Lung cancer mortality in France

Lung Cancer (2008) 59, 282—290

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/lungcan

REVIEW

Lung cancer mortality in France Trend analysis and projection between 1975 and 2012, using a Bayesian age-period-cohort model Daniel Eilstein ∗, Zo´ e Uhry, Tek-Ang Lim, Juliette Bloch Institut de veille sanitaire, 12, rue du Val d’Osne, 94415 Saint-Maurice Cedex, France Received 2 August 2007; received in revised form 8 October 2007; accepted 11 October 2007

KEYWORDS Age-period-cohort model; Bayesian analysis; Gibbs sampler; Lung cancer; Mortality projection; Markov Chain Monte Carlo methods

Summary Introduction: Lung cancer is currently the most common cancer in the world and as such is an important public health concern. One of the main challenges is to foresee the evolution of trends in lung cancer mortality rates in order to anticipate the future burden of this disease as well as to plan the supply of adequate health care. The aim of this study is to propose a quantification of future lung cancer mortality rates by gender in France until the year 2012. Methods: Lung cancer mortality data in France (1978—2002) were extracted from the National Statistics of Death and analyzed by 5-year age-groups and periods, using a Bayesian age-periodcohort model. Discussion: Between 1978 and 2002, female lung cancer mortality rate rises by 3.3% year−1 . For men, a slow increase is observed until 1988—1992 followed by a declining trend. In 1998—2002, age-standardized mortality rates were, respectively, 45.5 and 7.6 per 100 000 for males and for females. By 2008—2012 these figures would reach 40.8 (95% credibility interval (CI): 32.7, 50.0) and 12.1 (CI: 11.7, 12.6) per 100 000, respectively, which represents among women a 4.7% annual increase (CI: 4.5, 5.0). Results: Our results highlight the relevance of pursuing public health measures in order to cope more actively with tobacco smoking in the prevention strategy against lung cancer specifically among women. © 2007 Elsevier Ireland Ltd. All rights reserved.

Contents 1. 2.

Introduction............................................................................................................ Materials and methods .................................................................................................

∗

Corresponding author. Tel.: +33 1 55 12 53 87; fax: +33 1 41 79 67 68. E-mail address: [email protected] (D. Eilstein). 0169-5002/$ — see front matter © 2007 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.lungcan.2007.10.012

283 283

Lung cancer mortality in France 3.

4.

283

Results ................................................................................................................. 3.1. Males............................................................................................................ 3.2. Females ......................................................................................................... Discussion .............................................................................................................. Conflict of interest ..................................................................................................... Acknowledgement...................................................................................................... Appendix A. Autoregressive constraints on parameters ............................................................... References .............................................................................................................

1. Introduction Lung cancer was a relatively rare disease at the beginning of the century, but is now the most common cancer in the world. Different factors are suspected to be related to this disease, such as exposition to asbestos or to radon, or as food or genetic factors [1]. None of them, however, are as important as tobacco in the incidence and mortality of lung cancer. A considerable increase in tobacco consumption in industrialized countries over the past century has been observed. Trends in lung mortality are largely the results of this evolution, with an approximate 20 years lag due to the long latent period for this cancer. In France for the year 2000, lung cancer was the 2nd most frequent cancer among males and the 4th among females [2], representing 24% of deaths due to cancer in males and 8% in females [3]. Male lung cancer mortality have slightly increased, on average, over the past 20 years (+0.7% year−1 between 1980 and 2000), while female lung cancer mortality have increased significantly over the same period (+2.9% year−1 ) [2,3]. This evolution affects more particularly new generations [2,3] and is most likely largely related to the increase in tobacco consumption Considering the observed trends in lung cancer mortality, it is crucial for public health authorities and stakeholders to have a quantification of lung cancer mortality in the years to come. Indeed, this information is essential for anticipating medical needs, hence structuring the supply of medical care in order to cope efficiently with the burden of lung cancer. And in the future, a difference between projected and observed mortality would incline to search for explanations of the unexpected trend. This study proposes to project, by gender, lung cancer mortality rates and number of deaths in France for 10 years (2003—2012) based on deaths observed between 1978 and 2002.

2. Materials and methods Mortality data were obtained from the French National Statistics of Death (C´ epiDc/French National Institute of Health and Medical Research) from 1978 to 2002. We analyzed deaths from lung cancer (ICD8-ICD9: 162) among males and females aged over 20 years and living in the metropolitan France at the date of death. Population data, estimated from 1978 to 2002 and projected from 2003 to 2012, were obtained from the National Institute of Statistics and Economic Studies (Insee). Person-years were estimated from these population data [4]. Data were aggregated by sex, 5-year age ranges (20—24 to 95+ years) and 5-year periods (1978—1982 to 2008—2012). Mortality data were

286 286 287 288 289 289 289 290

observed from 1978 to 2002 and projected from 2003 to 2012. Projections were based on an age-period-cohort model (APC) [5,6]. This model is a Poisson regression [7], where the number of deaths is analyzed according to the age at death, the year of death and the year of birth (cohort). These models do not require knowledge of etiologic factors. The age effect accounts for the duration of the exposition to risk factors. Period effect represents factors that impact all individuals at the same time, regardless of their age (e.g. short-term exposition, behavior modification, therapeutic improvement). Cohort effect corresponds to an exposition which is specific to each generation. The model is: Yijk ∼ Poisson (mijk ijk ) and ln(ijk ) = ai + pj + ck Yijk is the number of deaths, mijk is the number of personyears, ijk is the mortality rate for age i, period j and cohort k, ai , pj and ck are, respectively, the effect of age i, period j and cohort k. Projection relies both on the extrapolation of the estimated age and cohort effects, and on the projection of period effects and of cohort effects for the new unobserved cohorts. Parameters were estimated using a Bayesian method with autoregressive constraints on parameters [8,9]. These constraints enable to smooth effects and to project period and cohort effects (see Appendix for details). It should be noted that only the two last cohorts (1983 and 1988) do not include observed mortality data. The analyses were performed by using BUGS software [10,11], in order to run Bayesian analyses, using a particular Markov Chain Monte Carlo method (Gibbs sampler) [12,13]. Results were obtained with 5000 simulations preceded by 4000 burn-out simulations. This method gets increasing interest in the literature [14—21]. Age-cohort and ageperiod-cohort models were compared considering the fit of model, the DIC deviance information criterion [22] and the sensitivity of the projections to the choice of the model. The mortality rates were age-standardized on the world population. All estimations are presented with their 95% credibility interval (CI) — Bayesian statistics consider credibility interval instead of confidence interval. The credibility interval is derived from the a posteriori distribution of the parameter given the data. A previous work has been carried on female lung cancer [20] where age-standardized mortality rates have been truncated (≥20 years), hence in order to compare previous results with the ones displayed below, the rates have to be multiplied by 0.6.

284

Table 1 model

Male age-specific and age-standardized lung cancer mortality rates per 100 000 in France, estimated (1978—2002) and projected (2003—2012) with an age-period-cohort

Age

1978—1982

1983—1987

1988—1992

1993—1997

1998—2002

2003—2007 CI

2008—2012 CI

20—24 25—29 30—34 35—39 40—44 45—49 50—54 55—59 60—64 65—69 70—74 75—79 80—84 85—89 90—94 ≥95 ASR

0.2 0.5 1.8 6.6 19.4 46.6 88.4 136.2 198.4 267.0 321.2 350.6 341.4 302.8 244.3 178.1 43.4* (43.1—43.7)†

0.1 0.5 2.0 7.3 21.2 47.0 90.4 150.7 209.0 271.9 341.6 382.1 379.8 342.0 281.5 221.0 46.0* (45.7—46.3)†

0.1 0.4 2.1 8.2 23.3 50.8 90.2 152.7 228.9 283.8 344.6 402.6 410.2 377.0 315.0 251.9 48.2* (47.9—48.5)†

0.1 0.3 1.5 7.8 24.7 52.9 92.7 144.7 220.3 295.3 341.8 385.8 410.5 386.8 329.7 267.2 47.9* (47.7—48.3)†

0.1 0.3 1.2 5.6 22.5 53.7 92.2 142.0 199.6 271.7 339.9 365.7 376.1 369.9 323.2 267.5 45.5* (45.3—45.8)†

0.1 0.2 0.9 4.2 16.2 48.9 93.9 141.7 196.4 246.4 313.3 364.4 357.1 339.6 309.8 263.2 43.2*

0.1 0.2 0.8 3.4 12.3 35.4 85.8 144.7 196.5 243.4 285.2 337.1 357.1 323.5 285.3 252.5 40.8*

* †

0.1—0.1 0.1—0.3 0.7—1.2 3.6—4.9 14.4—18.2 44.2—54.2 85.2—103.6 128.4—156.0 178.0—216.3 223.7—271.9 284.5—345.4 330.8—401.0 323.6—393.6 307.0—374.1 279.2—343.2 230.5—297.6 39.2—47.5

0.0—0.1 0.1—0.3 0.5—1.2 2.5—4.6 9.6—15.5 28.1—43.8 68.6—105.0 115.9—177.1 157.7—241.1 195.6—299.1 228.8—349.1 270.2—413.1 286.3—438.0 259.6—397.4 228.4—351.9 198.9—314.9 32.7—50.0

Mortality rates are age-standardized on the world population. (Credibility intervals).

D. Eilstein et al.

Age

1978—1982

1983—1987

1988—1992

1993—1997

1998—2002

2003—2007 CI

2008—2012 CI

20—24 25—29 30—34 35—39 40—44 45—49 50—54 55—59 60—64 65—69 70—74 75—79 80—84 85—89 90—94 ≥95

0.1 0.1 0.3 0.8 2.0 4.2 8.0 11.3 16.0 22.1 27.6 34.7 41.6 47.5 46.4 40.4

0.1 0.2 0.5 1.1 2.7 4.7 8.6 13.6 18.7 24.6 31.3 38.2 44.4 49.5 51.5 46.0

0.1 0.2 0.7 1.8 3.6 6.3 9.7 14.6 22.4 28.6 34.8 43.4 49.0 52.8 53.7 50.9

0.0 0.2 0.7 2.5 5.8 8.4 12.9 16.4 24.0 34.4 40.6 48.2 55.5 58.3 57.3 53.1

0.0 0.2 0.7 2.7 8.2 13.4 17.3 21.9 27.1 36.8 48.8 56.3 61.8 66.1 63.2 56.7

0.0 0.1 0.6 2.6 8.9 19.1 27.6 29.4 36.1 41.5 52.2 67.6 72.1 73.5 71.6 62.6

0.0—0.1 0.1—0.2 0.5—0.8 2.2—3.1 8.0—9.8 17.8—20.5 26.0—29.2 27.9—31.0 34.4—37.8 39.9—43.3 50.4—54.0 65.5—69.8 69.8—74.5 70.8—76.3 68.4—75.2 57.1—68.5

0.0 0.1 0.6 2.3 8.5 20.6 39.3 47.0 48.4 55.4 58.9 72.3 86.6 85.8 79.7 71.0

0.0—0.1 0.1—0.3 0.3—0.9 1.7—3.2 7.0—10.1 18.4—22.9 36.3—42.5 44.1—50.0 45.9—51.1 52.8—58.1 56.5—61.5 69.7—75.1 83.5—89.7 82.6—89.0 75.8—83.8 64.8—77.7

9.4—9.8

12.1*

11.7—12.6

ASR * †

4.0* (3.9—4.0)†

4.5* (4.4—4.5)†

5.2* (5.2—5. 3)†

6.2* (6.2—6.3)†

7.6* (7.5—7.7)†

9.6*

Lung cancer mortality in France

Table 2 Female age-specific and age-standardized lung cancer mortality rates per 100 000 women in France, estimated (1978—2002) and projected (2003—2012) with an age-cohort model

Mortality rates are age-standardized on the world population. (Credibility intervals).

285

286

D. Eilstein et al.

Fig. 1 Male age-standardizeda lung cancer mortality rate per 100 000 in France from 1978—1982 to 2008—2012 estimated with an age-period-cohort model. a Age-standardized on the world population.

3. Results

3.1. Males

Full APC model was used for men since the AC model presented a strong lack of fit and was unable to catch the observed reverse in the trend in the lung mortality rate. Furthermore, the Bayesian deviance information criterion (DIC) was strongly in favor of the APC model (882 vs. 1334). For women, the more parsimonious AC model was chosen since period effect added little to the fit of the model and did not change the projections (though the DIC was moderately in favor of APC model: 731 vs. 738).

Estimated (1978—2002) and projected (2003—2012) agespecific and age-standardized lung cancer mortality rates in men computed from the APC model are displayed by 5-year periods in Table 1. The age-standardized mortality rate is plotted in Fig. 1. From 1978 to 2002, the mortality rate trend is two-fold, first, from periods 1978—1982 to 1988—1992, an 11.0% increase (CI: 9.9, 12.0), i.e. +1.0% year−1 . Second, from periods 1988—1992 to 1998—2002, a −5.5% decrease (CI: −6.2,

Fig. 2 Male age-specific lung cancer mortality rate per 100 000 in France by cohort, estimated with an age-period-cohort model over the period 1978—2012.

Lung cancer mortality in France

287

Fig. 3 Female age-standardizeda lung cancer mortality rate per 100 000 in France from 1978—1982 to 2008—1012 estimated with an age-cohort model. a Age-standardized on the world population.

−4.6), i.e. −0.6% year−1 . This declining trend in lung cancer mortality rate is expected to continue between the periods 2003—2007 and 2008—2012, by −1.1% year−1 on average (CI: −3.2, 0.9). Which accounts for a −10.4% (CI: −28.1, 9.7) decrease between 1998—2002 period and 2008—2012 period. For the observation period (1978—2002), an increase by 36% in the annual number of deaths is recorded (15 224 in 1978—1982 to 20 771 in 1998—2002). This trend, due to the effect of ageing, is expected to continue and the annual number of deaths from lung cancer for men will most likely reach 22 200 in 2008—2012. Fig. 2 shows age-specific rates for the different birth cohorts (1879—1987 to 1984—1992); each age group is rep-

resented by a curve. For a given age, when analyzing the oldest to the youngest cohorts (i.e. following a curve from the left to the right), a continuous increase in the mortality rate is recorded, reaching an acme for people born between 1949—1962 cohort, followed by a decrease in the youngest cohorts.

3.2. Females Estimated (1978—2002) and projected (2003—2012) agespecific and age-standardized lung cancer mortality rates in women obtained from the AC model are presented by 5-year periods in Table 2.

Fig. 4 Female age-specific lung cancer mortality rate per 100 000 in France by cohort, estimated with an age-cohort model over the period 1978—2012.

288 The age-standardized mortality rate is plotted in Fig. 3. For the whole period, this rate rises strikingly with time. Female lung cancer mortality rate increased by 92.9% (CI: 88.6, 97.1) between periods 1978—1982 and 1998—2002, i.e. 3.3% year−1 . This increase is expected to carry on between periods 2003—2007 and 2008—2012, by +4.7% year−1 on average (CI: 4.5, 5.0). This represents a 59.1% (CI: 55.1, 63.2) increase between periods 1998—2002 and 2008—2012. Mortality rate for men was 11 times higher than for women in 1978—1982, 6 times higher in 1998—2002, and is predicted to be only 3.3 times higher in 2008—2012. The annual number of deaths from lung cancer among women increased by 122% between 1978—1982 and 1998—2002 (2016—4471), and should present an additional 73% increase by 2008—2012 reaching 7700. Considering the evolution of mortality rates according to birth cohort (Fig. 4), one observes that, for a given age, the rate increases markedly with birth cohorts and reaches a peak for women born around 1963. The rate slightly decreases afterwards for the youngest cohorts.

4. Discussion This study analyzed and projected for 10 years (2003—2012) lung cancer mortality trends in France, based on the observed mortality from 1978 to 2002, using a Bayesian approach of the age-period-cohort model. The model predicted an ongoing dramatic increase in the age-standardized lung cancer mortality rate for women and a slight decrease for men. Age-standardized rate should vary from 43.2 in 2003—2007 to 40.8 per 100 000 in 2008—2012 for men and from 9.6 to 12.1 per 100 000 for women, which represents, respectively, an 11% decrease and a 60% increase since 1998—2002. Similar evolutions have been observed in most industrialized countries [23—26]. Lung cancer mortality has known dramatic increases among women and for the most affected countries (UK, USA, Canada), the rates stabilize now at very high levels with age-standardized mortality rates ranging from 20 to 25 per 100 000. The beginning of a slight decline has been observed in the UK [26]. Among men, most developed countries have experienced a more or less important decrease in lung cancer mortality [23,24]. In 2000, lung cancer age-standardized mortality rate varied worldwide from 3.0 to 25.0 per 100 000 among female, and from 16.0 to 80.0 per 100 000 among male [23]. The projections in our work are based on the extrapolation in the future of a set of information already contained in the observed data (levels, trends, evolutionary structure). The projections obtained extend and amplify what have been observed previously. A reproach usually addressed to this type of prevision is to be unable to take into account modifications in the etiologic factors. This must be put into perspective when the involved factors present slow evolutions. It is true, however, that a sudden modification of past trends might cause unforeseen variations in lung cancer incidence and mortality. However, rationale for projections is to propose a baseline to evaluate different scenarios either when a change in behaviors occurred or because of specific public health policies adopted.

D. Eilstein et al. Despite the growing need for estimations of future cancer incidence and mortality rate, there is still few works in this field. Brennan and Bray have analyzed trends in lung cancer mortality in Europe, and have computed short-term projections with a Bayesian APC model [24]. Bayesian APC models have been compared to alternative methods on both incidence and mortality assessments/projections, and for different cancer sites [16,17]. Bashir and Esteve have computed a set of projections for various cancers in Finland [17]. The comparison between models has showed that Bayesian APC and classic AC models are equivalent and outperform models based on a simple extrapolation [27—30]. Bray has used forty different cancer incidence and mortality datasets to explore the behavior of Bayesian APC model and has compared it with alternative methods and has concluded that Bayesian models are a flexible and robust method to project cancer incidence and mortality [16]. Recently, Clements et al. proposed to use generalized additive models (GAM) to predict female lung cancer mortality in five countries and compared their model with Bayesian APC models [31,32]. They found that GAM model had better predictive performance than Bayesian APC models. The reason for this, according to the authors, is that APC Bayesian model relies on a linear extrapolation of period and cohorts effects outside the observed data while GAM model relies on smoothing splines providing a continuing curvature outside the observed data. They also outlined that Bayesian APC model produced wide credible intervals for projections as opposed to GAM model. However, the narrow confidence intervals obtained with the GAM model, or with the Bayesian AC model, might not account for the real uncertainty of the projections. Nevertheless the method proposed by Clements is a highly welcome contribution to projection models and must be tested on further datasets. It would be interesting to incorporate explicative factors in the model such as tobacco consumption, exposure to asbestos, to radon or any other lung cancer carcinogens such as arsenic compounds or silica. This would enable to assess different scenarios of lung cancer mortality longterm impacts based on changes in exposures. In France, detailed information about tobacco consumption is unfortunately scarce. A considerable work has been done recently to gather all available data on tobacco consumption [33]. However the models we use require even more detailed data. Despite the lack of more detailed data, APC models enable to have an acute quantification of lung cancer trends. Several works have attempted to compute lung cancer incidence or mortality projections incorporating information about tobacco consumption, using a variety of approaches. Brown and Kessler [34] have modeled and projected lung cancer mortality with an APC model including information about tobacco consumption. The cohort effect represented the tobacco consumption prevalence, while the period effect was replaced by a variable representing cigarettes tar concentration, assuming a linear effect for this variable. Holford et al. [35] have studied a complex model of lung cancer incidence based on the effects of age, period, cohort and on detailed information about tobacco consumption (prevalence of never smokers, former smokers and current smokers; mean duration of smoking among current and former smokers, etc.).

Lung cancer mortality in France Other kind of works predicted present and past lung cancer mortality from tobacco consumption histories, based on individual dose-response models for tobacco smoking, i.e. in the UK [36] and in the US white male population [37]. These methods enable to assess the impact of tobacco smoking on lung cancer mortality and also, by comparing predicted to observed lung cancer mortality, to appraise the influence of factors other than smoking. Those points are beyond the scope of our present work and are on our agenda for future research. Male lung cancer mortality slightly decreased since 1988—1992. The risk of dying from lung cancer decreased substantially for men born after 1960, as can be seen in Fig. 2. Undoubtedly, a period effect has also contributed to move downward the lung mortality rates in the recent past years. The cohort pattern can be related to the evolution of the prevalence of tobacco consumption [33]. The global cigarette consumption among men substantially decreased after 1990, which is encouraging and support the decline predicted in male lung cancer mortality [38]. Other aspects of tobacco consumption have also changed over time (switch from brown to blond tobacco and from handmade to manufactured cigarettes, decrease in tar) which likely tended to decrease lung cancer mortality as well. In the United States, Jemal et al. analyzed trends in lung cancer mortality from 1970 to 1997 using an age-period-cohort model [39]. They reported in particular a calendar-period decrease in lung cancer mortality after 1990 for both genders, which could be related to carcinogens in cigarettes reduction and to the increase in smoking cessation since the sixties. On the other hand, female lung cancer mortality presented a dramatic increase, which can be related to the development of tobacco consumption. The global cigarette consumption among women reached an acme in the nineties that is 20 years after the maximum observed in the UK or the US, and has stabilized since [38]. Most likely, the effect of this tobacco consumption increase will continue to affect female lung mortality within the next 10 years. The evolution of lung cancer mortality rates observed among youngest cohorts (born after 1960) are difficult to interpret and should be considered with caution. The observed rates present a high random variability due to small number of deaths and the effect of tobacco is not fully visible at these young ages due to the latency period of tobacco smoking. For the moment, it is too early to understand the dynamic evolution of lung cancer mortality for these young cohorts. However, they have little influence on the projections of the age-standardized mortality rate or of the total annual number of deaths. It should be noted that a slight decrease in mortality rates in young women has been observed in several European countries [26]. A similar phenomenon might be beginning to emerge in France as well. The model we used predicted, for the next 10 years in France, a dramatic increase in the age-standardized lung cancer mortality rate in women, and a slight decrease in men. In countries where tobacco consumption among women has developed earlier than in France (US, UK, etc.), female lung cancer mortality is currently equal or even higher than colorectal or breast cancer mortality and has become the first cause among female cancer deaths in several countries [23]. The lung cancer epidemic among French women, though it might not reach the level experienced

289 by other countries due to lower average tobacco consumptions among women [38], will continue in the years to come. The purpose of this study is to shed lights on a major public health issue, by taking into account the results discussed, public health stakeholders and the population may have a better understanding of the issue at stake with a better quantification of lung cancer in the next 10 years. As such, more tailored public health policies could be adopted in order to tackle this burden by acting on factors facilitating lung cancer and also to adapt the organization of health care planning appropriately.

Conflict of interest The authors have no conflict of interest to declare.

Acknowledgement This paper benefited from relevant comments and remarks from the reviewer, whom we are thankful.

Appendix A. Autoregressive constraints on parameters Let us denote i, j and k the age, period and cohort numeration and ai , bj and ck their corresponding effects. The autoregressive constraints on parameters are defined as follow. The conditional distribution of age effect ai given all others age effects ai  (i = i) is: ai |ai , i = i ∼ N(ai , 1/a ) and ai = E(ai |ai , i = i) is the conditional mean of age effect ai , 1/ a is the conditional variance (and  a the conditional precision, i.e. the inverse of variance). The conditional mean ai is expressed as: ai = (4ai−1 + 4ai+1 − ai−2 − ai+2 )/6, that is a cubic interpolation from the adjacent effects ai−1 , ai+1 , ai−2 , ai+2 . The conditional distribution of period effect bj given all preceding effects bj (j < j) is: bj |bj , j < j ∼ N(bj , 1/b ) and bj = E(bj |bj , j < j), with bj = 2bj−1 − bj−2 , which is a linear extrapolation from the two preceding effects bj−1 and bj−2 . Likewise to period effects, the conditional distribution of cohort effects is: ck |ck , k < k ∼ N(ck , 1/c ) and ck = E(ck |ck , k < k) with ck = 2ck−1 − ck−2 All precisions follow Gamma a priori distributions. These constraints allow to smooth effects and to project period and cohort effects. The two types of constraints (undirected for age effects; directed for cohort and period effects) are equivalent although written in a different way. The undirected form is preferred for the age effects since a directed form would induce dependence in the parameters estimation to the youngest age effects, which can be poorly estimated due to the small number of cases.

290

References [1] Beckett WS. Epidemiology and etiology of lung cancer. Clin Chest Med 1993;14(1):1—15. ´volution [2] Remontet L, Buemi A, Velten M, Jougla E, Esteve J. E de l’incidence et de la mortalit´ e par cancer en France de 1978 ` 2000 [Report]. Francim/Hˆ a opitaux de Lyon/Inserm/Institut de veille sanitaire, September 2002. [3] Remontet L, Est` eve J, Bouvier AM, et al. Cancer incidence and mortality in France over the period 1978—2000. Rev Epidemiol Sante Publique 2003;51:3—30. [4] Hill C, Doyon F. Age in completed years versus age reached during the year. Rev Epidemiol Sante Publique 2005;53:205— 8. [5] Clayton D, Schifflers E. Models for temporal variation in cancer rates. I. Age-period and age-cohort models. Stat Med 1987;6:449—67. [6] Clayton D, Schifflers E. Models for temporal variation in cancer rates. II. Age-period-cohort models. Stat Med 1987;6: 469—81. [7] McCullagh P, Nelder JA. Generalized linear models. 2nd ed. London: Chapman and Hall; 1989. [8] Breslow NE, Clayton DG. Approximate inference in generalized linear mixed models. J Am Stat Assoc 1993;88:9—25. [9] Berzuini C, Clayton D. Bayesian analysis of survival on multiple time scales. Stat Med 1994;13:823—38. [10] Spiegelhalter D, Thomas A, Best N, Gilks W. BUGS 0.5 Bayesian inference using Gibbs sampling manual (version ii); 1996. [11] Spiegelhalter D, Thomas A, Best N, Gilks W. Ice: non-parametric smoothing in an age-cohort model. In: BUGS 0.5 examples volume 2 (version ii); 1996. p. 30—6. [12] Gilks WR, Richardson S, Spiegelhalter DJ. Markov Chain Monte Carlo in practice. New York: Chapman & Hall Interdisciplinary Statistics; 1996. p. 486. [13] Bray I. Application of Markov Chain Monte Carlo methods to projecting cancer incidence and mortality. J R Statist Soc 2002;51:151—63. [14] Eilstein D, H´ edelin G, Schaffer P. Incidence of colorectal cancer in Bas-Rhin, trend and prediction in 2009 (French). Bull Cancer 2000;87:595—9. [15] Bray I, Brennan P, Boffetta P. Projections of alcohol- and tobacco-related mortality in Central Europe. Int J Cancer 2000;87:122—8. [16] Bray I. Application of Markov chain Monte Carlo methods to projecting cancer incidence and mortality. Appl Stat 2002;51:151—64. [17] Bashir SA, Est` eve J. Projecting cancer incidence and mortality using Bayesian age-period-cohort models. J Epidemiol Biostatist 2001;6:287—96. [18] Bray I, Brennan P, Boffetta P. Recent trends and future projections of lymphoid neoplasm—–a Bayesian age-period-cohort analysis. Cancer Causes Control 2001;12:813—20. [19] Eilstein D, H´ edelin G, Schaffer P. Cervical cancer in Bas-Rhin: trend and prediction of the incidence in 2014 (French). J Gynecol Biol Reprod 2002;31:28—33.

D. Eilstein et al. [20] Eilstein D, Uhry Z, Cherie-Challine L, Isnard H. Lung cancer mortality among women in France. Trend analysis and projection between 1975 and 2014, with a Bayesian age-cohort model. Rev Epidemiol Sante Publique 2005;53:167—81. [21] Eilstein D, Quoix E, H´ edelin G. Incidence of lung cancer in BasRhin, France: trends and projection for 2014. Rev Mal Respir 2006;23:117—25. [22] Spiegelhalter D, Best N, Carlin B, Va,der Linde A. Bayesian measures of model complexity and fit. J Royal Stat Soc Ser B 2002;64:583—639. [23] Web site IARC (http://www-depdb.iarc.fr/who/menu.htm). [24] Brennan P, Bray I. Recent trends and future directions for lung cancer mortality. Eur B J Cancer 2002;87:43—8. [25] Bray F, Tyczynski JE, Parkin DM. Going up or coming down? The changing phases of the lung cancer epidemic from 1967 to 1999 in the 15 European Union countries. Eur J Cancer 2004;40:96—125. [26] Bosetti C, Levi F, Lucchini F, Negri E, La Vecchia C. Lung cancer mortality in European women: recent trends and perspectives. Ann Oncol 2005;16:1597—604. [27] Hakulinen T, Dyba T. Precision of incidence predictions based on Poisson distributed observations. Stat Med 1994;13:1513—23. [28] Dyba T, Hakulinen T, P¨ aiv¨ arinta L. A simple non-linear model in incidence prediction. Stat Med 1997;16:2297—309. [29] Dyba T, Hakulinen T. Comparison of different approaches to incidence prediction based on simple interpolation techniques. Stat Med 2000;19:1741—52. [30] Osmond C. Using age, period and cohort models to estimate future mortality rates. Int J Epidemiol 1985;14:124—9. [31] Clements MS, Armstrong BK, Moolgavkar SH. Lung cancer rate predictions using generalized additive models. Biostatistics 2005;6:576—89. [32] Clements MS, Armstrong BK, Moolgavkar SH. RE: ‘‘Bayesian projections: what are the effects of excluding data from younger age groups’’. Am J Epidemiol 2006;164:293—4 (Author reply: 293—4). [33] Hill C, Laplanche A. Le tabac en France: les vrais chiffres. La documentation franc ¸aise. Paris, Mai; 2004. [34] Brown CC, Kessler LG. Projections of lung cancer mortality in the United States: 1985—2025. J Natl Cancer Inst 1988;80:43—51. [35] Holford TR, Zhang Z, Zheng T, McKay LA. A model for the effect of cigarette smoking on lung cancer incidence in Connecticut. Stat Med 1996;15:565—80. [36] Lee PN, Forey BA. Trends in cigarette consumption cannot fully explain trends in British lung cancer rates. J Epidemiol Commun Health 1998;52:82—92. [37] Swartz JB. Use of a multistage model to predict time trends in smoking induced lung cancer. J Epidemiol Commun Health 1992;46:311—5. ´volution de la consommation de cigarettes [38] Hill C, Laplanche A. E en France par sexe, 1900—2003. Bull Epidemiol Hebdomadaire 2005;21—22:94—7. [39] Jemal A, Chu KC, Tarone RE. Recent trends in lung cancer mortality in the United States. J Natl Cancer Inst 2001;93: 277—83.