Incidence and mortality trends for prostate cancer in 5 French areas from 1982 to 1996
Article first published online: 9 OCT 2001
Copyright © 2002 Wiley-Liss, Inc.
International Journal of Cancer
Volume 97, Issue 3, pages 372–376, 20 January 2002
How to Cite
Chirpaz, E., Colonna, M., Menegoz, F., Grosclaude, P., Schaffer, P., Arveux, P., Lesec'h, J. M., Exbrayat, C. and Schaerer, R. (2002), Incidence and mortality trends for prostate cancer in 5 French areas from 1982 to 1996. Int. J. Cancer, 97: 372–376. doi: 10.1002/ijc.1603
- Issue published online: 27 DEC 2001
- Article first published online: 9 OCT 2001
- Manuscript Accepted: 29 JUN 2001
- Manuscript Revised: 23 APR 2001
- Manuscript Received: 13 OCT 2000
- Association Grenobloise d'Aide à la Recherche en Oncologie
- prostate cancer;
- incidence and mortality trends;
- age-period-cohort effects
After an increase in the 1980s, incidence and mortality for prostate cancer in North America or England and Wales started to decrease in the early 1990s. The reasons for this evolution are widely debated, notably the importance of early detection. This study describes trends of prostate cancer incidence and mortality in 5 areas in France, where practices of early detection for this cancer are widely used. The 5 French administrative areas, covered by a population-based registry, have a total population of approximately 1,700,000 men. Incidence data from these registries were studied for the period 1982–1995, and mortality data were provided by the Institut National de la Santé et de la Recherche Médicale (INSERM) for the period 1982–1996. Age-Period-Cohort models by Poisson regression were created to characterize these trends. Between 1982 and 1995, 14,699 cases of prostate cancer were registered by the 5 registries under consideration. After a little intensification of the increase in 1987, undoubtedly due to early detection (notably using Prostate-Specific Antigen), the trend of the incidence seems to reverse from 1993. Mortality increased monotonically from 1982–1990 by an average of 1.8% per year, before decreasing annually by an average of 3.3% until 1996. Poisson regressions indicated a period effect on both incidence and mortality data; a small, but significant, cohort effect exists for incidence evolution, showing that elements such as etiologic factors may have an influence. Until results of randomized studies on mass screening are available, the question of individual screening remains; improved knowledge of risk factors could be interesting. © 2001 Wiley-Liss, Inc.
As in many developed countries, adenocarcinoma of the prostate is the most commonly diagnosed cancer among French men. In 1995, incidence (crude rates) was estimated at 93.7 per 100,000 person-years (PY), while lung and colorectal cancer are the second and third most common cancers with rates of 66.2 and 64.1 per 100,000 PY, respectively.1
Prostate cancer became the most frequently diagnosed cancer in males after an important increase in detection, notably in relation with the development of screening by serum Prostate-Specific Antigen assay (PSA) and digital rectal examination.2–5 This screening remains controversial for several reasons: the main reasons are incomplete knowledge regarding the natural history of this disease,6 the problems linked to the uncertainties concerning treatment indications in the early stages of the disease6–8 and the fact that this screening has not proved effective in terms of public health.9, 10 To our knowledge, only 1 randomized study has been published on the effect on mortality of mass screening for prostate cancer11 in which the authors found a decrease of 67.1% in prostate cancer mortality in the screened group compared with the control group. This study has been largely criticized because of the low participation rate in the screened group, which leads the authors not to analyse results in intent-to-screen.
Recently, some studies performed on the SEER Program data or on other population-based registries in the USA have shown a decrease in age-standardized incidence rates of prostate cancer, beginning in the early 1990s.12, 13 This decrease in incidence was preceded by a decrease in specific mortality in the USA (SEER Program12). Similar patterns concerning the mortality have been observed in Canada14 and in England and Wales.4 Even though causes of this evolution remain controversial, particularly the influence of PSA screening,4, 12, 14–17 we decided to study trends of prostate cancer incidence and mortality in a French population, where individual screening by PSA has been widely used since the mid-1980s.
POPULATIONS AND METHODS
This study was carried out in the populations of 5 French administrative areas covered by population-based registries (Bas-Rhin, Calvados, Doubs, Isére and Tarn), covering in 1989 (mid-year of our period study) nearly 1,700,000 men (representing approximately 6% of French men). From the 10 general cancer registries of France, we chose the 5 that were able to give us valid incidence data from 1982, i.e., approximately 5 years before the introduction of PSA in daily practice in this country. These incidence data were available until 1995. We used data for localization 185 from the International Classification of Diseases for Oncology (ICD-O, 1st edition); because more than 95% of prostate carcinoma cases are adenocarcinomas, we included all the histologies.
Mortality data (main cause of death, prostate cancer) are from the Institut National de la Santé et de la Recherche Médicale (INSERM) for the same areas from 1982 to 1996. We used population data from the Institut National de Statistiques et d'Evaluation (INSEE); since populations are given for 1 January each year, we calculated the number of PY for year γ as the mean of the populations for years γ and γ+1.18
We used direct standardization to study evolution of incidence and mortality eliminating age effects, using the European population as reference.18 These age-standardized rates were calculated with the classical 5-year segmentation of age, from zero to 85 years old and over.
We completed the analysis by Poisson regressions with the objective of determining whether period and/or cohort effects have influenced these evolutions. The period effect is defined by a change in the slope of linear trends in calendar-period effects (to develop cancer or to die); it is generally in favor of the influence of a factor with short evolution time, which simultaneously affects every age-specific rate. It is classically the case for the evolution of therapeutic or diagnostic methods. The cohort effect is defined by a change in the slope of linear trends in birth-cohort effects and is classically imputable to the evolution of exposure to an event with a long latency time, such as exposure to a risk or prognostic factor. To determine whether a cohort or period effect (or both) is predominant in the evolution of incidence or mortality, we used the method described by Clayton and Schifflers.19, 20 Briefly, the principle of this method is to perform successively age, age-period, age-cohort and age-period-cohort models using Poisson regression (maximum likelihood method), and to choose the model with regard to its deviance and to the contribution of the different variables (log likelihood ratio test). If Rijk is the rate (incidence or mortality) for the ith of age interval in the jth of calendar year periods, where k indexes the corresponding birth-cohort, the usual age-period-cohort model is
where αi are the age effects, βj are the calendar-period effects and χk are the birth-cohort effects. Considering the number of cases and to have sufficient periods to study these evolutions, we used age classes (45–89 years old) and periods (1982–1996) of 3 years. For the incidence, since data for 1996 was unavailable, the 1994–1996 period was represented by 2 years, 1994 and 1995. In the case of an age-period-cohort model well-fitted to data, because the 3 parameters are not separately identifiable, Maximum Likelihood estimation requires the specification of a parameter constraint: our estimates were obtained under the constraint that the first and the last cohort effects are zero. Other constraints might have lead to other results for the magnitude or the direction of the linear trends in birth-cohort effects or in calendar-period effects, but the changes in the slope are independent of the chosen constraint and can be studied by using identifiable differences in linear contrasts;21 for the change in the slope of birth-cohort effects at the 3 years birth-cohort interval centered by k, we used the contrast defined by
We used the Poisson regression to evaluate annual percentage change of the rates on log-linear scale, adjusted on age, fitting age-period models in which the calendar-period is entered as a continuous variable (= age-drift model). All confidence intervals are at 95% and the level of significance for the tests is the classical level at 5%.
During the period of 1982–1995, 14,699 cases of prostate cancer were registered by the 5 registries considered. The mean age at diagnosis for the period was 73.0 years [72.9–73.1]. This age decreased between 1982 and 1995, particularly since 1990; for the 5 areas, the means are 73.5, 73.4 and 72.5 for the periods 1982–1986, 1987–1990 and 1991–1995, respectively.
The evolution of incidence rates for the 5 areas can be divided into 3 periods (Fig. 1): between 1982 and 1987, there was an increase of 32.2% of standardized incidence rates (from 42.6 to 56.3 per 100,000 PY). The second period, 1987–1993, saw this increase intensify to 65.9% (reaching a standardized rate of 93.4 per 100,000 PY). From 1993–1995, the trend seems to reverse, with standardized rates for 1994 and 1995 at 87.0 and 87.3 per 100,000 PY, respectively. The annual percentage change estimated by Poisson regression was +5.8, +7.3 and −3.1 for the 1982–1987, 1987–1993 and 1993–1995 periods, respectively (Table I). For 1987–1993, the model is not well-fitted to data [deviance = 86.9 with 54 degrees of freedom (df), p=.01] because the acceleration of the increase includes 2 phases: the first between 1987 and 1989, with an annual percentage change of +18.5% and a second until 1993 with a slower increase of +4.5% per year. If we compare the trends between the different periods, introducing in the age-drift model an interaction term (between the drift term and a variable determining the different periods), the observed trend for the 1987–1993 period is statistically different from the trend of the 1993–1995 period (p<10−3) but not from the period of 1982–1987 (p=0.16); if we compare the trends of the 1982–1987 and 1987–1989 periods, the difference is statistically significant (p<10−3).
|Period||RR||95% CI||p Test for trend||p Deviance|
|1982–1987||1.057||1.039–1.077||p < 10−3||p = 0.10|
|1987–1993||1.073||1.062–1.085||p < 10−3||p = 0.01|
|1993–1995||0.969||0.935–1.005||p = 0.09||p = 0.02|
The incidence trends by area are fairly similar, except for the Doubs where the 1982 rate was the lowest. It has been affected by the more important increase between 1982 and 1993, reaching a level similar to the other areas in the last year.
The decrease in incidence of prostate cancer since 1993 seems to concern only age groups of 70 years old and over (Fig. 2).
The evolution of the specific mortality from prostate cancer is constituted by 2 phases (Fig. 3). The first concerns 1982–1990 with an annual increase estimated at 1.8% (p for trend =0.003). This period is followed by 1990–1996 where there is an inversion of the trend with a decrease of 3.3% per year (p for trend <10−3). As for incidence, this evolution is relatively homogeneous between the 5 areas. The analysis by age groups shows that the decrease began more precociously for age groups who died younger (Fig. 4). Death rates began to decrease for those aged 85 and over only in 1994, and this is relatively homogeneous for all areas (data not shown).
The results of the modelisations by Poisson regression of age-period-cohort effects are summarized in Table II. For the incidence, only the age-period-cohort model is well-fitted to data. Therefore, both period and cohort effects seem to have had an influence. The period effect concern certainly the inversion of the trend in 1993 (see Fig. 6); it cannot be studied by evaluating a contrast in calendar-period parameters because of a too small number of calendar-periods. The cohort effect seems to concern an acceleration of the increase in the risk of contracting the disease from 1 birth-cohort to the next, for the generations born since around 1920–1930 (Fig. 5). The birth-cohort effects curve (Fig. 6) shows a slope change around the 3 years birth-cohort interval centered on the 1925 year, and the contrast C defined in Eq. (1) takes the significant positive value of 4.38 [2.87;5.87], confirming the concave shape of this birth-cohort curve.
|Models||Incidence deviance (df)/p||Mortality deviance(df)/p|
|Age||961.9(60)/p < 10−3||80.2(60)/p = 0.04|
|Age-drift||566.6(59)/p < 10−3||69.8(59)/p = 0.16|
|Age-period||144.5(56)/p < 10−3||48.1(56)/p = 0.77|
|Age-cohort||523.1(42)/p < 10−3||60.0(42)/p = 0.16|
|Age-period-cohort||47.8(39)/p = 0.16||33.2(39)/p = 0.73|
For the mortality, age seems to explain a large part of the evolution (deviance=80.2 for 60 df). So the addition of information leads to a model well-fitted to data. If we compare age-period and age-cohort models to the age-drift model by the log-likelihood ratio test, the addition of the cohort effect does not significantly improve the fit of the model when the addition of the period effect is very significant. So, trends in mortality can be described by an age-period model with drift effect, the period effect traducing the inversion of the trend around 1990 (Table III).
As already described in several countries, we observed a significant increase in incidence of prostate cancer in the 1980s for the 5 French areas included in the study. The main reason for this increase adjusted on age is probably the improvement of the detection for this cancer with the apparition of new diagnostic techniques (ultrasonography, transuretral resection and PSA). Of course, a part of this increase is certainly artificial, reflecting an improvement of the coverage of the registries over the period. However, to minimize this “exhaustivity effect,” the registries included in the study have been chosen given their first year of activity, which might be anterior to 1980, excepted for the Tarn, which began his registration in 1982; Figure 1 shows that the incidence rates in this area are roughly similar to the others. The little acceleration in the incidence trend that began in the mid-1980s, already described in several countries,4, 5, 12, 13 is probably the result of the screening using PSA.
The stabilization of incidence, with a possible initiation of a decrease (less clear than in the SEER Program) can be attributed to several factors. PSA is very likely a major cause for this decrease in at least 2 ways. The first is the decrease in the proportion of screened persons: Legler and collaborators17 observed on a SEER program population Medicare beneficiary that the evolution of the incidence of prostate cancer is linked to the rate of first-time PSA procedures; this proportion decreased from 1992, which closely follows trends of prostate cancer incidence. This aspect is difficult to study in France because PSA does not have its own code in the health insurance nomenclature. Nevertheless, the decrease in indications of PSA in screening in recent years is probable, insofar as this screening has been highly criticized. The second possible explanation is the depletion of the pre-clinical pool of prostate cancer.16 These 2 explanations certainly played a role in the evolution of incidence in our 5 French areas, but their respective importance is difficult to determine.
After a slow and regular increase during the 1980s, prostate cancer mortality began to decrease in 1990 in the 5 areas of our study. This trend appeared sooner than in the USA (1991–1992),12 Canada (1991)14 or the United Kingdom (1992).4 As for incidence evolution, causes of lower mortality are very controversial. Indeed, the role of early diagnosis is discussed by several authors, notably because of the small delay between the beginning of the increase of incidence due to detection by PSA and the beginning of the decrease in mortality.12, 14, 16 The natural history of this disease is long, and the lead-time due to screening by PSA is evaluated at a range of 3–9 years.22 Etzioni and collaborators22 showed recently by a modelisation performed on SEER Program data that a decrease in mortality imputable to PSA screening is possible if we make the hypothesis of a short lead-time (3 years). Now, the delay separating the beginning of the incidence increase and the beginning of the mortality decrease in this study is nearly the same as (or even lower than) the one we observed from our data. A possible sign of a benefit of PSA testing is the reduction of prostate cancer mortality rates to levels below those observed before the beginning of the use of this test.16, 23 From 1985–1990, the age-adjusted mortality rate increased from 30.1 per 100,0000 men to 35.4 per 100,000 men, before decreasing to 28.5 per 100,000 men in 1996. Another argument in favor of a positive effect on mortality of the early detection of prostate cancer is the decrease in the incidence of metastatic stages concomitant with the decrease in mortality observed in this SEER Program data.12 This decrease is superior to that observed for other stages and has not been preceded by an increase after the introduction of PSA, as for local or loco-regional extension. Unfortunately, clinical data for our 5 areas are not available to study these evolutions according to clinical stages. Menegoz and collaborators3 reported stability on incidence of metastatic cases at diagnosis for the area of Isère until 1990, when loco-regional stages increased. This does not support screening because, in this area, the decrease in mortality began in 1988, but the lack of follow-up in time and sample fluctuations make it difficult to draw conclusions.
On the other hand, the comparison between prostate cancer mortality evolutions in the USA and England (including Wales)4 shows few differences between the 2 countries although PSA screening is less common in the UK than in the USA. The decrease of the specific mortality by prostate cancer appeared in an incidence increase period; in the absence of evolution in methods of death codification, the second hypothesis to explain the reduction in deaths from prostate cancer is the improvement of therapeutic practices. Indeed, in the mid-1980s, there was an extensive development of radical prostatectomy and an improvement in radiotherapy techniques; in the same way, LH-RH agonist, treatment of advanced stages, appeared around 1990. Given that radical prostatectomy and radiotherapy are early stage treatments, and that LH-RH agonist represents an alternative (most certainly interesting) to hormono-suppression, it is difficult to attribute all the decrease in prostate mortality entirely to therapeutic advances, when confronted by the short delays between these treatment developments and the mortality evolution.
Age-period-cohort models showed a cohort effect for incidence evolution. This effect is classically a result of the influence of the evolution to exposure to one or several risk factors. Several authors found a cohort effect for mortality data in Europe, with an increase in the risk of death from one birth cohort to the next,24, 25 but this effect has not yet been described for incidence data. In Denmark, the increase in incidence without important evolution in diagnosis proceedings (particularly with little development of early diagnosis) leads Brasso and collaborators26 to conclude a “true increase” in incidence. Risk factors for prostate cancer remain poorly understood; immigration data indicate genetic predisposition, but the influence of the style of life could also play a part.27 An “age-calendar-period interaction,” i.e., differences among age groups in the magnitude or direction of their calendar-period trends, can lead to a significant birth-cohort effect in age-period-cohort analyses.28 Such interaction is possible for prostate cancer incidence, in regard to variation by age of PSA screening rates, and particularly after the different recommendations on PSA screening indications published since the beginning of the 1990s; nevertheless, the curves in Figure 2 are not in favor of this hypothesis.
In conclusion, we observed a decrease of specific mortality for prostate cancer in France. This decrease appeared during an incidence increase period, without evolution of the cause of death codification, and resulting of a period effect; these facts are in favor of the improved management of prostate cancer, but the respective weights of early detection development and treatment evolutions in this improvement have still to be determined. Mass screening has not proved effective in terms of public heath, but a few randomized studies are in progress29, 30 and we have to wait for their results before drawing conclusions. With regard to individual screening, the problem is to be able to identify persons at high risk, in order to reduce a new state named “PSAdynia,” disabling anxiety resulting from the knowledge of the patient's serum PSA.31 A few risk factors are known (family history of prostate cancer32 and ethnic group33), but many questions remain unanswered, notably concerning the aspects of the way of life (diet, tobacco etc.).34, 35 The cohort effect found on incidence in our 5 French areas indicates the necessity of further research in this domain of prostate cancer etiology.
- 18Statistical methods in cancer research, IV: descriptive epidemiology. IARC Scientific Publications 128. Lyon: International Agency for Research on Cancer, 1994;, , .
- 31Eradication of a disease: how we cured asymptomatic prostate cancer. Lancet Oncol May, 17–9, 2000..