were then recalculated taking these additional cases into account. We were unable to review enough subjects with these other official causes of death to also consider sex and age at death in this analysis.
To determine how the incidence of PLC changed in our cohort since 1987, we calculated the annual age-, city-and sex-adjusted incidence from 1988 to 1994, the latest year for which tumor registry data were complete. Because we were not able to conduct pathology review of these more recent subjects, these incidence rates could not be adjusted for false-negative DCO errors. However, we were able to adjust for most false-positive DCO errors by determining the year when adjusting for these DCO errors had the same effect as discarding DCO cases altogether. By discarding DCO cases of PLC diagnosed before this time, we were able to analyze the combined 1958 -1994 dataset.
We estimated incident rates of hepatocellular carcinoma and cholangiocarcinoma using a cause-specific joint analysis of a detailed cross-tabulation of cases and person-years at risk. 22 Because histologic subtype was only known for about 40% of PLC cases, we could not directly calculate incidence rates of HCC and CC, but we were able to estimate the ratio of the two subtypes over time taking into account sex, age and city. Based on the assumption that PLC cases with known histologic subtypes were representative of cases with unknown subtypes, we applied these ratios to our Poisson model of PLC incidence from 1958 to 1987, which was adjusted for false-positive and false-negative DCO errors. Thus, we were able to calculate incidence rates of HCC and CC over this time period.
Statistical tests and 95% confidence limits were likelihood based. Analyses were performed using the Epicure software (Hirosoft, Seattle, WA). We arrived at the best fitting model by forward selection followed by backward elimination. Adequacy of the final fitted model was assessed by plotting the fitted incidence rates against parameter estimates for points based on grouping the continuous variables. All p-values presented in this report were two-tailed.

Death certificate accuracy
As shown in Table I, there were a total of 1,105 subjects whose death certificate causes of death were listed as liver cancer, chronic hepatitis, cirrhosis, pancreatic cancer or other diseases and for whom clinical or histologic records were also available for review. Only 51% of death certificates listing liver cancer as cause of death were confirmed by the pathology review. This percentage varied with exact cause of death: 89% of 98 deaths attributed specifically to primary hepatocellular carcinoma, hepatoblastoma or hemangiosarcoma (ICD 155.0) were confirmed as PLC, compared with 17% of 30 deaths attributed to primary cholangiocarcinoma (ICD 155.1) and 46% of 573 deaths attributed to liver cancer not specified as primary or with unknown histology. Figure 1 displays for the two sexes the probabilities of accurate death certificate PLC diagnoses by age and year of death. About two-thirds of death certificates for males correctly diagnosed PLC, compared with about one-third of those for females, a difference that was statistically significant (p Ͻ 0.001). The accuracy of death certificate PLC diagnoses improved significantly over the period 1958 to 1987 for both sexes, improving particularly rapidly in later years (the coefficient for year squared was positive: p ϭ 0.016). At the beginning of the study in 1958, 73% of males and 91% of females with death certificates specifying PLC as the cause of death were found to have died of other causes, but by 1987 these false-positive rates had declined to 9% and 26%, respectively. There was a significant interaction between gender and age (p Ͻ 0.001) with death certificates becoming more inaccurate with older ages at death for males but not for females.
As shown in Table II, we calculated by 2-to 6-year periods the percentages of subjects reported as dying from chronic hepatitis or cirrhosis or from pancreatic cancer whose true cause of death was PLC. Although we included as cases in the incidence analysis the 52 subjects with other death certificate diagnoses who were found FIGURE 1 -Accuracy of primary liver cancer (PLC) causes of death on death certificates by gender. Plots show the probability that when liver cancer was listed as the underlying cause of death it correctly identified a PLC. The curve on each grid shows the 50% accuracy level. Probabilities were estimated by logistic regression including measures of sex, age, and year of death in the following model: where the estimated parameters (95% confidence intervals) were as follows: intercept (␣ sex ): males, 0.53 (0.14, 0.93); females, Ϫ1.54 (Ϫ2.05, Ϫ1.07); age (␤ sex ): males, Ϫ0.066 (Ϫ0.092, Ϫ0.042); females, 0.0144 (Ϫ0.0132, 0.0432); year (␥ 1 ): 0.12 (0.09, 0.14); year 2 (␥ 2 ): 0.0038 (0.0007, 0.0068).  Yes  356  51  45  63  4  10  52  18  457  No  345  49  26  37  37  90  240  82  648  Total  701  100  71  100  41  100  292  100  1,105 during review to have PLC, there were too few subjects in any of these other diagnostic categories for us to adjust the numbers of PLC cases within calendar year strata for these errors. In the initial years of the study, small percentages of subjects who were diagnosed as dying from pancreatic cancer were found to have died from PLC, but by 1975 this diagnostic error no longer occurred. In all, 10% of such subjects were missed PLC cases. In contrast, high percentages of subjects whose death certificates said they died from chronic hepatitis or cirrhosis were found throughout the study to have died from PLC; about two-thirds of subjects with these diagnoses were missed PLC cases. In order to adjust DCO causes of death for false-positive and false-negative errors it was necessary for us to assume that the accuracy rate of cause of death determinations was the same for DCO subjects and for subjects whose death certificates were accompanied by additional clinical records. However, this was not always true. Among subjects whose death certificates recorded PLC as cause of death, reports of RERF autopsies, which were completed independently of death certificates, were available for 28%, and the pathology panel agreed that 54% of these subjects died of PLC. In comparison, reports of outside (non-RERF) autopsies were available for 12% of such subjects, and 77% were in agreement with the pathology assessment, a significantly greater proportion (Fisher's exact test, p ϭ 0.0002). This suggests that the results of outside autopsies, and probably other clinical information as well, were used to help determine causes of death. Thus, because clinical records were less likely to be available for DCO subjects, false-positive errors may be greater for them than we report.
For subjects with PLC for whom either cirrhosis or chronic hepatitis was listed as the cause of death on death certificates, the pathology panel found that official causes of death were more likely to be wrong for those receiving autopsies outside RERF (82% were in fact PLC), as opposed to receiving the independent RERF autopsy (66% were PLC), a difference that was not statistically significant (Fisher's exact test, p ϭ 0.23). This suggests that the availability of additional clinical information did not improve the accuracy of death certificates reporting cirrhosis and chronic hepatitis as cause of death.
Analysis I: PLC incidence after adjustment for false-positive death certificates Table III shows the adjustments we made to the numbers of DCO subjects to correct for errors in causes of death attributed to PLC. In the early part of the study, about half of all liver cancer cases were DCO cases, but after 1964 this percentage dropped, ranging from 20 -30% afterwards. As the false-positive rate declined over the study period, the proportion of liver cancer DCO cases retained in the incidence analysis correspondingly increased, from 7.7 of 49 (16%) in 1958 to 1964 to 28.4 of 36 (79%) in 1985 to 1987. Correcting for false-positive DCO errors, we estimated that a total of 89.7 (46%) of the 195 DCO subjects listed as having died from liver cancer were in fact PLC cases. Adding these cases to the 523 confirmed PLC cases, there were a total of 612.7 cases used in the Poisson regression incidence analysis depicted in Figure 2, which shows the incidence of PLC over time adjusting for age, city and gender. Figure 2 also illustrates the effect on incidence rates of our adjustment for false-positive DCO subjects and, alternatively, how these rates would look if DCO cases were excluded. Correcting for false-positive DCO causes of death reduced the incidence of PLC by about half in the early years of the study but had less effect in later years when confirmed cases predominated and DCO causes of death were more accurate. In the early years of the study, the effect of adjusting for false-positive DCO causes of death was similar to that of excluding DCO cases altogether, but beginning about 1975, death certificate accuracy was high enough that adjustment had little impact on incidence rates. However, even then there were a substantial number of DCO cases, so excluding all DCO cases rather than adjusting for false-positive errors would have downwardly biased incidence rates. Without this adjustment, the estimated incidence of total PLC would have been 8.1 ϫ 10 Ϫ5 in 1960 and 24.4 ϫ 10 Ϫ5 in 1985, a ratio of 3.0; in comparison, with these adjustments, the corresponding rates were 4.4 ϫ 10 Ϫ5 in 1960 and 23.7 ϫ 10 Ϫ5 in 1985, a ratio of 5.4. Thus, without adjustment for false-positive DCO cases, the increase in PLC incidence would have been underestimated by about 40%. On the other hand, excluding all DCO cases resulted in corresponding rates of 3.8 ϫ 10 Ϫ5 in 1960 and 21.4 ϫ 10 Ϫ5 in 1985, a ratio of 5.6. Thus, ignoring all DCO cases would have resulted in overestimating the increase in PLC incidence rates by about 4%.

Analysis II: PLC incidence after adjustment for false-negative death certificates
As shown in Table III and Figure 2, adjusting for PLC cases missed because of false-negative death certificate causes of death approximately doubled the number of PLC cases detected during the study. Figure 2 shows that the largest relative effect of this adjustment was in 1965-1980, when the adjustment tended to straighten this portion of the incidence curve, resulting in a poorer fit of the linear-quadratic model relating incidence to calendar year and a better fit of the quadratic spline model with two segments and joining point at 1980. 23 All incidence rates depicted in Figure  2 show a rapid increase in PLC incidence starting at about 1980, the time when more sensitive liver cancer diagnostic methods were introduced in Japan. After adjustment for both false-positive and false-negative DCO errors, the estimated PLC incidence was 7.4 ϫ 10 Ϫ5 in 1960 and 32.0 ϫ 10 Ϫ5 in 1985, a 4.3-fold increase. As shown in Figure 2, until 1975 our DCO excluded incidence curve was very similar to our false-positive DCO adjusted incidence curve. From 1975 to 1987 our adjusted curve was more similar to the curve derived from including, rather than excluding, DCO cases of PLC. To derive the incidence curve illustrated in Figure 3, we first excluded DCO cases of PLC diagnosed before 1975 and then supplemented the remaining cases with the PLC cases diagnosed from 1988 to 1994 but not reviewed by our pathology panel. The resulting incidence curve provided no evidence of further increase in PLC rates, and, in fact, these data suggested that PLC rates peaked in our cohort in 1993 and have declined since then. This decrease was not statistically significant, perhaps because it was based on only the final 2 years of case ascertainment.

Temporal changes in HCC and CC incidence rates
Of the 329 cases of PLC that had histological review by our pathology panel, 280 were hepatocellular carcinoma, 48 were cholangiocarcinoma and 1 was a hepatoblastoma; the other 11 cases were reviewed by pathologists not on our review panel. Using our best estimate of PLC incidence, which includes adjustment for both false-positive and false-negative DCO errors, the observed rates of HCC and CC were fit by the model described in Table IV. All terms included in the model were significant at the 0.05 level except the sex-by-ln(age) term, which was included because it was highly significant in the overall PLC incidence model. The model showed that HCC and CC had significantly different relationships with regard to sex, log age at diagnosis and year of diagnosis. The relative risk for males versus females was 8.07 for HCC and 2.25 for CC, a gender difference that was statistically significant (p ϭ 0.002). The key finding was that whereas HCC incidence rates significantly increased over time (relative risk: 1.05 per year; p Ͻ 0.001), CC rates declined slightly but not significantly (relative risk: 0.99 per year). This finding is also illustrated in Figure 4, which graphs the approximated incidence rates for HCC and CC by calendar year. The crude rate of HCC was 6.4 (95% confidence limits: 4.51, 9.51) times higher than that of CC (p Ͻ 0.001). The close similarity of the lines for HCC and PLC reflect the fact that most PLC in this study was HCC. DISCUSSION Our results show that death certificates listing liver cancer as cause of death have been highly inaccurate. Our results are in   agreement with those of Percy and associates, 9 who found that the more specific the ICD code, the greater its accuracy. In their analysis of 1973 to 1985 SEER data, they found that 83% of deaths attributed to ICD code 155.0 (primary liver cancer) were correct. In comparison, when we restricted our study to such subjects diagnosed in the same years, we found that 92% of death certificates were correct, suggesting that the situation was similar in the United States and Japan. We found that death certificate causes of death for males were significantly more accurate than those for females. One possible explanation for this gender difference in error rates relates to the previously noted observation that the male-to-female ratio of HCC is consistently higher in studies in which causes of death are confirmed. 7 Because of the increased incidence of PLC in males, the more even male-to-female sex ratio for tumors with liver metastases results in a higher proportion of non-PLC and DCO errors for females. 7 Another possible explanation for the difference in death certificate error rates for males and females is the propensity for ovarian tumors to metastasize to the liver, 24 combined with the difficulty clinicians have in distinguishing ascites associ-ated with PLC from malignant ascites associated with ovarian cancers. 25 Our results show that false-positive DCO errors almost doubled the estimated incidence rate of PLC in 1960 but declined steadily over time and were negligible by 1985. We conclude that studies examining pre-1975 PLC incidence in Japan and other developed countries need to take these errors into account. In our study, adjusting for these DCO errors had approximately the same effect as discarding pre-1975 DCO cases of PLC altogether. A recent French study of HCC followed a similar strategy, disregarding pre-1980 death certificates because of high error rates. 26 We found that the death certificate false-positive rate for PLC causes of death was significantly reduced when autopsies were performed by outside physicians and information was available for cause of death determinations. The error rate was higher when autopsies were performed by RERF after cause of death was determined, indicating that the availability of clinical information improves PLC cause of death accuracy. Thus, our false-positive results should be conservative, meaning that DCO false-positive rates for PLC are even greater than we report. Our results suggest that the availability of such clinical data does not improve the false-negative rate of death certificates for detecting PLC, possibly because of the difficulty in distinguishing chronic hepatitis and cirrhosis from PLC as a cause of death.
By comparing diagnostic information to the cause of death listed on the death certificate, our approach implicitly assumes that for subjects with PLC who died, liver cancer was their underlying cause of death. Without this assumption, it would not be possible to use clinical or surgical diagnostic results to assess the accuracy of death certificates. Our assumption is supported by the fact that PLC is rapidly fatal: in an earlier study of A-bomb survivors, we found that the mean time of survival from diagnosis with HCC, which accounts for about 85% of PLC in this cohort, ranged from 5.6 months for cases with cirrhosis to 8.9 months for cases without cirrhosis (unpublished data). Given the rapid time to death associated with this disease, it seems reasonable to assume that liver cancer was the underlying cause of death for persons who died with it.
Although the false-positive error rate declined over time, as of 1987, we found no decline in the proportion of PLC cases that were falsely attributed to cirrhosis or chronic liver disease on death certificates. However, our latest percentage (1985)(1986)(1987) was based on just six cases and may not be accurate. Overall about two-thirds of such diagnoses were in fact PLC. Because of the 1 Subtype-specific incidence rates and relative risks for the intercept and risk factors are presented because they were significantly different (p Ͻ 0.001) for HCC and CC. Plotted rates represent rates for mean the age of 48.4 years for males and 51.4 years for females for 1971, the mid-year of the study.-2 Although not significant, this term was included because the gender ϫ ln(age) interaction was statistically significant in the model for PLC overall. large numbers of deaths attributed to these causes, this was the major source of death certificate error in our study and adjusting for it approximately doubled the estimated number of PLC cases. The problem of clinically distinguishing cirrhosis from HCC is well known, and investigators continue to search for ways to solve it. 10,11 Our findings are in agreement with an earlier study of A-bomb survivors that found that only 33 (55%) of 60 cases with RERF autopsy diagnoses of liver cancer had death certificates recording these as the cause of death. 8 In our study we estimated that only 612.7 (54%) of 1132.6 cases of PLC had death certificates recording the correct cause of death.
The PLC incidence curves presented in Figure 2 show a particularly rapid increase in the PLC incidence beginning in the early 1980s, when improved liver cancer screening methods of ultrasound and CT scanning were introduced in Japan. Figure 5 illustrates how PLC incidence rates would look for our cohort if we were to extrapolate backwards from the average incidence for the last 3 years of follow-up (1985)(1986)(1987) using the rate of increase prior to 1980 (i.e., the slope of the lower segment of the regression spline).
In our study the effect of ignoring DCO cases of PLC until 1975 when the false-positive rate had declined and including later DCO cases generated incidence rates very similar to those calculated using our adjustment procedure. Taking into account false-negative death certificates as well as false-positive ones further decreased the ratio of PLC incidence between 1960 and 1985 from 5.4 to 4.3, which was higher than the 3.0 ratio generated when no adjustments were made. Taking into account diagnostic methods introduced in the early 1980s further reduced this ratio to 2.4. A Danish study, which compared PLC incidence from 1948 to 1952 with incidence from 1978 to 1982, found that adjusting for diagnostic errors reduced the relative increase over this time from 3.4 to 1.4. 27 Recent studies conducted in the United States, 1 United Kingdom, [3][4][5] France 26,28 and Japan 29 have examined changes in liver cancer incidence occurring from the mid-1970s to the mid-1990s without considering or only partially considering the changes in death certificate and diagnostic accuracy occurring over this time period. Our study suggests that in Japan and probably in other developed countries, PLC incidence rates up to the mid-and late-1970s were concurrently inflated by the erroneous classification on death certificates of metastatic cancer as PLC and deflated by the omission of PLC cases missed until the introduction of new diagnostic methods in the early 1980s. A high proportion of deaths due to PLC were misdiagnosed as cirrhosis or chronic hepatitis; these errors continued until our pathology review ended in 1987 and did not appear to decline with the advent of improved diagnostic methods in the early 1980s.
Since we found that temporal changes in the incidence of HCC and CC were very different, changes in PLC incidence rates over time will also reflect the histologic makeup of PLC cases in each country. It should be noted that our study is unlike previous investigations in that we studied a single cohort: residents of the 1945 atomic bombings who were alive in 1950 to 1952. Thus, the secular changes observed in this cohort may not be directly relevant to the secular trends observed in other studies.
We found a slight downward curvature in the pre-1980 portion of the incidence curve, so that removing the further, post-1980 increase in incidence resulted in a leveling off of incidence rates around 1985; lack of further increase in incidence was confirmed by analysis of the PLC cases reported to the tumor registries between 1988 and 1994. Because post-1987 cases were not reviewed, we were unable to adjust rates for false-negative DCO errors. However, this is unlikely to explain the post-1990s decline in HCC incidence rates that we observed since false-negative cases would have been added at all time points, and if anything, fewer would be added in later years with the advent of better diagnostic methods. The pre-1990 increase in the incidence of HCC that we report is consistent with the previously reported increase in Japanese HCC rates from 1968 to 1984 30 and with the increase from 1975 to 1995 reported by the Japanese National Cancer Center. 6 Nishioka and associates 30 found that the incidence rate of HCC patients who were positive for the hepatitis B surface antigen remained constant from 1968 to 1984 and that in 1990, of 105 HCC cases testing negative for hepatitis B in five areas of Japan, 76% were infected with the hepatitis C virus.
In our study of hepatitis C virus infections among A-bomb survivors with HCC, we found that the prevalence of this virus increased from 14.3% for cases dying from 1961 to 1967 to 44.7% for cases dying from 1961 to 1987 (unpublished data). Thus, it is likely that the increase in HCC rates we report is attributable to increased numbers of patients infected with the hepatitis C virus, not the hepatitis B virus. It is not clear whether the post-1990s leveling off and slight decline in HCC rates that we report are a function of decreased numbers of hepatitis B virus-infected or of hepatitis C virus-infected cases. Since the incidence rate of CC remained constant throughout our study period, the incidence of this disease does not appear to be greatly affected by changes in the prevalence of either virus.
We conclude that changes in false-positive and false-negative cause of death errors on death certificates and in liver cancer diagnostic methods have a strong impact on recorded liver cancer incidence rates and, specifically, on rates of HCC and CC. Taking these errors into account, HCC incidence rates rose substantially in our Japanese cohort from the late 1950s to the mid-1980s. The relative increase ranged from 2.4-to 4.3-fold, depending on how much of the post-1980 increase was attributable to improved diagnostic methods. The incidence of HCC peaked in our cohort around 1993 and appears to have declined since then. In contrast, the incidence of cholangiocarcinoma remained constant from 1958 to 1987. FIGURE 5 -Estimated incidence of primary liver cancer (PLC) from 1958 to 1987, assuming that the post-1980 increase in excess of the annually increasing increment was due to improved diagnostic methods. The dotted line indicates the incidence of primary liver cancer adjusting for false-positive and false-negative causes of death for death-certificate-only subjects. The solid line shows how this incidence curve would look if the cases that were presumably missed because of poorer diagnostic methods were added at 1985 and earlier time points.