The number of cancer deaths in the United Kingdom by cancer site, sex, 5-year age group and year for the 35-year period from 1971 to 2005, and population numbers for the same years and age–sex groups were provided by Cancer Research UK on the basis of separate national data from England, Scotland, Wales and Northern Ireland.

The R-based software, ‘Nordpred', developed at the Norwegian Cancer Registry for projections of nordic incidence and mortality rates (

Møller *et al*, 2002), was used to project mortality rates in 2006–2025. The calculations are on the basis of the mortality rates from 1971 to 2005, aggregated into 5-year time periods from 1971–1975 to 2001–2005.

Future numbers of deaths were projected by multiplying the projected age-specific rates by the population projections for 2006–2025.

For projecting future rates the Nordpred software uses the age-period-cohort (APC) models, deriving the relevant parameters from the past observations and using them to estimate future rates. Here, the age, period and cohort function as pseudo indicators for factors that have influenced past trends, such as exposure to risk factors, treatment or screening affecting certain age groups, periods or cohorts. Within the basic APC model, the Nordpred software allows the choice of several options for modelling and projecting incidence and mortality rates. For this study, we used the model found by the Nordic research group to give the most accurate predictions of future numbers of 20 different cancer types (

Møller *et al*, 2003). Thus, instead of using the standard exponential link function in poisson regression, the power of 5 was used. The reasoning behind this is that the power function levels of a potential exponential increase (so as to keep projections closer to the observed rates). The model can be written as

where

*R*_{ap} is the mortality rate for age group

*a* and period

*p*,

*A*_{a} is the age effect of age group

*a*,

*D* is the common linear drift of period and cohort,

*P*_{p} is the non-linear period effect and

*C*_{c} is the non-linear cohort effect, where

*c*=

*p*−

*a*. Note that the 5-year periods are numbered consecutively so that

*p* increases by 1 every 5 years; similarly,

*a* counts 5-year age groups.

Those age groups for which less than 50 deaths were observed during any 5-year period were not included in any period in the modelling.

The age component was projected directly when estimated. For age groups not included in the model, the mean observed mortality for 1996–2005 was used for projections.

The linear drift *D* was projected into the future, but with two modifications. The so-called cut trend system was used, in the sense that instead of adding *D* for each future period, we added *D*, 3/4 *D*, 1/2 *D* and 1/4 *D* in the four future 5-year periods. The reasoning behind this was that whatever might have caused the increase or decrease in the past is not likely to continue unchanged indefinitely; rather some attenuation of such influences seems more plausible. In addition, Nordpred checks for significant curvature in the trends of the observed rates, and if this was present, the change in the last 10 years (recent slope), instead of the average change in the whole period (average slope), was used as the drift component in the future projections. The rationale is that the factors responsible for the most recent trends (when there has been a change) are more likely to be an influence on future rates, rather than those operative in the more distant past.

The non-linear cohort component was projected for estimated cohorts. For young cohorts not estimated, the last estimated cohort component was used.

The projections assume that the last non-linear period component applies to all future periods.

To estimate future rates for all cancers combined, the ‘other cancers' category (i.e., other than the 21 sites studied) was created, and projected rates were calculated, as for the other sites. The total (‘all cancers') was estimated as the sum of the numbers of deaths at the individual sites (including ‘other cancers') divided by the population numbers.

The projected numbers of deaths were calculated using the projected mortality rates and the population projections from the

Government Actuary's department (2007). In the tables, the numbers for 2023 are presented using the average population projections and the mortality rates for 2021–2025. The numbers for 2003 are presented using the average observed numbers for the corresponding 5-year period, 2001–2005.

Mortality rates are presented as age-standardised rates, according to the European standard population (

Doll and Smith, 1982).

In 1984, the rules used to interpret secondary causes of death changed. In 1993, the rules from before 1984 were used again, and, in addition to this, the extent to which non-specific causes of death automatically generated medical enquiries changed. Estimates of the effects of these changes on disease-specific mortality rates have been published (

Office of Population Censuses and Surveys, 1996). Age-specific mortality rates for the period 1971–1992 were adjusted to the coding practices from 1993 onwards using the factors in Table A.3 and B in the publication mentioned above (

Office of Population Censuses and Surveys, 1996).