The observed data were annual incidence rates, for GB, by cancer site, sex and 5-year age group from 1975 to 2007. The rates were converted to numbers of cases for the United Kingdom by multiplying by the ratio of the respective populations. Compared with the actual incidence (number of recorded cases) in the United Kingdom in 2007, the difference between observed and modelled numbers of cases (using GB rates applied to the UK population) was <1.5% for all sex/site combinations. In , we present the projected changes in ASRs and in the numbers of cases of cancer between 2007 and 2030. These can be used to split the change in numbers between changes due to changing rates and changes due to changing population. Thus, for instance, the 11% increase in male stomach cancer cases is accompanied by a 33% fall in rates, so that the effect of the population change is to increase numbers by 66% (=[100+11]/[100−33]−1).
We have deliberately shown results on a linear rather than a logarithmic scale because of the greater uncertainties in predicting cancers in the young. For ovarian cancer, for instance, cancers in young women are likely to be germ cell with very different risk factors from epithelial cancer in older women; therefore, there is no reason to suspect that cohort effects seen in those aged 15–29 years will carry forward to ages 35–49 years. Rates of melanoma, ovarian cancer, endometrial cancer (corpus uterus) and testicular cancer (not shown in ) are all non-negligible in those aged 25–49 years, and our model is seen to fit these observed data well. For other sites, one would need to more carefully adjust the model used if the interest was specifically on rates in the population in this age group.
Future predictions depend on multiple assumptions, but the basic premise is that past trends, affecting as they do the risk of cancer in specific generations and/or time periods, will be carried forward into the future. We modified this basic assumption in one important respect – we attenuated the ‘drift' component of the observed changes in rates by 8% each year; thus in 2017–2018 years the annual drift is just 43% of what it was between 1975 and 2007, and by 2030 it was just 15%. The idea of attenuating the drift, so as not to assume that increases or decreases continue forever, was proposed and shown to be empirically useful by Møller et al (2002)
, although whereas they used arithmetic damping, we chose to use a geometric damping.
We compared our results with those based on projections using the Nordpred package and its default power-5 link. The Nordpred package requires input of past data for 5-year time periods, and produces projections for periods of the same duration. The ‘drift' component in Nordpred is reduced by 25% for each 5-year period after the first. The results are a little different from those based on our modelling using single years of observation, and the magnitude of the deviation is similar to that seen when current observed incidence rates are compared with those projected from past data using Nordpred. We also allow the reader to compare our model projections using power-5 and exponential link functions. The latter is used for comparison because it is the default for Poisson regression and because the age and period effects can be interpreted as relative risks. It is not used for the main projections because it can lead to extreme results, as the model is extrapolated further into the future. It is of course impossible to judge in advance which of the many projections (by sex, age group and cancer site) will prove to be the more accurate.
The uncertainty associated with these predictions does not concern sampling error (which would be small when based on the relatively large numbers of cases under study), but the unquantifiable bias when trends in some cancers behave in a manner that is inconsistent with the assumptions of the statistical analysis of past rates.
In the 24-year period 1984 to 2007, the overall age-standardised incidence rate for cancer has increased in GB in both sexes, although in the most recent 10-year period (1998–2007) they have remained fairly constant (). Male rates increased slightly but have now returned to those of the late 1990s level. Female rates have increased slightly (by around 3%). Future projections suggest that the decline in male rates will continue, whereas rates in women will peak, then start to decrease during the current decade (2010–2019). The reduction in smoking prevalence has been associated with declines in lung cancer rates in males throughout the period, and, from around 2004 in females (Cancer Research UK, http://info.cancerresearchuk.o
, accessed December 2010). Our projections implicitly assume that the reduction in smoking will continue, but do not explicitly model smoking rates or take account of the most recent changes in smoking prevalence. Indeed, changing rates of smoking over the last decade will continue to affect lung cancer rates into the next decade due to the cohort effect of smoking cessation – ex-smokers have lower rates of lung cancer compared with current smokers even (particularly) many years after cessation. Nevertheless, whereas the reduction in male smoking in the United Kingdom in the last quarter of the twentieth century was dramatic, the more recent changes have been more modest and while the rates of lung cancer are predicted to continue to fall, the numbers will begin to increase () as the population grows and ages. Similarly, we do not consider the likely impact of colorectal screening, HPV vaccination or of accelerated changes in obesity. The predicted impact of fecal occult blood testing on colorectal cancer incidence is small, as this form of screening is intended primarily to diagnose cancer early. Nevertheless, the lead time from screening will change the age-specific rates (Parkin et al, 2008
), and with the introduction of flexible sigmoidoscopy it is likely that the bowel screening programme will result in a noticeable fall in the incidence of distal cancers. The impact of HPV vaccination (introduced in women aged 12–18 years in 2008) on cancer rates by 2030 will be quite small, as the total burden of HPV-related cancers by age 40 years is small relative to the lifetime burden of cancer in women.
After a long period of increasing incidence, age-standardised incidence rates of breast cancer have begun to decline since 2005, and this trend is projected to continue, reflecting in particular declines in the age group 55–64 years. Some (if not all) of the decline in breast cancer in this age group is undoubtedly due to the substantial reduction in the use of HRT in recent years (Parkin, 2009
Other notable changes in rates include the continued steady fall in stomach cancer in both men and women; the increase in number of rarer cancers including melanoma, non-Hodgkin's lymphoma, kidney, liver and orophaynx in both men and women; the increase in oesophageal cancer particularly in women; and the increase in cancer of the corpus uterus. Although it is not possible in a paper such as this to consider in detail the likely reasons for past changes in rates of particular cancers or the likely impact of changes to the health service over the next decade, the standard approach for obtaining projections at any site can be used as a basis for more detailed epidemiological study of such trends one site at a time.
Møller et al (2007)
published projections of cancer incidence for England, for the period 2004–2020, based on observed rates in 1974–2003, using the Nordpred package. Although incidence rates in men were projected to decline by 7% between 2001 and 2020 (driven largely by the declining rates of lung cancer), female rates were projected to increase by the same amount, with significant increases in the risk of breast cancer in particular. Apart from the somewhat different populations studied (our projections were based on rates from GB, not just England (86.2% of the British population in 2007)), the availability of 4 years additional data (2004–2007) and the slight differences in methodology, there are two important sources of difference. First, Møller et al
‘projected' the future incidence of cancer of the prostate by assuming that rates would remain at the level observed in 1999–2003. This almost certainly results in an estimate of future burden of cancer of the prostate that is too low, and, as this cancer accounted for almost one-quarter of all new cancer cases in men in 2007, an underestimate too of the overall cancer burden. There is considerable uncertainty in predicting prostate cancer incidence, which is being driven not only by an inherent increase in the risk of the disease, but also by the over-diagnosis (and over-treatment) as a consequence of testing with PSA. There is little information available on the extent of PSA testing in the United Kingdom. In Scotland, the PSA testing first came into use in 1989, and the rate of testing accelerated rapidly after 1991 (Brewster et al, 2000
); the slowly increasing rates of incidence in the United Kingdom greatly accelerated at this time (). We have attempted to capture the underlying increase in incidence (pre-testing) and assumed a testing effect remaining as it was in 2004–2007. This is of course almost certainly wrong, but, equally certainly, likely to provide a more realistic future estimate than the assumption of no change, as we do take into account the underlying increase in rates and we believe that PSA testing, which is currently much less common than in the USA, will most likely increase. Second, Møller et al (2007)
projected forward the trends in breast cancer incidence observed in 1994–2003. The changes in incidence in recent years due to patterns of breast cancer screening by age, and the decrease in use of HRT (Parkin, 2009
) mean that their assumption of a continuing increase in incidence rates (~1% annually) is almost certainly too pessimistic. In our prediction, the age-standardised incidence for breast cancer will decline by 8% between 2007 and 2030.
Projections of cancer incidence, although inherently subjective and unreliable, do provide a necessary baseline for future planning of cancer resources and against which preventive interventions can be judged. The methodology used here minimises the subjectivity and provides a framework for such planning.