Trends in mortality for cancers, comparing multiple- and underlying-cause rates, in an English population 1979–1999

In compiling official mortality statistics, rules for selecting the underlying cause of death have changed twice in the last 20 years in England. Mortality statistics for most types of cancer were not greatly affected, but there were significant effects on coding for cancers of colon, liver, breast, prostate, testis and bladder, and for lymphoma and leukaemia.


METHODS
The former Oxford National Health Service Region covered a total population of about 2.5 million people. All causes of death on each death certificate in this population were coded for the years 1979 -1999, for the Oxford record linkage study (Goldacre et al, 2000), using the 9th ICD revision codes. In each year, the underlying cause of death was selected and coded according to the prevailing national rules. Following convention, all certified causes of deathunderlying cause plus causes elsewhere on the certificate -are termed 'mentions'.
The database was searched for records with mentions of malignant neoplasms . We analysed the data in three time periods defined by changes to coding rules, 1979 -1983, 1984 -1992 and 1993 -1999, to calculate the percentage of mentions that were coded as the underlying cause of death in each period. We tested changes in these percentages between time periods by w 2 tests on the numbers on which the percentages were based. We then calculated changes over time in annual mortality rates, standardised by applying the age-specific rates in the Oxford region population in each year to the European standard population. We calculated the average annual percentage change over time in mortality rates for mentions by fitting linear regression models to the logarithms of the death rates. We calculated changes in rates for mentions using every individual year. We calculated rates for underlying cause using individual years before and after the rule 3 changes, that is, omitting data for 1984 -1992. We omitted these years to determine whether, if the discontinuity in the application of the rule was ignored, trends in mentions and underlying cause were similar. Table 1 shows the extent to which underlying cause statistics underestimate the total mortality ascribed to cancers. For example, in 1993 -1999 mortality from all cancer was 10% higher, and that from breast cancer was 21% higher, than that judged from data on underlying cause alone.

RESULTS
Comparing periods with different coding rules (Table 1), the ratio of mentions to underlying cause showed statistically significant and noteworthy differences for cancer of the colon, liver, breast, prostate, testis, bladder, Hodgkin's disease, non-Hodgkin's lymphoma, lymphoid and myeloid leukaemia (we defined noteworthy as a significant change in the ratio of more than 5%). Statistically significant, but numerically smaller, changes in the ratios between the periods with different coding rules were seen for cancers of the oesophagus, pancreas and lung. Changes in Table 1 Mention-based mortality rates per 100 000 population; ratio of death with a mention of each cancer to death certificates with the cancer as the underlying cause of death; and average annual change in mortality rates with 95% confidence interval (95% CI in italics)
Trends in mortality for cancers MJ Goldacre et al the ratios coinciding with rule changes can be interpreted using prostate cancer as an example: between 1979 -1983 and 1984 -1992, the ratio of mentions to underlying cause decreased from 1.39 to 1.19, or, in other words, in the later of the two periods a higher percentage of certificates that included prostate cancer were coded with prostate cancer as the underlying cause. This meant that underlying-cause mortality rates for prostate cancer (the only rates available from routine statistics) exaggerated the rise from 1979 -1983 to 1984 -1992. The subsequent increase in the ratio of mentions to underlying cause, from 1.19 in 1984 -1992 to 1.30 in 1993 -1999, means that a lower percentage of cases with prostate cancer were coded with the disease as the underlying cause in the later than in the earlier of these two periods. In this example, the trends for underlying-cause and mentions converge comparing the first and second periods, and diverge comparing the second and third periods (Figure 1). Age-standardised death rates, expressed as mentions per 100 000 resident population, showed a significant increase over time for cancers of the oesophagus (in men only), liver, malignant melanoma, prostate, kidney, and non-Hodgkin's lymphoma. Significant decreases over time were found for cancers of the lip, stomach, colon, rectum, gall bladder, pancreas, lung (in men only), breast, bladder, Hodgkin's disease and non-Hodgkin's lymphoma. Omitting the middle time period, these trends based on underlying cause were similar to those based on mentions.

DISCUSSION
The ratios of mentions to underlying cause show the extent to which statistics for underlying cause, alone, have underestimated the population toll of cancer mortality. For certain cancers, studies of trends in rates that use underlying cause only, starting or finishing in years between 1984 and 1992, may be misleading. In general, however, the impact of the rate changes on cancer was small. Particularly, if data for the years 1984 -1992 are ignored, rates show similar trends whether analysed for mentions or for underlying cause. As we have shown, even over a period as short as two decades, many cancers show significant upward or downward trends.
Consideration needs to be given to the distinction between dying with a cancer and dying from a cancer. Particularly for cancers in which therapy has been successful, the certifying doctor may consider that the cancer, though appropriate to certify as a contributing cause, was not the underlying cause of death. As treatment improves, for some cancers it can be expected that there will be a trend away from certifying the cancer as the underlying cause of death in affected people who die. Testicular cancer and Hodgkin's lymphoma are examples (Table 1). Trends in mortality statistics for cancer, based on underlying cause alone, may be particularly difficult to interpret when the period studied coincides with changes in practice in the selection of the underlying cause.

CONCLUSION
Multiple-cause coding of mortality in England has become routine since 1993. In analysing cancer mortality rates, consideration should be given to rates based on mentions as well as underlying cause, particularly when the analysis crosses periods of change to selection rules.