How does smoking impact mortality rate in older age : investigation based on data across Europe

(1)

How does smoking impact mortality rate in older age:

investigation based on data across Europe

Diandian Yi

0 u T1 T2 T3 T4=T T5 t→ U(t)

Master’s Thesis to obtain the degree in Actuarial Science and Mathematical Finance University of Amsterdam

Faculty of Economics and Business Amsterdam School of Economics Author: Diandian Yi Student nr: 11089199

Email: yi.diandian@hotmail.com Date: August 11, 2016

Supervisor: MSc. Andrei Lalu Second reader: PhD. Servaas van Bilsen

(2)

(3)

How does smoking impact mortality rate in older age — Diandian Yi iii

Statement of Originality

This document is written by Student [Diandian Yi] who declares to take full responsi-bility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(4)

(5)

How does smoking impact mortality rate in older age — Diandian Yi iii

Abstract

We aim to find the trend of impact from smoking on mortality rate in older age in this paper. To quantify the impact, we employed the PGW method which is developed by Preston, Glei, and Wilmoth to calculate deaths attributable to smoking in 2010. This method not only used lung cancer mortality as the basic indicator of the damage caused by smoking in a particular population, but also investigate the macro-level statistical association between lung cancer mortality and mortality from all other causes of death to get a more comprehensive quantification. We used death cause data from Eurostat with this method to look at the trends across Europe with particular focus on Western and Northern European regions. We found that smoking would raise the relative risk of smoker to non-smoker (ratio of smoker’s mortality rate to nonsmoker’s mortality rate) to a higher and higher level along with people getting older until the relative risk of smoker to non-smoker reaches a peak, and then, the relative risk diminishes. That means after smoker live until the age with peak relative risk, their mortality rate would become more and more similar to nonsmoker. If we looked at the trend in Western/Northern European countries, for male, the peak in relative risk of smoker to non-smoker emerges at 70-74 years old and is around 2.20 times (smoker’s mortality rate/nonsmoker’s mortality rate), and then the relative risk declines to around 1.5 times when people grow to 80-84 years old; for female, the peak emerges at 55-70 years old and is around 2.60 times, and then relative risk declines to around 1.3 times when people grow to 80-84 years old. If we looked at the trend in the whole Europe area, for male, the peak emerges at 60-65 years old and is around 2.8 times, and then relative risk declines to around 1.6 times when people grow to 80-84 years old; for female, the peak emerges at 55-65 years old and is around 1.7 times, and then relative risk declines to around 1.0 times when people grow to 80-84 years old. The stronger similarities we found between different genders in these patterns across Western/Northern European areas than that across the whole Europe can be explained by the much more convergence between male and female in smoking behavior in Western/Northern European countries than that in the rest of Europe. And the change of level among different investigation year suggests that women’s prevalence of smoking is increasing while men’s is decreasing while the heaviness of smoking in both genders are increasing.

Keywords Mortality, Cigarette smoking, Europe, PGW method, Relative risk of smoker to

(6)

Preface

Yestoday, after I finished my thesis writing, I read a description of thesis which I think to be very accurate: “Thesis is a collection of pages filled with words, equations, figures and tables, that hopefully conveys some useful knowledge for certain experts in some field of research. However, for the author it usually represents much more than that. It represents more like an end of a journey.”1 Yes, I still can recall the first day in last September that I came to this beautiful city and started my academic journey in UvA. On that day I still dont know that I will experience one unforgettable year in which I have great opportunity to connect and develop what I have learnt in the past about Econometrics, Finance and risk management and also learn brand new things in using R, Matlab and Machine Learning to manipulate empirical data to link theory to real business.

I learnt a lot from courses and tutorial arranged by Prof. Rob kaas, Prof. Michel H. Vellekoop, Dr. Simon Broda, Dr. Servaas van Bilsen, Dr R. Perez Ribas, Dr Zhenzhen Fan and so on. Besides, during this year, I also made friends with my classmates who are from different countries, and enjoyed an open and international atmosphere in my program. I believe these experience will be all very rewarding to me.

In the beginning of 2016, Im very fortunate to have an opportunity to become an intern in Achmea Reinsurance. Working and communicating with colleagues in Achmea make me know more about business in insurance industry. And the most important, idea of this thesis comes from Achmea. Elaine Turner, Gijs Kloek, John Bastiaansen, Joost de Jongh, Paul van der Velden and Yolande Pernot-Odekerken, thank you for providing data information and recommending great papers in this field, and for your encouragement and advice during the toughest part of this journey. Hopefully my paper can convey some useful knowledge in this field for you experts.

Many thanks go to Andrei Lalu and Servaas van Bilsen. During the process of writ-ing thesis, your professional suggestions help me a lot.

Special thanks go to my parents. It is so sweet to talk to you even only on facetime. Looking forward to meeting you at home.

And most of all, to Zichen Deng, thank you for your gentle encouragement, your love and good humour, and for celebrating the near-finalisation of this thesis with me so many times in recent months.

1_{Quote from Dr. Eelco Doornbos’ PhD thesis (2010, Delft/Leiden).}

(8)

(9)

Chapter 1

Motivation and Introduction

Factors impacting mortality rate and how they affect people’s mortality rate has be-come a focal point of research in demographics, the health care sector, pension fund and insurers of life products. At the same time, as people’s life expectancy has been increasing rapidly over the last decades, the trend of these factors impacting on people with older age also becomes an interesting problem because the large number of elderly aged people begins to draw society’s attention to their life quality and insurance com-pany may also consider the accurate pricing of life insurance product for older people. Actually there have been lots of literature about factors impacting mortality rate at every specific age, how mortality rate develop when people gets older, and so on, but the effect of interaction between factors such as smoking and aging is rarely discussed. Extensive evidences in the literature document that factors like whether people smoke or not will impact their mortality rates a lot. However, when people gets older and older, does this factor still have same impact on people at an older age in the same way as people at a young age? Our research in the current thesis is motivated by such question and concern. The smoking impacted mortality rate in older age also links to a hot topic: the process of aging and how biological and people’s behavior factors in-teract in the individual trajectory from good health to death. The process of aging is hard to separate from deteriorating health, both being susceptible to environmental and social factors. Our study tries to find out whether getting older also implies more bio-logical determination or, on the contrary, continues and accumulates people’s behavior influences.

Besides concerns from academic standpoint and public attention to older people life quality, this research also has its practical application value. In practice, insurance com-panies price their life insurance products based on people’s mortality rate. Regarding to different group of people with different risk characteristics such as non-smoking, com-panies apply the percentage methodology to adjust mortality rate from basic mortality table i.e., a mortality table for general population. To explain this in detail, insurers multiply the general mortality rate with a percentage (for example: 70%) to get the mortality rate of people who don’t smoke since they have lower risk. Because the per-centage applied for every specific age is the same, this implies an assumption that a kind of risk affect people in different age at the same level. To measure this effect more accurately and build more sophisticated pricing system to charge fair premium, clari-fying the trend of factors impacting people’s mortality rate with their age gets older is very meaningful.

Based on data accessibility, we choose smoking as our main analyzing factor. As previous researches have demonstrated a clear link between the cigarette smoking and a excess risk in mortality rate, plenty of population data sets have included information about whether people smoke or not. However, the specific methodology to calculating smoking-attributable risk is still left to more discussion. In thesis, we will consider these methods in literature reviews and choose the indirect method to indicate the smoking

(10)

2 Diandian Yi — How does smoking impact mortality rate in older age

risk.

This research adopt the data from Eurostat which provides mortality information on developments over time in the underlying causes of death and also highlights geo-graphical differences. Using the indirect method, we inspect not only the trend of the effect of smoking on mortality rate on older age people, we also interpret the parameter to conduct a relatively comprehensive comparison among different geographic area in Europe and among Europe and U.S using existing mortality rate table from U.S..

The remainder of our paper is as follows. Chapter 2 describes the literature, where we review existing literature about factors impacting mortality rate. Regarding to smok-ing, we will go through crucial research about this factor, emphasize their methods to measure smoking attributable risk, and look at their models to quantifying these factors’ impact. Latest research of mortality rate at older age also would be listed.

Chapter 3 describes our model structure. Chapter 4 introduces the data set we use and investigate the interaction between smoking and aging. In this part we apply empirical population data from Europe to look at the trend. Based on proven causal effect of smoking and high mortality rate, we build a statistical model to fit our data to explain the trend of interaction of these two effects on mortality rate.

Chapter 5 presents main results, where we investigate the trend differences between various geographic areas. Chapter 6 concludes and looks into the future in this field.

(11)

Chapter 2

Literature Review

Smoking is one of the most important factor to consider when we talk about the factors impacting mortality rate. And the way and amount smoking impacting people’s mor-tality rate has been discussed by many studies. Generally speaking, cigarette smoking increases the risk of dying from many different causes of death. According to the crite-ria used by the U.S. surgeon general for establishing a causal relationship, these causes include lung cancer, many other forms of cancer, cerebrovascular disease, chronic ob-structive pulmonary disease, and coronary heart disease (U.S. Surgeon General, 2004). In the academic world, there involves a process of development of different methods to quantify the impact of smoking on mortality rate. To identify the mortality risks associated with smoking, the Cancer Prevention Study II (CPS-II) conducted the largest study as the prospective cohort study that can be used to compare the death rates of current smokers and former smokers with the death rates of those who never smoked regularly. It has tracked mortality among a cohort numbering 1.2 million individuals when the study began in 1982. This study includes information like people’s smoking behavior, and classifies people’s smoking status according to how heavily people smoke. Based on these information, researches has been conducted using cohort studies to estimate the number of deaths in a population that are attributable to smoking. This calculation is conventionally made by comparing the actual number of deaths in a particular age-sex group in the population with the number that would have occurred if everyone had had the death rates of lifetime nonsmokers in that category. Mokdad et al. (2004) used this method to estimate that 435,000 deaths were attributable to smoking in the U.S. in 2000.

But as the number of deaths attributable to smoking was estimated directly from cohort studies, several biases exist in this cohort studies. For example, the classification of smoking status among participants may be imprecise because smoking behavior often varies over time whereas in cohort studies smoking status is typically identified at base-line and assumed constant thereafter. And the smoking categories themselves impose a rigid frame on what can be blurry patterns of behavior. What’s more, we are not sure that wether such studies are available in many populations for which attributable risk estimates are sought.

To fill these gaps, Peto, Lopez et al. (1992) developed an ingenious method. The method used observed death rates from lung cancer as an indicator of the population’s cumulative smoking exposure to replace the directly measured smoking behavior based on self-reported information. Peto et al. thus translated observed lung cancer death rates for a given population into an estimate of the smoking impact ratio by referring to the difference between lung cancer death rates for smokers and nonsmokers in CPS-II. This scalar is then used to adjust the cause-specific relative risks for smokers versus nonsmokers from CPS-II in order to derive a population-specific estimate of the risk attributable to smoking for other smoking-related causes of death. This technique heav-ily depends on the assumption that CPS-II estimates of lung cancer death rates for

(12)

smokers and nonsmokers and relative risks for other causes of death can be applied to other countries and across time. And considering smoking is somehow a self-selected be-havior, the mortality differential between smokers and nonsmokers may be confounded with the effects stemming from other risk factors. Thus, later versions of this method involve some discount in scaling.

Using this method, many researches have been done based on data from other group of population. Staetsky (2009) applied this method to look at the diverging trends in female old-age mortality. The populations this research consider are from France, Japan, Denmark, the United States, and the Netherlands. Because Mesle and Vallin (2006) and Rau et al. (2008) discovered two contrasting patterns of change in mortality of old-age females: a pattern of a large decrease in mortality exhibited by France and Japan and a pattern of a smaller decrease, stability or a certain increase in mortality shown by Denmark, the United States, and the Netherlands. Staetsy (2009) estimated the mortality rate attributable to smoking, and found that the removal of smoking related mortality makes the pace of improvement in old-age female mortality considerably more similar across countries participating in a comparison. That means different patterns of smoking behaviour in these countries partly explain the emerged contrasting patterns in old-age females mortality.

Fanny Janssen and Joop de Beer (2016) projected future mortality in the Nether-lands taking into account mortality delay and smoking. In their research, the role of smoking is outlined by estimating the share of mortality attributable to smoking, for which they use an adjusted version of the indirect Peto-Lopez-technique. They also in-vestigated the smoking-attributable mortality in Poland and Netherlands to explain the regional differences for respective countries (2012 and 2014).

Based on this version of indirect method, Preston, Glei, and Wilmoth (2010) (PGW, hereafter) developed an alternative to the Peto-Lopez method for calculating deaths attributable to smoking. Although in this method, lung cancer mortality is still used as the basic indicator of the damage caused by smoking in a particular population, they do not use the relative risks from CPS-II or any other study. Instead, they investigate the macro-level statistical association between lung cancer mortality and mortality from all other causes of death. This approach is motivated by the expectation that lung cancer mortality is a reliable indicator of the damage from smoking and that such damage has left a sufficiently vivid imprint on other causes of death (Samuel H. Preson et al., 2010). The key similarity of the two indirect methods — Peto-Lopez and PGW — is that the lung cancer death rate is interpreted as an indicator of the damage from smoking within a population. Lung cancer is a unique condition because it is so closely tied to one behavioral risk factor. According to CPS-II, smoking was responsible for more than 90% of lung cancer deaths among men and more than 70% among women (Andrew Fenelon, Samuel H. Preston, 2012). But the PGW method further investigates the impact of smoking when people died of from cause other than lung cancer by looking into the statistic relationship between lung cancer mortality rate and mortality rate from other cause other than lung cancer. Andrew Fenelon and Samuel H. Preston (2012) used this method to estimate the number of deaths attributable to smoking based on data from U.S. states. They calculated smoking-attributable fractions for the 50 states and the United States as a whole in 2004, and estimated the contribution of smoking to the high adult mortality of the southern states. They found smoking-related mortality explains as much as 60% of the mortality disadvantage of southern states compared with other regions.

For mortality trends in old age population, Fannny Janssen(2005) found that dif-ferent from the general observed decline in old-age mortality, for the Netherlands and Norway there have been reports of stagnation in the decline since the 1980s. They looked at cause-specific mortality to see for which causes of death the recent stagnation is most apparent. And they found smoking has had a marked influence on the trends in old-age

(13)

How does smoking impact mortality rate in older age — Diandian Yi 5

mortality.

This paper will apply an updated version PGW method to look into the trend of relative risk of smoker to nonsmoker (mortality rate of smoker/mortality rate of non-smoker) for people older than 65 based on country-level data across Europe.

(14)

Chapter 3

Model

We are interested in specifying relative risk of smoking to non-smoking for specific gender and age groups to see its impact’s trend. The relative risk of smoking in this paper will be calculated as the ratio of the mortality rate of smoker to the mortality rate of non-smoker.

3.1 Model Assumptions

To quantify this ratio, we construct model based on below assumptions:

We assume that smoking behavior impact people’s mortality rate in two ways: one is impacting mortality by raising mortality rate among people dying from lung cancer, measured by the ratio of the difference of death rate of smoker and the overall lung cancer death rate to the overall lung cancer death rate; the other is impacting mortality by raising mortality rate in people dying from cause other than lung cancer, and measured by the ratio of the difference of death rate of people smoking and dying from from cause other than lung cancer and death rate of people.

We also assume that the CPS-II estimates of lung cancer death rates for smokers and nonsmokers can applied to other populations. That means we will use these two estimates as benchmark when we assess relative increase or decrease of mortality rate of population of interest to lung cancer mortality rate for nonsmokers or smokers. We also used these two estimates to quantify the mortality rate of people died from other cause after we investigated the macro-level statistical association between lung cancer mortality and mortality from all other causes of death based on the data of population of interest.

3.2 Model Structure

We calculate relative risk of smoker to nonsmoker as:

RR_{smoker/nonsmoker}= RR_{smoker/population}× RR_{population/nonsmoker}, (3.1)

3.2.1 Relative risk of population of interest to nonsmoker

For RR_{population/nonsmoker}, as Figure3.1displayed, in the first impacting way, the ratio of the difference of observed lung cancer death rate of population of interest and lung cancer death rate of nonsmoker from CPS-II to observed lung cancer death rate of population of interest can be shown like:

AL=

ML− ML∗

ML

, (3.2)

where AL is the proportion of population mortality attributable to smoking risk, ML

is the observed lung cancer death rate of population of interest, and M_L∗ is the lung 6

(15)

Relative risk of smoker to population

Aggregate two impact together Impact on mortality by raising mortality rates among people dying from lung cancer

Impact on mortality by

raising mortality

rates among people dying from causes other

than lung cancer

Measured by the ratio of the difference of observed lung cancer death rate of population of interest and lung cancer death rate of nonsmoker from CPS-II to observed lung cancer death rate of population of interest.

Measured by the ratio of the difference of the death rate from other cause of population of interest and death rate from other cause for nonsmoker to observed death rate from other cause of population of interest. The death rate from other cause will be estimated using statistical relationship between death rate from lung cancer and death rate from other cause.

Assume smoking impacts mortality in two

ways

Figure 3.1: Model structure for calculating RR_{population/nonsmoker}

cancer death rate among lifelong nonsmokers. We use estimates of lung cancer death rate among lifelong nonsmokers from CPS-II for respectively gender and age group as M_L∗ in our paper. Detailed data information would be elaborated in Data chapter.

In the second impacting way, we employ a standard hazards model to investigate the statistic relationship between death rate from lung cancer and death rate from other cause. The model specification is:

ln(MO) = βLML+ βaXa+ βsXs+ βtT + βaL(ML× Xa), (3.3)

where a indexes the age group, s the country, t the investigation year. MOis the

mortal-ity rate among people dying from other causes than lung cancer, Xaand Xsare dummy

variables for age group and state respectively, we also include a linear time trend (T) as well as interactions between lung cancer mortality and age group (ML× Xa). We

estimate the statistical relationship among the population of interest between variables and MO using this equation. When we replace MLin this equation with ML∗, we expect

to get estimate of the mortality rate among nonsmokers dying from other causes than lung cancer with other variables equal.

The ratio of the difference of the death rate from other cause of population of interest and death rate from other cause for nonsmoker to observed death rate from other cause of population of interest is found by below equation.

AO=

eβL0(ML)_{− e}βL0(M ∗ L)

eβL0(ML) , (3.4)

where β_L0 is the model coefficient for lung cancer including age interactions (β_L0 = βL+ βaL). In the standard hazards model we employed in previous part, we know that

eβ0L(ML)is the death rate from cause other than lung cancer when mortality rate were set

at observed level, and eβ0L(M ∗

L)is the death rate from cause other than lung cancer when

mortality rate were set at the level for never-smokers. This AO would be the increase

part of population’s mortality from causes other than lung cancer.

The attributable fraction for total mortality is a weighted average of the attributable fractions for lung cancer and other causes:

A = ALDL+ AODO

D , (3.5)

where DL and DO are deaths from lung cancer and other causes, respectively, and

D is total deaths. Thus, A here suggests the proportion of death number among the population of interest that is attributable to smoking risk.

(16)

And following above logic, we can get the relative risk of normal population group (consists of both smoking and nonsmoking people) to that of nonsmoking people as:

RRpopulation/nonsmoker =

D

D − A × D = 1

1 − A, (3.6)

3.2.2 Relative risk of smoker to population of interest

Relative risk of smoker to population

Aggregate two impact together Impact on mortality by raising mortality rates among people dying from lung cancer

Impact on mortality by

raising mortality

rates among people dying from causes other

than lung cancer

Measured by the ratio of the lung

cancer death rate of smoker from CPI-II to observed lung cancer

death rate of population of interest.

Measured by the ratio of the death rate from other cause for smoker to the death rate from other cause of population of interest. The death rate from other cause will be estimated using statistical relationship between death rate from lung cancer and death rate from other cause.

Assume smoking impacts mortality in two

ways

Figure 3.2: Model structure for calculating RRsmoker/population

For RRsmoker/population, as Figure3.2displayed, in the first impacting way, the ratio

of lung cancer death rate of smoker from CPS-II to the observed lung cancer death rate of population of interest can be shown like:

ALS =

M_S∗ ML

, (3.7)

where M_S∗ is the lung cancer death rate among smokers and ML is the observed lung

cancer death rate of population of interest. Therefore, ALS is the ratio of the lung cancer

mortality rate the observed population would have if everyone of them be assumed to be smoker to observed lung cancer mortality rate of the population. We use estimates of lung cancer death rate among smokers from CPS-II for respectively gender and age group as M_S∗ in our paper1. Data information is discussed with more details in Data chapter.

In the second impacting way, we employ a same standard hazards model (Eq. (3.3)) as RRpopulation/nonsmoker part. We only changed the equation to calculate the ratio of

the death rate from other cause for smoker2 _{to the death rate from other cause of} population of interest: AOS = eβ0L(M ∗ S) eβL0(ML), (3.8)

where β_L0 is the model coefficient for lung cancer including age interactions (β_L0 = βL+βaL). The ratio of the mortality rate the observed population would have if everyone

of them be assumed to be smoker to observed mortality rate of the population is: AS=

ALSDL+ AOSDO

D = RRsmoking/population, (3.9)

1_{To make it clear, as M}∗

L is mortality rate among nonsmoker dying from lung cancer and M ∗ S is mortality rate among smoker dying from lung cancer, according to our assumption that smoking would increase mortality rate, we will have the relationship: MS∗ > M

∗

L. Because there are both nonsmokers and smokers among any population group, we can say there exists the relationship: MS∗ > ML> M

∗ L. And also we have: 0 <AL<1, and ALS > 1.

2

Like we have done in last section, when we replace ML in this equation with MS, we expect to∗ get estimate of the mortality rate among smokers dying from other causes from lung cancer with other variables equal.

(17)

where same as above elaboration, DL and DO are deaths from lung cancer and other

causes, respectively, and D is total deaths. And A should be larger than one.

3.2.3 Relative risk of smoker to nonsmoker

Combining the relative risk of normal population group (consists of both smoking and nonsmoking people) to that of nonsmoking people and the raised proportion by smoking compared to normal population group, we can get the relative risk of smoking of smoker to nonsmoker group as:

RRsmoker/nonsmoker=

1

(18)

Chapter 4

Data and Estimation

4.1 Data

4.1.1 Data from CPS-II

According to the model structure, we assume that the CPS-II estimates of lung cancer death rates for smokers and nonsmokers can be applied to other populations. We use these two estimates as the benchmark when we assess relative increase or decrease of mortality rate of population of interest to lung cancer mortality rate for nonsmokers or smokers. We also use these two estimates to quantify the mortality rate of people who died from other causes after we investigated the macro-level statistical association between lung cancer mortality and mortality from all other causes of death based on the data of population of interest. We apply the lung cancer death rates (by sex and 5-year age group) among those observed among individuals in the CPS-II study (1982-1988) who never smoked regularly (used as M_L∗) or smoked (used as M_S∗) (Thun et al., 1997). These age-specific lung cancer death rates are presented in Table4.1.

Table 4.1: Age-specific lung cancer death rates (per 1,000) Lifelong Nonsmoker Smoker

Age Men Women Men Women

50-54 0.06 0.06 1.15 0.65 55-59 0.05 0.07 2.06 1.20 60-64 0.12 0.12 3.61 1.77 65-69 0.22 0.17 5.82 2.86 70-74 0.35 0.31 9.09 3.10 75-79 0.52 0.33 11.18 4.00 80-84 0.89 0.58 12.28 4.18

4.1.2 Data across Europe

We use the statistics data on causes of death from Eurostat1, which provide mortal-ity information on developments over time in the underlying causes of death and also highlight geographical differences. The newest data set consists of annually death data from 37 countries and districts between 1994 and 2013 and is clustered by: a shortlist of 86 causes of death based on the International Statistical Classification of Diseases and Related Health Problems (ICD), developed and maintained by the World Health Organization (WHO); sex; age; geographical region (NUTS level 2).

Statistics on causes of death are based on two pillars: medical information contained on death certificates, which may be used as a basis for ascertaining the cause of death;

1_{http://ec.europa.eu/eurostat/web/health/causes-death/data/main-tables}

(19)

How does smoking impact mortality rate in older age — Diandian Yi11

the coding of causes of death following the International Statistical Classification of Diseases and Related Health Problems (ICD). The medical certification of death is an obligation in all member states. The information provided in the medical certificate is used to code the cause of death.

To calculate the impact smoking casts on lung cancer death rate (first assumed impact way mentioned in model part) and estimate parameters of the model (second assumed impact way mentioned in model part, model about statistic relationship be-tween lung cancer mortality and mortality from other cause), we used annual data by sex and 5-year age groups (50-54,...,80-84) for 37 countries and districts from 1994 to 2013.

To look into the variation within areas in Europe, we also use annual data by sex and 5-year age groups for 11 western and Northern countries2from 1994 to 2013. We compare these two estimates based on two samples above and look into the trend based on data across Europe and based on data of 11 Western and Northern European countries.

4.2 Estimation

Based on the model structure, we estimate the relationship between lung cancer mor-tality and mormor-tality from other causes using the following Poisson Model:

E[M_a,s,tO |M_a,s,tL , Xa, Xs, T, Ma,s,tL ×Xa] = exp(βLMa,s,tL +βaXa+βsXs+βtT +βaL(Ma,s,tL ×Xa)),

(4.1) where a indexes the age group, s the country, t the investigation year. Thus, we con-trol for age group fixed effects (Xa), country fixed effects (Xs), a linear time trend of

investigation year (T ) and effects of interactions between lung cancer mortality and age group (M_a,s,tL × Xa). Ma,s,tL and Ma,s,tO are death rates for lung cancer and other

causes, respectively. Since M_a,s,tO is constructed as count data, we use poisson regres-sion to estimate the model, where we assume that M_a,s,tO given the vector of regressors

M_a,s,tL , Xa, Xs, T, Ma,s,tL × Xa is Poisson distributed with density

f (M_a,s,tO |M_a,s,tL , Xa, Xs, T, Ma,s,tL × Xa) =

eµa,s,t_µM O a,s,t a,s,t

M_a,s,tO ! (4.2)

and mean parameter µa,s,t = exp(βLMa,s,tL + βaXa+ βsXs+ βtT + βaL(Ma,s,tL × Xa)). We

estimate the Poisson model using conditional (fixed effects) quasi-maximum likelihood estimator (QMLE). (More details are presented in in Appendix A and the likelihood function of poisson model is presented in equation 6.4.) The second reason of using poisson regression is that the death rates M_a,s,tO are non-negative, leading to skewed distribution. Other methods, such as OLS with Log Transformation, can also be used here. However, interpretation of the magnitude of the effect of regressors would be difficult (see Cameron et al., (2013)). We also access the robustness of specification by considering the Negative Binomial Model, where the Poisson Model is a nested special case.3 We only report results of poisson regression here, because Negative Binomial Models yield almost identical results. Comparing to the Negative Binomial Model, the Poisson Model is also more robust to distributional misspecification provided that the conditional mean is specified correctly (see Cameron et al., (2013) and Wooldridge (2002)).

2

We choose countries according to the overlap of European Union members before 2004 and Western and Northern Europe defined by United Nations Statistics Division geoscheme, including Belgium, Germany, Denmark, Spain, Finland, France, Ireland, Netherlands, Portugal, Sweden, and U.K.

3

In the Negative Binomial Model, the variance to mean ratio is defined as 1 + δ where δ is the over dispersion parameter. If δ = 0, the model collapses to a Poisson where the variance equals the mean.

(20)

We apply the model with the data set across Europe and also the subset of the data across Western and Northern Europe respectively. We estimate models for male and female separately.

4.2.1 Estimation based on data across Europe

Appendix B and Appendix C show detailed estimation results based on data across Europe for male and female respectively. Here we only show key estimates here and discuss the goodness of fit.

Table 4.2presents estimated coefficients from the model in Eq. (4.1). If exponenti-ated, they can be interpreted as the proportional increase in the death rate of causes other than lung cancer associated with an increase in the lung cancer death rate of 1 per 1,000, all else being equal. Coefficients are smaller at higher ages, reflecting both higher death rates overall and more varied factors influencing mortality.

Table 4.2: Estimated coefficients (βL+ βaL × Xa) for lung cancer death rate

by age and sex

Age Men Women

55-59 0.342 0.176 60-64 0.226 0.148 65-69 0.143 0.092 70-74 0.080 0.058 75-79 0.045 0.010 80-84 0.035 0.001

Note: Estimated using poisson regression in Eq. (4.1) using data across Europe. The exponential of the above coefficients represents the proportional increase in the death rate for other causes associated with a 1 per 1,000 increase in the lung cancer death

rate. Thus they are calculated as βL+ βaL× Xa, include both coefficient of Ma,s,tL

and coefficient of interaction term from Eq. (4.1). As a dummy variable, Xa is 1 for

specific age group, is 0 for other age group. That means for every age group, βaLis the

corresponding coefficient for the specific age group when we looked at the interaction term.

As we consider the fitness of the estimation, by conducting analysis of deviance, we compared nested models to see if variables are significant. According to ANOVA table

4.34 _{we will show below, we found those fixed effects and effects of interaction terms} both improve model’s fitting effectively.

We find big variations between different regions in Europe. For example, in estima-tion results of female sample5_{, coefficients of countries and districts in western Europe,} such as Germany, France, and Netherlands, are positive; whereas that of Eastern Europe countries are more likely to be negative. The phenomena imply that the cross-region variation in Europe is of great interest. Thus we estimate the model using sub-sample including only Western and Northern Europe countries to see whether there are hetero-geneous results between different regions.

4.2.2 Estimation based on data of Western and Northern Europe

Appendix D and Appendix E show detailed estimation results based on data of Western and Northern Europe for male and female respectively.

4

We conduct analysis of deviance which is used to compare two nested models. The implication behind this analysis is similar to other criterions, such as AIC or BIC (see Kaas et al., (2008)). Added terms (in this paper, they are successively lung cancer mortality variable, age group fixed effects, country fixed effects, investigation year linear trend and interaction terms compared to the null model) not only increase the goodness of its fit but also maintain the concision of the model. We also compute AIC and BIC for models, and the results turn out to be the same outcome as the analysis of deviance.

(21)

Table 4.3: ANOVA table for estimation based on data across Europe Df Deviance Resid. Df Resid. Dev

Female NULL 3889 85128.00 Lung Cancer 1 10473.39 3888 74654.61 Age 6 67049.21 3882 7605.40 Country 36 5790.00 3846 1815.41 Year 1 1139.67 3845 675.73

Lung Cancer × Age 6 28.80 3839 646.93 Male NULL 3891 103961.54 Lung Cancer 1 53310.20 3890 50651.35 Age 6 38360.09 3884 12291.26 Country 36 8756.70 3848 3534.56 Year 1 1611.43 3847 1923.13

Lung Cancer × Age 6 518.38 3841 1404.75

Table4.4 presents estimated coefficients from the model in Eq. (4.1) based on data of Western and Northern Europe. They show similar trend as Table 4.2.

Table 4.4: Estimated coefficients (βL+ βaL × Xa) for lung cancer death rate

by age and sex

Age Men Women

55-59 0.232 0.553 60-64 0.154 0.384 65-69 0.102 0.251 70-74 0.070 0.169 75-79 0.044 0.098 80-84 0.028 0.059

Note: Estimated using poisson regression in Eq. (4.1) using data of Western and Northern Europe. The exponential of the above coefficients represents the proportional increase in the death rate for other causes associated with a 1 per 1,000 increase in the lung cancer death rate.

According to ANOVA table 4.5 we will show below, we found adding both fixed effects and interaction terms can effectively improve our model’s fitting.

Compared the sub-sample estimation with estimation based on the full Europe sam-ple, we found the coefficient of year variable for female based on data across Europe is -0.0234, and that for male is -0.0192; whereas the coefficient of year variable for female based on data of Western and Northern Europe is -0.0270, and that for male is -0.0240. Exponentiated coefficients implies that the mortality from other causes than lung cancer relation becomes decreasing every year, and the decreasing speed is much larger when it comes to male in Western and Northern Europe. More comparisons will be made in next chapter when RR has been calculated.

(22)

Table 4.5: ANOVA table for estimation based on data of Western and Northern Europe Df Deviance Resid. Df Resid. Dev

Female NULL 1203 20842.52 Lung Cancer 1 5582.70 1202 15259.82 Age 6 14419.24 1196 840.58 Country 10 392.48 1186 448.10 Year 1 334.74 1185 113.36

Lung Cancer × Age 6 46.18 1179 67.18 Male NULL 1203 29083.50 Lung Cancer 1 18625.99 1202 10457.52 Age 6 9256.36 1196 1201.16 Country 10 499.37 1186 701.79 Year 1 560.71 1185 141.08

(23)

Chapter 5

Results Interpretation

According to the model setting, we combine estimation results and data about lung cancer mortality of smoker/non-smoker from CPS-II to calculate the relative risk of smoker to nonsmoker (ratio of mortality rate of smoker to nonsmoker) respectively based on data across Europe and Western/Northern Europe1.

5.1 Trend of Relative Risk in Europe

Figure 5.1and Figure5.2show the age pattern of relative risk of smoker to nonsmoker based on data across Europe:

2.0 2.5

Y50-54 Y55-59 Y60-65 Y65-69 Y70-74 Y75-79 Y80-84 Age group

Relateive risk of smoker (to nonsmoker)

factor(Year) 1994 1999 2004 2009 2013

RR of smoker (to nonsmoker) for male

Figure 5.1: Trend of Relative Risk of smoker/nonsmoker with aging for male Corresponding adjustment coefficients (smoker/non-smoker) for male at different age groups for different investigation years are shown in Table 5.1:

Corresponding adjustment coefficients (smoker/non-smoker) for female at different age groups for different investigation years are shown in Table 5.2:

5.2 Trend of Relative Risk in Western and Northern

Eu-rope

Figure 5.3and Figure5.4show the age pattern of relative risk of smoker to nonsmoker based on data in Western/Northern Europe:

1

Other than analysis in this chapter, more intermediate results, such as RRsmoker/population and RRpopulation/smoker, would be shown in the Appendix F.

(24)

1.2 1.4 1.6 1.8

factor(Year) 1994 1999 2004 2009 2013

RR of smoker (to nonsmoker) for female

Figure 5.2: Trend of Relative Risk of smoker/nonsmoker with aging for female

Table 5.1: Relative risk of male smoker/nonsmoker for different age gruop 1994 1999 2004 2009 2013 Y50-54 2.47 2.30 2.41 2.39 2.42 Y55-59 2.64 2.61 2.61 2.68 2.70 Y60-64 2.83 2.76 2.77 2.85 2.91 Y65-69 2.71 2.68 2.72 2.76 2.84 Y70-74 2.34 2.36 2.40 2.47 2.48 Y75-79 1.81 1.82 1.84 1.88 1.90 Y80-84 1.60 1.61 1.62 1.64 1.65

Table 5.2: Relative risk of female smoker/nonsmoker for different age gruop 1994 1999 2004 2009 2013 Y50-54 1.52 1.56 1.54 1.61 1.71 Y55-59 1.57 1.61 1.68 1.70 1.78 Y60-64 1.59 1.61 1.65 1.71 1.75 Y65-69 1.54 1.56 1.61 1.66 1.70 Y70-74 1.33 1.34 1.36 1.39 1.41 Y75-79 1.14 1.14 1.16 1.18 1.19 Y80-84 1.06 1.05 1.06 1.07 1.08

(25)

How does smoking impact mortality rate in older age — Diandian Yi17 1.6 1.8 2.0 2.2 2.4

factor(Year) 1994 1999 2004 2009 2013

RR for male in Western/Northern Europe

Figure 5.3: Trend of Relative Risk of smoker/nonsmoker with aging for male in West-ern/Northern Europe

1.6 2.0 2.4 2.8

factor(Year) 1994 1999 2004 2009 2013

RR for female in Western/Northern Europe

Figure 5.4: Trend of Relative Risk of smoker/nonsmoker with aging for female in West-ern/Northern Europe

(26)

Corresponding adjustment coefficients (smoker/non-smoker) for male at different age groups for different investigation years in Western/Northern Europe are shown in Table5.3:

Table 5.3: Relative risk of male smoker/nonsmoker for different age gruop in West-ern/Northern Eruope 1994 1999 2004 2009 2013 Y50-54 2.00 1.97 2.05 2.03 2.12 Y55-59 2.03 2.07 2.10 2.19 2.19 Y60-64 2.11 2.15 2.18 2.25 2.27 Y65-69 2.14 2.18 2.23 2.27 2.34 Y70-74 2.17 2.20 2.24 2.30 2.36 Y75-79 1.81 1.82 1.86 1.90 1.93 Y80-84 1.51 1.50 1.52 1.50 1.55

Corresponding adjustment coefficients (smoker/non-smoker) for female at different age groups for different investigation years in Western/Northern Europe are shown in Table5.4:

Table 5.4: Relative risk of female smoker/nonsmoker for different age gruop in West-ern/Northern Eruope 1994 1999 2004 2009 2013 Y50-54 2.21 2.17 2.08 2.24 2.50 Y55-59 2.61 2.48 2.60 2.65 2.77 Y60-64 2.42 2.39 2.45 2.50 2.60 Y65-69 2.46 2.43 2.47 2.57 2.61 Y70-74 1.85 1.83 1.90 1.92 1.94 Y75-79 1.56 1.57 1.61 1.64 1.65 Y80-84 1.30 1.30 1.31 1.32 1.33

In general, smoking raises people’s mortality rate in the trend that the relative risk to non-smoker people becomes increasingly severe when people gets older until it reach a peak, and then, the relative risk diminishes. That means after people live until the age of peak relative risk, their mortality rate would become more and more similar to nonsmoker. Based on the data’s restriction, we only apply the data set with people aged below 84. For example, if we looked at the trend in Western/Northern countries, for male, the peak emerges at 70-74 years old and at 2.20 times, and then relative risk declines to around 1.5 times when people gets to 80-84 years old; for female, the peak emerges at 55-70 years old and at around 2.60 times, and then relative risk declines to around 1.3 times when people gets to 80-84 years old. For the whole Europe area, for male, the peak emerges at 60-65 years old and at around 2.8 times, and then relative risk declines to around 1.6 times when people gets to 80-84 years old; for female, the peak emerges at 55-65 years old and at around 1.7 times, and then relative risk declines to around 1.0 times when people gets to 80-84 years old. The impact of smoking on older age people obviously is not a linear trend, and more sophisticated adjustments at different age are more reasonable.

From above estimates, we find that extensive regional variations exist between West-ern/Northern European countries and the rest of Europe in relative risk of smoker to non-smoker. We find that the difference of relative risk between different gender in rest of Europe is much bigger than that in Western/Northern European countries. When we looked at the data across the Europe, the relative risk for male ranges from 1.5 times to 3 times, and the relative risk for female ranges from 1 times to 1.8 times. However, the relative risk for male in Western/Northern European countries ranges from 1.5 times

(27)

to 2.3 times, and for female this range is from 1.3 times to 2.6 times. This could be ex-plained by the much more convergence between male and female in smoking behavior in Western/Northern European countries than that in the rest of Europe. But along with the development of globalization and people’s behavior mode converges increasingly, we can make the forecast that the future variation should become smaller.

5.3 Implication of Change in RR’s Level among Different

Investigation Year

By breaking down the relative risk of smoker to non-smoker:

RR_{smoker/nonsmoker} = 1

1 − A×AS = RRnormalpopulation/non−smoker×RRsmoker/normalpopulation, (5.1) The formal part which is RRnormalpopulation/non−smoker is impacted by both the

prevalence of smoking in normal population and the heaviness of smoking for smok-ers; whereas RR_{smoker/non−smoker} is only impacted by the heaviness of smoking for smokers. Based on data across Europe, we also looked at the trend of relative risk of normal population to non-smokers:

1.2 1.4 1.6 1.8

Relateive risk of population (to nonsmoker)

factor(Year) 1994 1999 2004 2009 2013

RR of population/nonsmoker for male

Figure 5.5: Trend of Relative Risk of population/nonsmoker with aging for male across Europe

Compared to the plot showing the trend of relative risk of smoker to nonsmoker, we find that for different years of investigation, the level of relative risk of normal popula-tion to nonsmoker for male decreased from 1994 to 2003, while the level of relative risk of normal population to nonsmoker for female increased from 1994 to 2003. Because the heaviness of smoking both slightly increased based on the trend of male and female relative risk of smoker to nonsmoker, these trends together suggest that women’s preva-lence of smoking is increasing while men’s is decreasing while the heaviness of smoking is increasing.

(28)

1.05 1.10 1.15

Relateive risk of population (to nonsmoker)

factor(Year) 1994 1999 2004 2009 2013

RR of population/nonsmoker for female

Figure 5.6: Trend of Relative Risk of population/nonsmoker with aging for female across Europe

(29)

Chapter 6

The Future

This paper analyzes the trend of smoking impact on mortality rate of people older than 65. Applied data investigated from 1994 to 2013, we figured out the pattern shown last chapter. But these age group people who are older than 50 in our paper started to smoke early. If we care about the impact smoking will cast on people smoke now and become older in the future, we need to consider projection mortality adjustment. And because the smoking behavior between gender and age group are different now, more adjustment may required to be incorporated into our projection model.

Besides, this paper employed and developed PGW method(Preston, Glei and Wilmoth, 2010) to specify and quantify attributable smoking risk. Based on country-level death cause data, we looked at trend in different regions including several countries and dis-tricts. If in the future there can be data in zip code district level in one country, we can use this method to figure out relative risk adjustment for a country’s mortality table.

References

Andrew Fenelon, Samuel H. Preston, (2012). “Estimating Smoking-Attributable Mortality in the United States.” Demography, 49, 797-818.

Janet Currie, Hannes Schwandt, (2016). “Mortality Inequality: The Good News from a County-Level Approach.” Journal of Economic Prespectives, 30(2), 29-52.

Silvia Balia, Andrew M. Jones, (2008). “Mortality, Lifestyle and Socio-Economic Status.” Journal of Health Economics, 27(1), 1-26.

J. N. Hobcraft, J. W. McDonald and S. O. Rutstein, (1984). “Socio-Economic Factors in Infant and Child Mortality: A Cross-National Comparison.” Population Studies, 38(2), 193-223.

Evalyn N. Grant, Christopher S. Lyttle and Kevin B. Weiss, (2000). “The Relation of Socioeconomic Factors and Racial/Ethnic Differences in US Asthma Mortality.” American Journal of Public Health, 90(12), 1923-1925.

Seppo Koskinen, Tuija Martelin, (1994). “Why are Socioeconomic Mortality Differ-ences Smaller among Women than among Men?” Social Science Medication, 38(10), 1385-1396.

Leonid A. Gavrilov, Natalia S. Gavrilova, (2011). “Mortality Measurement at Advanced Ages: A Study of the Social Security Administration Death Master File.” N Am

(30)

Actuar J, 15(3), 432-447.

Peto, R., Lopez, A.D., Boreham, J., Thun, M., and Heath, C., Jr. (1992). “Mortality from tobacco in developed countries: Indirect estimation from national vital statis-tics.” Lancet, 339(8804), 1268-1278.

Preston, S. H., Glei, D. A., and Wilmoth, J. R. (2010). “A new method for estimating smoking-attributable mortality in high-income countries.” International Journal of Epidemiology, 39, 430-438.

Mokdad, A.H., Marks, J.S., Stroup, D.F., and Gerberding, J.L. (2004). “Actual causes of death in the United States, 2000”. Journal of the American Medical Association, 291(10), 1238-1245.

Staetsky, L. (2009). “Diverging trends in female old-age mortality: A reappraisal.” Demographic Research, 21(30), 885-914.

Mesl, F. and Vallin, J. (2006). “Diverging trends in female old-age mortality: The United States and the Netherlands versus France and Japan.” Population and Development Review, 32(1), 123-145.

Rau, R., Soroko, E., Jasilionis, D., and Vaupel, J.W. (2008). “Continued reductions in mortality at advanced ages.” Population and Development Review, 34(4), 747-768. Magdalena M. Muszynska, Agnieszka Fihel and Fanny Janssen (2014). “Role of

smoking in regional variation in mortality in Poland.” Addiction, 109, 1931-1941. Fanny Janssen, Valentin Rousson and Fred Paccaud(2015). “The role of smoking in

changes in the survival curve: an empirical study in 10 European countries.” Annals of Epidemiology, 25(2015), 243-249.

Fanny Janssen, Joop de Beer(2016). “Projecting future mortality in the Netherlands taking into account mortality delay and smoking.” Conference of European Statisti-cians, working paper 18.

Kaas, R., M.J. Goovaerts, J. Dhaene, and M.Denuit (2008). ‘ ‘Modern Actuarial Risk Theory—Using R”, 2nd edition, Springer, Heidelberg.

Cameron, A. Colin, Pravin K. Trivedi (2013). ‘ ‘Regression Analysis of Count Data”, 2nd edition, Cambridge University Press, Cambridge.

Wooldridge, Jeffrey M. (2002). ‘ ‘Econometric Analysis of Cross Section and Panel Data”, 1st edition, The MIT Press Cambridge, Massachusetts, London, England. U.S. Surgeon General. (2004). ‘ ‘The Heath Consequences of Smoking: A Report of

the Surgeon General.”, Washington, DC: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, National Center for Chronic Disease Prevention and Health Promotion, Office on Smoking and Health.

Thun, M.J., Day-Lally, C., Myers, D.G., Calle, E.E., Flanders, W.D., Zhu, B., et al.(1997). ‘ ‘Trends in tobacco smoking and mortality from cigarette use in cancer prevention studies I (1959 through 1965) and II (1982 through 1988).” In D.M.

(31)

Burns, L. Garfinkel, and J.M. Samet (Eds.), ‘ ‘Changes in Cigarette-Related Disease Risks and Their Implications for Prevention and Control, Smoking and Tobacco Control” (pp.305-382). NIH publication 97-4213. Bethesda, MD: Cancer Control and Population Sciences, National Cancer Institute, U.S. National Institutes of Health. Fanny Janssen (2005). ‘ ‘Determinants of Trends in Old-Age Mortality–Comparative

Studies among Seven European Countries over the Period 1950 to 1999”, Thesis Erasmus MC, University Medical Center Rotterdam, The Netherlands.

Public Health England. (2015). “Recent Trends in Life Expectancy at Older Ages.” https://www.gov.uk/government/uploads/system/uploads/attachment_ data/file/499252/Recent_trends_in_life_expectancy_at_older_ages_2014_ update.pdf, consulted on April 16, 2016.

Natalia S. Gavrilova, Leonid A. Gavrilov (2014). “Mortality Trajectories at Extreme Old Ages: A Comparative Study of Different Data Sources on U.S. Old-Age Mor-tality.” http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4318539/, consulted on March 17, 2016.

(32)

Appendix A: Poisson regression

To save spaces, we present details of the poisson regression using general notations here. Following Cameron et al., (2013), we assumes that yi given the vector of regressors xi

is independently Poisson distributed with density f (yi|xi) =

e−µi_µyi i

yi!

, yi = 0, 1, 2..., (6.1)

and mean parameter µi = exp(x0_iβ), where where β is a k × 1 parameter vector. This

model implies that the conditional mean is given by

E[yi|xi] = exp(x0iβ). (6.2)

It also implies a particular form of the conditional variance

V ar[yi|xi] = exp(x0iβ), (6.3)

which is equal to the conditional mean. Therefore, the log-likelihood function is L(β) = n X i=1 {yix0iβ − exp(x 0 iβ) − ln yi!}. (6.4)

To match general notations with our model in Chapter 4, yi we use here is Ma,s,tO in

Chapter 4 and xi used here represents Ma,s,tL , Xa, Xs, T, Ma,s,tL × Xa in Chapter 4. We

implement the estimator using the routine GLM from R. As argued in Cameron et al., (2013), the estimator of β in the GLM model can be interpreted as a QMLE, meaning that it is an MLE based on a possibly misspecified density. Although it should be clear that in general the terms QML have different meanings.

(33)

Appendix B: Estimation results

for estimation based on data

across Europe for female

(34)

Estimate Std. Error z value Pr(>|z|) (Intercept) 48.3275 1.3840 34.92 0.0000 lungDeath 0.2587 0.1464 1.77 0.0772 ageY55-59 0.3524 0.0537 6.56 0.0000 ageY60-64 0.7387 0.0498 14.83 0.0000 ageY65-69 1.2037 0.0459 26.23 0.0000 ageY70-74 1.7535 0.0440 39.82 0.0000 ageY75-79 2.3973 0.0423 56.73 0.0000 ageY80-84 3.0435 0.0418 72.87 0.0000 geoAT -0.3532 0.0759 -4.65 0.0000 geoBE -0.3425 0.0774 -4.42 0.0000 geoBG 0.2659 0.0748 3.55 0.0004 geoCH -0.5542 0.0766 -7.24 0.0000 geoCY -0.2144 0.0783 -2.74 0.0062 geoCZ 0.0024 0.0752 0.03 0.9740 geoDE -0.2968 0.0842 -3.53 0.0004 geoDK -0.2286 0.0784 -2.92 0.0035 geoEE 0.0204 0.0751 0.27 0.7860 geoEL -0.2624 0.0756 -3.47 0.0005 geoES -0.5052 0.0763 -6.62 0.0000 geoFI -0.3608 0.0762 -4.73 0.0000 geoFR -0.5906 0.0786 -7.51 0.0000 geoFX -0.6081 0.0770 -7.90 0.0000 geoHR 0.0879 0.0763 1.15 0.2493 geoHU 0.0886 0.0755 1.17 0.2408 geoIE -0.2275 0.0776 -2.93 0.0034 geoIS -0.4675 0.0790 -5.92 0.0000 geoIT -0.4838 0.0790 -6.13 0.0000 geoLT 0.0170 0.0751 0.23 0.8207 geoLU -0.3656 0.0761 -4.81 0.0000 geoLV 0.1364 0.0751 1.82 0.0693 geoMK 0.3900 0.0754 5.17 0.0000 geoMT -0.1774 0.0755 -2.35 0.0188 geoNL -0.3550 0.0821 -4.32 0.0000 geoNO -0.4064 0.0763 -5.32 0.0000 geoPL -0.0646 0.0781 -0.83 0.4085 geoPT -0.2517 0.0757 -3.33 0.0009 geoRO 0.2447 0.0753 3.25 0.0012 geoSE -0.4307 0.0762 -5.65 0.0000 geoSI -0.2003 0.0756 -2.65 0.0081 geoSK 0.1058 0.0751 1.41 0.1593 geoUK -0.2990 0.0777 -3.85 0.0001 geoLI -0.5607 0.0959 -5.85 0.0000 geoRS 0.3883 0.0828 4.69 0.0000 geoTR -0.0534 0.0877 -0.61 0.5427 year -0.0234 0.0007 -33.94 0.0000 lungDeath:ageY55-59 -0.0830 0.1650 -0.50 0.6150 lungDeath:ageY60-64 -0.1111 0.1514 -0.73 0.4631 lungDeath:ageY65-69 -0.1669 0.1466 -1.14 0.2548 lungDeath:ageY70-74 -0.2003 0.1455 -1.38 0.1685 lungDeath:ageY75-79 -0.2486 0.1452 -1.71 0.0869 lungDeath:ageY80-84 -0.2579 0.1450 -1.78 0.0753

(35)

Appendix C: Estimation results

for estimation based on data

across Europe for male

(36)

Estimate Std. Error z value Pr(>|z|) (Intercept) 40.1373 1.0625 37.78 0.0000 lungDeath 0.5311 0.0398 13.35 0.0000 ageY55-59 0.2987 0.0487 6.13 0.0000 ageY60-64 0.6757 0.0472 14.33 0.0000 ageY65-69 1.1308 0.0466 24.24 0.0000 ageY70-74 1.7011 0.0446 38.11 0.0000 ageY75-79 2.2960 0.0424 54.20 0.0000 ageY80-84 2.8576 0.0395 72.41 0.0000 geoAT -0.2759 0.0629 -4.39 0.0000 geoBE -0.3948 0.0647 -6.10 0.0000 geoBG 0.2558 0.0622 4.11 0.0000 geoCH -0.4374 0.0634 -6.90 0.0000 geoCY -0.3120 0.0654 -4.77 0.0000 geoCZ -0.0614 0.0627 -0.98 0.3275 geoDE -0.2671 0.0693 -3.85 0.0001 geoDK -0.2412 0.0631 -3.82 0.0001 geoEE 0.0913 0.0626 1.46 0.1443 geoEL -0.4386 0.0633 -6.93 0.0000 geoES -0.4277 0.0632 -6.76 0.0000 geoFI -0.2266 0.0631 -3.59 0.0003 geoFR -0.4322 0.0646 -6.70 0.0000 geoFX -0.4282 0.0635 -6.75 0.0000 geoHR 0.0139 0.0637 0.22 0.8274 geoHU 0.0105 0.0628 0.17 0.8673 geoIE -0.2073 0.0632 -3.28 0.0010 geoIS -0.4321 0.0637 -6.78 0.0000 geoIT -0.4782 0.0656 -7.29 0.0000 geoLT 0.1029 0.0624 1.65 0.0990 geoLU -0.3585 0.0634 -5.66 0.0000 geoLV 0.2001 0.0625 3.20 0.0014 geoMK 0.1994 0.0630 3.17 0.0015 geoMT -0.2581 0.0630 -4.10 0.0000 geoNL -0.4213 0.0680 -6.20 0.0000 geoNO -0.2903 0.0629 -4.61 0.0000 geoPL -0.0693 0.0649 -1.07 0.2855 geoPT -0.1214 0.0627 -1.94 0.0527 geoRO 0.1874 0.0627 2.99 0.0028 geoSE -0.2915 0.0630 -4.62 0.0000 geoSI -0.1678 0.0628 -2.67 0.0076 geoSK 0.0947 0.0626 1.51 0.1300 geoUK -0.3248 0.0632 -5.14 0.0000 geoLI -0.3653 0.0773 -4.73 0.0000 geoRS 0.1817 0.0695 2.61 0.0090 geoTR -0.1979 0.0732 -2.70 0.0069 year -0.0192 0.0005 -36.26 0.0000 lungDeath:ageY55-59 -0.1895 0.0446 -4.25 0.0000 lungDeath:ageY60-64 -0.3048 0.0414 -7.36 0.0000 lungDeath:ageY65-69 -0.3878 0.0405 -9.58 0.0000 lungDeath:ageY70-74 -0.4507 0.0401 -11.25 0.0000 lungDeath:ageY75-79 -0.4865 0.0400 -12.18 0.0000 lungDeath:ageY80-84 -0.4966 0.0399 -12.44 0.0000

(37)

Appendix D: Estimation results

for estimation based on data of

Western and Northern Europe

for female

Estimate Std. Error z value Pr(>|z|) (Intercept) 55.1502 2.8946 19.05 0.0000 lungDeath 0.7629 0.3600 2.12 0.0341 ageY55-59 0.2996 0.1255 2.39 0.0170 ageY60-64 0.6861 0.1147 5.98 0.0000 ageY65-69 1.1440 0.1075 10.64 0.0000 ageY70-74 1.7014 0.1036 16.42 0.0000 ageY75-79 2.3658 0.1010 23.43 0.0000 ageY80-84 3.0691 0.0997 30.78 0.0000 geoDE 0.0519 0.0483 1.07 0.2825 geoDK -0.0254 0.0491 -0.52 0.6052 geoES -0.1267 0.0346 -3.67 0.0002 geoFI -0.0133 0.0325 -0.41 0.6825 geoFR -0.2226 0.0381 -5.84 0.0000 geoIE 0.0275 0.0418 0.66 0.5108 geoNL -0.0691 0.0465 -1.49 0.1371 geoPT 0.1244 0.0328 3.79 0.0002 geoSE -0.1183 0.0329 -3.59 0.0003 geoUK -0.0612 0.0438 -1.40 0.1624 year -0.0270 0.0014 -18.69 0.0000 lungDeath:ageY55-59 -0.2099 0.3827 -0.55 0.5834 lungDeath:ageY60-64 -0.3792 0.3577 -1.06 0.2892 lungDeath:ageY65-69 -0.5115 0.3543 -1.44 0.1488 lungDeath:ageY70-74 -0.5938 0.3542 -1.68 0.0937 lungDeath:ageY75-79 -0.6648 0.3547 -1.87 0.0609 lungDeath:ageY80-84 -0.7044 0.3545 -1.99 0.0469 29

(38)

Appendix E: Estimation results

for estimation based on data of

Western and Northern Europe

for male

Estimate Std. Error z value Pr(>|z|) (Intercept) 49.5467 2.2797 21.73 0.0000 lungDeath 0.3985 0.1639 2.43 0.0151 ageY55-59 0.3699 0.1227 3.01 0.0026 ageY60-64 0.7563 0.1261 6.00 0.0000 ageY65-69 1.2012 0.1166 10.30 0.0000 ageY70-74 1.7050 0.1054 16.18 0.0000 ageY75-79 2.2868 0.0999 22.89 0.0000 ageY80-84 2.8944 0.0954 30.34 0.0000 geoDE 0.1337 0.0423 3.16 0.0016 geoDK 0.1271 0.0265 4.79 0.0000 geoES -0.0566 0.0289 -1.96 0.0503 geoFI 0.1384 0.0308 4.50 0.0000 geoFR -0.0463 0.0347 -1.34 0.1818 geoIE 0.1642 0.0291 5.64 0.0000 geoNL -0.0162 0.0350 -0.46 0.6422 geoPT 0.2323 0.0402 5.77 0.0000 geoSE 0.0560 0.0423 1.33 0.1848 geoUK 0.0414 0.0273 1.52 0.1297 year -0.0240 0.0011 -21.28 0.0000 lungDeath:ageY55-59 -0.1664 0.1860 -0.90 0.3708 lungDeath:ageY60-64 -0.2444 0.1721 -1.42 0.1556 lungDeath:ageY65-69 -0.2965 0.1657 -1.79 0.0735 lungDeath:ageY70-74 -0.3285 0.1640 -2.00 0.0451 lungDeath:ageY75-79 -0.3547 0.1636 -2.17 0.0302 lungDeath:ageY80-84 -0.3707 0.1636 -2.27 0.0235 30

(39)

Appendix F: Intermediate

Results

We show intermediate results here, such as RRsmoker/populationand RRpopulation/nonsmoker

based on data across Europe; RRsmoker/population and RRpopulation/nonsmoker based on

data from Western/Northern Europe.

Table 6.1: RRsmoker/population based on data across Europe

male female 1994 1999 2004 2009 2013 1994 1999 2004 2009 2013 Y50-54 1.43 1.36 1.48 1.55 1.68 1.39 1.43 1.39 1.41 1.49 Y55-59 1.38 1.45 1.49 1.53 1.61 1.43 1.44 1.46 1.47 1.51 Y60-64 1.47 1.54 1.64 1.67 1.69 1.39 1.45 1.47 1.47 1.48 Y65-69 1.55 1.60 1.71 1.75 1.79 1.42 1.42 1.47 1.48 1.46 Y70-74 1.62 1.62 1.69 1.75 1.78 1.26 1.26 1.29 1.30 1.29 Y75-79 1.44 1.46 1.46 1.49 1.54 1.11 1.12 1.13 1.13 1.14 Y80-84 1.38 1.40 1.39 1.40 1.40 1.05 1.05 1.06 1.06 1.06

Table 6.2: RRpopulation/nonsmoker based on data across Europe

Male Female 1994 1999 2004 2009 2013 1994 1999 2004 2009 2013 Y50-54 1.72 1.69 1.64 1.55 1.44 1.09 1.09 1.11 1.14 1.14 Y55-59 1.92 1.80 1.75 1.75 1.68 1.10 1.12 1.15 1.15 1.18 Y60-64 1.92 1.79 1.69 1.70 1.73 1.14 1.11 1.12 1.17 1.18 Y65-69 1.75 1.67 1.59 1.57 1.59 1.09 1.09 1.10 1.12 1.16 Y70-74 1.45 1.46 1.42 1.41 1.39 1.06 1.06 1.06 1.08 1.09 Y75-79 1.26 1.25 1.26 1.26 1.24 1.02 1.02 1.03 1.04 1.05 Y80-84 1.16 1.15 1.16 1.17 1.18 1.01 1.01 1.01 1.01 1.01

Table 6.3: RRsmoker/population based on data from Western/Northern Europe

Male Female 1994 1999 2004 2009 2013 1994 1999 2004 2009 2013 Y50-54 1.50 1.48 1.52 1.55 1.69 1.79 1.79 1.66 1.68 1.87 Y55-59 1.46 1.50 1.49 1.57 1.61 1.94 1.95 1.93 1.87 1.93 Y60-64 1.47 1.53 1.55 1.61 1.64 1.81 1.86 1.87 1.81 1.81 Y65-69 1.49 1.53 1.62 1.65 1.72 1.89 1.92 1.95 1.95 1.90 Y70-74 1.59 1.58 1.66 1.74 1.79 1.52 1.53 1.56 1.57 1.55 Y75-79 1.45 1.43 1.46 1.51 1.58 1.42 1.39 1.40 1.40 1.41 Y80-84 1.32 1.30 1.32 1.32 1.35 1.26 1.24 1.24 1.23 1.21 31

(40)

Table 6.4: RR_{population/nonsmoker} based on data from Western/Northern Europe

Male Female 1994 1999 2004 2009 2013 1994 1999 2004 2009 2013 Y50-54 1.34 1.33 1.35 1.31 1.25 1.23 1.21 1.25 1.34 1.34 Y55-59 1.39 1.38 1.40 1.39 1.36 1.35 1.27 1.35 1.41 1.44 Y60-64 1.44 1.41 1.40 1.40 1.39 1.34 1.28 1.31 1.38 1.43 Y65-69 1.43 1.42 1.38 1.37 1.36 1.30 1.27 1.27 1.32 1.37 Y70-74 1.37 1.39 1.35 1.32 1.32 1.22 1.20 1.21 1.22 1.25 Y75-79 1.25 1.27 1.27 1.26 1.22 1.10 1.13 1.15 1.17 1.17 Y80-84 1.14 1.15 1.15 1.17 1.15 1.03 1.05 1.06 1.08 1.09

How does smoking impact mortality rate in older age : investigation based on data across Europe