Scottish Government

Cancer in Scotland: Sustaining Change
Cancer Incidence Projections for Scotland (2001-2020)
An Aid to Planning Cancer Services

INTRODUCTION AND SUMMARY

It is predicted that there will be over 168,000 cases of cancer diagnosed in adults in Scotland during 2016-20 (around 33,700 cases per annum) compared to around 26,400 cases per annum actually recorded during 1996-2000. Most types of cancer are predicted to increase in numbers of new cases. Notable exceptions are stomach, lung and cervical cancers, all of which are predicted to decline. Overall, it is predicted that there will be a 28% increase in the number of people diagnosed with cancer over the next 20 years (approximately 1.4% per year).

An increase of around 650 cases a year is expected in the under 75's and 1,200 in those aged 75 and over. Most of the predicted increase is attributed to a growing number of elderly people in the population. However, the individual risk of some types of cancer is also predicted to increase.

MATERIALS

Data from the Scottish cancer registry (extracted April 2004) were used to examine trends in historical cancer incidence rates (1961-2000). Population estimates (1961-2001) and 2002-based population projections (2002-2020) were obtained from the General Register Office (Scotland). All malignancies excluding non-melanoma skin cancer (C00-C96 excluding C44) were considered. The most common 19 malignancies were considered separately (covering 90% of all malignancies), and the remainder combined into a category labelled "Other and unspecified". Only persons aged 35 or more years were included in the analyses, with the exception of melanoma skin cancer (ages 20+), cervical cancer (25+), testicular cancer (20+) and Hodgkin's disease (15+). Due to a recent (late 1990's) change in the coding of bladder cancers it was necessary to combine malignant, in situ and uncertain behaviour bladder tumours.

METHODOLOGY

Generalised linear models (so called age-period-cohort or APC models (Clayton and Schifflers, 1987)) were fitted to the observed data. The following models were considered: age effect only; age effect+linear period effect or 'drift'; age+period effects; age+cohort effects; full age+period+cohort effects model. The models were analysed for each site and gender and the best (most parsimonious) model was then used to make predictions for the future using the GRO(S) population projections.

The models were fitted using five year age groups (35-39, ...., 80-84, 85+) and five year time periods (quinquennia: 1961-65, ....., 1996-2000). Overlapping birth cohorts were then calculated as based on the age groups and time periods. Earlier age groups were considered for testes, uterine cervix, malignant melanoma and Hodgkin's disease. The modelling was performed using SAS Release 8.02.

In applying the best model to the future time periods an assumption was made that the age effects would remain the same in the future, so that the fitted age parameters could be used for the projection. Future period and cohort effect parameters were considered in six different ways (these are fully described in the Cancer Scenarios report (Scottish Executive Health Department, 2001)). Three of these allowed drift, the linear trend by period or cohort, to continue into the future. The other methods did not allow period and cohort trends to continue unchanged into the future.

As noted in Cancer Scenarios, the limitations of the procedure used here must be stressed. Statistical predictions for values outside the range of observed data used in a model should always be interpreted with caution. There are theoretical reasons why the predictions may be fallible and, in addition, changes may occur in the future which are independent of the historical data and so cannot be identified in the statistical analyses. These include changes in the exposure of the population to risk factors either in the new time periods or in birth cohorts who were too young to be included in the analyses.

RESULTS

Overall, it is predicted that there will be over 168,000 cases of cancer diagnosed during 2016-20 (around 33,000 cases per year) (see Table 1 and Figure 1). An increase of around 650 cases a year is expected in the under 75's and around 1,200 cases a year in those aged 75 and over ( see Tables 2a and 2b). Most of predicted increase is a result of a growing elderly population ( see Figure 2).

1 ^{This does not take into account a short to medium term increase in breast cancer cases expected as a result of extending the Scottish breast screening programme upper age range from 64 to 70.}

DISCUSSION

The numbers of new cases of cancer seen in a population are products of individual risks of developing cancer and the numbers of individuals actually at risk. The present analysis suggests that the dominating feature determining cancer incidence in Scotland in the next twenty years will be growing numbers of older people. Although an individuals' risk will increase for many cancers, there are notable exceptions: lung cancer risk is continuing to decline in men and is expected to begin to decline very soon in women; cervical cancer incidence is expected to fall further as screening and potentially other preventive measures become more effective; finally, the reduction in risk in stomach cancer is part of a long-term trend which has been seen in many developed countries.

The projections presented in this report are clearly of practical significance to the planning of cancer services in Scotland. How robust are the data? Appendix A describes a validation exercise in which the methodology was applied to historical data for 1961-1990 in order to estimate incidence for 1991-2000. The latter was then compared to the incidence actually observed for this period. Overall, the results were encouraging for all cancers in aggregate: the estimates were within 2% of the observed data. However, for some of the less common cancers the estimates and observed incidence differed by 30% or more. This tended to be due to interventions which occurred after the period of the observed data used in the model. For example, improvements in the cervical screening programme in the early 1990s. For the more common cancers such as breast, lung and colon-rectum, the results are more satisfactory. Therefore we believe that use of the projections in aggregate and for the major cancers is suitable for planning purposes, but would urge caution in use of data for less common cancers in isolation.

Advances in the methodology for calculating projections have been made since the preparation of the Cancer Scenarios report in 2000. A similar project was undertaken in New Zealand (New Zealand Ministry of Health, 2002). This suggested that the APC models used in the analyses presented here can give projections that are too high for cancers where the rate is increasing relatively fast. This will be investigated before the next set of projections are published. Specifically, the New Zealand approach involved use of:

• Generalised linear models (APC models as used in this report; Clayton and Schliffers, 1987) with drift applied using Osmond technique (Osmond, 1985)

• Non-parametric generalised addative models (GAM) (Wahba, 1990; Hastie and Tibshirani, 1990)

• Non-linear models (Dyba and Hakulinen, 2000)

The projections would then be based on all the models passing the initial testing phase using a model averaging process. This would comprise of the mean of the modelled number of cases of the models selected and fitted for each cancer site. In the New Zealand report this average gave a better fit than any of the component models in their empirical testing.