Issues in estimating smoking attributable mortality in Israel Free

Abstract

Introduction

Smoking is the biggest single preventable risk factor for mortality in developed countries,¹ annually causing close to five million deaths worldwide.² At the turn of the 21st century, Israel's male smoking prevalence (33%) was similar to that of neighboring Egypt (33.0%), Italy (32.4%), Denmark (32.0%) and Norway (31.0%), but was considerably lower than neighboring Jordan (48.0%), Japan (52.8%), Turkey (62.5%), Russian Federation (63.2%) and China (66.9%). On the other hand, Sweden (19.0%), the USA (25.7%), Finland (27.0%), Canada (27.0%) and the UK (27.0%), all had considerably lower male smoking prevalence than Israel.² For females, Israel's smoking prevalence (24%) was similar to that of Turkey (22.0%), Canada (23.0%) and the UK (26.0%), exceeded by Denmark (29.0%) and Norway (32.0%), but was considerably higher than developing countries like Egypt (1.6%), China (4.2%), Russian Federation (9.7%), Jordan (10.0%) and developed countries like Japan (13.4%), Sweden (19.0%), Finland (20.0%) and the USA (21.5%).

In Israel, with population 7.3 million in mid-2008,³ the first and only estimate of smoking attributable mortality (SAM) was submitted as evidence in 2001 to the national Gillon commission⁴ to decrease damage caused from smoking. The estimated deaths from active smoking in 1999 were 9527 and a further 1385 deaths of fetuses, children, spouses and workmates from passive smoking. These estimates assumed a 9-year lag period for each diagnosis, ignored elevated risks in ex-smokers and did not use age-specific categories.

In contrast to deaths which are clearly attributable to a given factor, for example, accidental deaths (including automobile accidents, suicides, drowning), deaths due to smoking are harder to identify. Despite this difficulty, smoking has been implicated as a factor in >90 causes of death. SAM and smoking attributable morbidity have been estimated using various approaches.^5–9 Peto's ‘indirect’ method¹⁰ used lung cancer rates to retroactively estimate smoking prevalence. Malarcher,⁹ calculated attributable fractions adjusted for age and other potential confounders. Thun¹¹ used the Cox proportional hazard model, incorporating a wide array of potential confounders. McNulty¹² used smoking status reports from death certificates. Our study chose the widely accepted population-attributable fraction (PAF) method,¹³ which uses relative risks (RR), prevalence and estimates of disease-specific mortality to estimate SAM.

The US Centres for Disease Control¹⁴ provides an online user-friendly computational program, called SAMMEC (Smoking Attributable Mortality, Morbidity and Economic Costs) to produce estimates of tobacco-related mortality.

Like many other countries other than the USA,¹⁵ we planned to use the SAMMEC computational program to obtain updated estimates of SAM in Israel. This had the advantage of allowing relatively fast computation, and enabling comparisons of our estimates with similarly produced estimates in other countries. Israel has not conducted large cohort studies of smoking and health on the general population, so disease-specific RR specific to Israel are not available. Instead, we relied on US-based RR that have been extensively used to estimate SAM, both globally,^16,17 and in various countries.¹⁵ Limitations of using US-based CPS-II estimates have been noted by other authors and apply to our work as well.¹⁶

However, serious limitations of the SAMMEC tool (which had been previously described on the SAMMEC website^5,14) became apparent. First, the lack of flexibility in deciding which diseases enter into the calculations. The diagnosis list¹⁴ is based on the 2004 US Surgeon General's Report,⁵ omitting diagnoses such as diabetes which have subsequently been proven to be smoking-related.¹⁸ Secondly, the estimates of disease risk due to smoking used in the SAMMEC calculation¹⁴ are based on data from two major US studies (Cancer Prevention Studies I and II), that were carried out in the 1980s. The enormous body of US and international literature on health risks due to tobacco use, which includes major studies¹⁹ and compendiums of studies,²⁰ has not been utilized. Because of the large variability in reported estimates of RR of diseases due to smoking,²¹ the advantage of using RR derived from more than one study⁶ has not been realized.

Furthermore, the standard SAMMEC approach does not consider the latency period that occurs between exposure to the risk factor (tobacco) and tobacco-induced mortality. For example, stroke deaths in smokers in 2003 are caused by decades of prior tobacco use, and are not a reflection of current smoking prevalence. Estimates of SAM using the CDC SAMMEC approach and presented in the 2004 US Surgeon General's Report⁵ are based on current levels of smoking. In contrast, other researchers have taken varying lag times into account in their calculations.²² This is particularly important because both the prevalence of smoking in most countries changes over time.

Finally, the SAMMEC approach calculates SAM only for adults aged >35 years. While SAM is low in the 20–35-year-old age group, it is not non-existent and should be included.

Thus, while SAMMEC provides a tool which is simple to use, it may provide inaccurate estimates of mortality due to smoking.

One of the possible reasons for underestimating or overestimating SAM is the possibility of errors resulting from the omission of lag-times. Among its various calculations, this article quantifies the extent of underestimation due to ignoring lag-times in the Israeli population.

This article aims to estimate the mortality from active smoking in Israel, using the following methods:

• Based on the SAMMEC set of RR and diseases applied to persons aged >20 years, with and without allowance for lag-times.

• Based on an expanded list of diseases making allowances for lag-times.

Methods

Prevalence

Representative smoking prevalence surveys in Israel began in 1974, when male and female prevalence were 50 and 35%, respectively.²³ Prevalence has gradually declined to 28.9 and 17.8% among males and females, respectively, in 2007.²⁴

Age (in 5-year groups from 20 to 85, then >85) and gender-specific estimates of smoking prevalence from 1974 to 2003 were obtained from unpublished and published surveys.^25–27 The results of a linear regression performed on the data for the years 1974–2000 were used to estimate prevalence in years when no survey occurred (i.e. 1977, 1978, 1980, 1982, 1984–86, 1990, 1998, 1999) and extrapolate to the prevalence from 1959 to 1973.

Age- and gender-specific prevalence data of former smokers were drawn from raw data analysis of the 2003/4 household survey carried out by the Central Bureau of Statistics.²⁶ Therefore, estimates on former smoking prevalence from 1959 to 2002 were based on a lagged model that took equally into account both the absolute decrease in smoking prevalence (i.e. a large decrease would mean there are more former smokers) and year-specific smoking prevalence (i.e. if smoking prevalence was 100%, then there would be no former smokers).

Mortality

Population³ and national mortality data from the last three available years (2000, 2001, 2003) were used to calculate diagnosis, age- and gender-specific mortality rates. These rates were then applied to the population structure of 2003, to provide estimates of diagnosis, age- and gender-specific mortality.

Selection of diagnoses and RR

For comparative purposes, we chose diagnoses contributing to SAM and diagnosis and gender-specific RR for current and former smokers:

• used in the SAMMEC model¹⁴ (table 1).

• based on the literature (table 2), including diagnoses taken from previous calculations of costs attributable to smoking in the USA,^27–30 Australasia^31,32 and the UK³³ augmented with information from a recent article.¹⁸ For the literature-based diagnoses, RR for former smokers were only available for a few cases¹⁸ so these were estimated from the proportional excess risk of former to current smokers used in SAMMEC model for the same diagnosis or wider disease category.

Lag-factors

We took into account the fact that smoking rates between 1959 and 2003 contribute to SAM in 2003, due to lag factors, as follows.

We assumed persons who died younger had shorter lag-times than persons who died at an older age. Estimates of the lag-time where made by subtracting the disease and gender-specific average age at death by the gender-specific average age of smokers.

We assumed a lag time for all diagnoses of 2.5 years for persons aged 20–24, 5 years for 25–29-year-olds, 7.5 years for 30–34-year-olds and 10 years for 35–39-year-olds. These shorter times reflect the fact that the lengthy lag-times for active smoking (e.g. 25–40 years) are non-feasible in young persons. A disease-specific age-related linear factor was applied to the 40–54-year age group (i.e. with lower than average time-lag). The average disease-specific lag-time was assumed to occur in the 55–59-year-age group. The average lag time estimates that we used varied from brain cancer (23.7 years), lung cancer (30.5 years), colon cancer (35.2 years) to 40.3 and 40.4 years for chronic obstructive pulmonary disease (COPD) and coronary artery disease, respectively. A disease-specific age-related linear factor was also applied to the >60 age group (i.e. with increasing above average lag-time) within the constraints of achieving the overall average disease and gender-specific time lag.

Next, for each diagnostic, age and gender group, smoking prevalence was obtained for the year relating to 2003 less the lag-time. For example, males aged 75–79 had an estimated lag-time of 34 years for mortality due to bronchitis or emphysema, so estimates were obtained from smoking prevalence rates for persons who were between 40 and 44 years old in 1969 (i.e. 2003 less 34 years).

Calculation of SAM

This was based on the formula applied to each diagnosis, age and gender category and finally aggregated

where

and

First, estimates from the basic SAMMEC model, were expanded slightly by applying the RR to all those aged 20–34 in addition to those aged >35. Estimates of SAM were calculated using the above formula based on un-lagged smoking prevalence rates for 2003 (table 3). A second SAMMEC estimate (table 3) and literature-based estimate (table 4) were calculated taking into account the lag-factors described in the previous section. Using identical methodology, including adjustments for population size, mortality rates and prevalence, we calculated an additional estimate of SAM for 1980 for the expanded literature-based table (including lag times).

The elasticity of SAM with respect to changes in excess risk (or protection) was based on the formula (% change in SAM)/(% change in excess risk) where excess risk is defined as (RR – 1).

Results

The un-lagged and lagged SAMMEC models estimate 3859 and 6275 deaths, respectively, attributable to smoking in 2003 in Israel (table 3). In the lagged model, cardiovascular diseases (account for 38.1% of SAM), malignant neoplasms (34.2%), respiratory diseases (17.0%) and cerebrovascular diseases (10.8%). Four specific diagnoses alone accounted for 69.0% of SAM: ischemic heart disease (24.9%), lung cancer (including trachea and bronchus, 20.5%), chronic airway obstruction (12.9%) and cerebrovascular disease (10.8%).

Lagged calculations based on smoking-related diseases and RR obtained from the literature (table 4) estimated SAM to be 8664 persons, including 172 fewer deaths, due to the protective effect of smoking on Parkinson's disease. Cardiovascular causes accounted for 42.6% of SAM, malignant neoplasms (29.1%) and respiratory diseases (13.8%). Six specific diagnoses alone accounted for 79.3% of SAM: coronary artery disease (25.1%), lung cancer (14.1%), COPD including emphysema (11.3%), diabetes (10.3%), cerebrovascular disease (5.5%) and myocardial infarction (5.2%). Inclusion of persons aged 20–34 accounted for an additional 68 deaths attributable to smoking, representing 0.8% of the total SAM. SAM in 1980 was estimated to be 7201 persons.

A sensitivity analysis, carried out by reducing the lag-time in the model by one-third, estimated that there would be 7997 smoking-attributable deaths, a decrease of 7.7% from the 8664 deaths of the original estimate.

We examined the elasticity of the estimated SAM in response to changes in the excess RR (or protection). A 1% increase in excess RR would cause a 0.37% rise in the SAM estimate from 8664 to 8699 deaths. A 10 or 30% increase or decrease across all RR would give SAM estimates ranges of 8324–8971 and 7515–9507, respectively.

Discussion

Using Israeli data, we have shown that estimation of SAM is influenced by three issues, besides the choice of the lower bound of the age range: the set of diseases included in the calculations, estimates of RR and the year of smoking prevalence. Estimates obtained through a SAMMEC-based approach underestimated mortality due to smoking. Introducing more appropriate lag times increased the SAM estimate based on the SAMMEC set of RR and diseases by 62.6% from 3859 to 6275. The expanded literature-based table (including lag times) estimated annual SAM to be 8664 deaths. This estimate is 20.3% higher than an estimate of 7201 of SAM in 1980, because the large population increase (from 3.9 to 6.7 million between 1980 and 2003), outweighed the 29% relative decrease in (lagged) smoking prevalence rates as well as technologically driven decreases in case fatality rates. Note that the 1980 estimates are likely to be less accurate as they are totally dependent on linear extrapolations of smoking estimates prior to 1973.

The ratio of SAMMEC-based US estimates of 440 000 SAM deaths³⁶ to Israeli SAM estimates after adding in estimates of 1100–1200 deaths from passive smoking (based on the Gillon Commissions estimates adjusted for prevalence decreases and increased compliance to smoking in public laws), reflect the over 40-fold smaller population size (6.7 million versus 291 million in 2003) and the slightly higher smoking prevalence rates of smoking in Israel.

Around half (50.3%) of the increase from the un-lagged SAM estimate of 3859 deaths (table 3) to the final SAM estimate of 8664 deaths (table 4) is attributable to the use of lag times. Use of estimates of risk obtained from a wider and more recent pool of studies accounted for a further 28.2%, while 20.0% is attributable to the inclusion of additional diseases whose relationship to tobacco use has been recently established (e.g. Diabetes). The remaining 1.5% was attributable to the widening of existing diseases categories.

The largest specific sources of the increased estimate of SAM are as follows:

• An increase in cardiovascular-related mortality (excluding cerebrovascular events) of 1306 deaths from 2389 (SAMMEC) to 3695 due primarily to the higher RR of the literature-based estimates.

• An increase of 2416 deaths due to introducing lagged prevalence data.

• The inclusion of diabetes as a SAM category accounted for an additional 896 deaths. This link was published in 2005,¹⁸ subsequent to the most recent update of the disease categories in SAMMEC.¹⁴

• An increase in the estimate of 133 respiratory and 66 stomach cancer deaths due primarily to the higher RR of the literature-based estimates.

• The use in the literature-based estimates of the wider category of leukemia (ICD 10, c91-c95) instead of acute myeloid leukemia (ICD 10, c92) together with higher female RR raised mortality estimates by 175 from 24 to 199.

• The inclusion of extra disease categories from the literature estimates such as 45 deaths from cancer of the ureter (ICD 10, c66), 176 deaths from unspecified cancer (ICD 10, c79, c80, z51.1, z54.2) and 46 deaths from diseases of digestive system, accounts for the remainder of the differences in mortality, despite the compensation for reduced mortality from Parkinson's disease.³²

Using appropriate lag-times is a crucial and complex element of accurate estimation of SAM. If tobacco use has already peaked and is declining, then using the current prevalence levels will underestimate the true mortality caused by tobacco. In Israel, where tobacco use has been decreasing for several decades, using an un-lagged approach produced a SAM underestimate of 38.5% on the SAMMEC categories.

The average lag-times used in our study are consistent with the finding of a recent major British study¹⁹ that showed the full effects of tobacco on national mortality rates can take more than 50 years to mature. However, using lag-times of one-third shorter duration, produced an estimate that was 7.7% lower than the baseline estimate.

Since we have no data on tobacco use in Israel at the beginning of the 1959–73 period, our use of linear extrapolation to estimate the prevalence during the period can be questioned. Use of quadratic extrapolation gave an estimate of SAM of 8785. The small difference of only 1.4% can be explained partly by the inherent linearity of the data and partly by the fact that most of the lagged estimates (especially among males) did not require use of estimated prevalence prior to 1974 and if they did, it was mainly for data from 1967 onwards.

A small underestimation of disease burden may be generated due to us excluding mortality of persons <20 years, still lower than SAMMEC’s underestimation that excludes persons <35 years.

To improve our model, we could incorporate additional age-specific RR. A partial adjustment incorporating gender- and age-specific RR for heart attacks, strokes, pneumonia, influenza as well as lung, breast, colon, cervical and prostate cancer,⁶ reduced our SAM by 3.8% to 8334.

The SAM estimates were sensitive to the values used for excess RR (point elasticity at baseline estimate = 0.37). However, it is likely that errors in the diagnosis-specific RR would be in both directions thus reducing the impact on the final SAM estimates.

SAM can be used to calculate PYLL (Potential Years of Life lost) due to active smoking. Combined with estimates of morbidity, these provide a basis for the calculation of the total disease burden attributable to smoking in terms of QALYs (Quality adjusted life years). Subsequently the QALY estimates of the burden can be multiplied the effectiveness, compliancy and coverage rates to estimate QALYs saved by various interventions. By integrating the QALY data with data relating to intervention costs and treatment costs saved, estimates of the cost per QALY of each intervention can be made in order to help decision makers prioritize what interventions they wish to fund.

We estimate that smoking causes 22.5% of mortality in Israel, similar to the 19.5% caused by SAM in the USA,^34,35 the higher Israeli estimate reflecting the slightly higher Israeli prevalence rates.

A number of issues would benefit from future research.

First, RR of tobacco based on meta-analyses would provide more accurate estimates than single study estimates. This would render redundant our use of published SAM studies that provided an ‘expanded set of diseases’, eliminating any bias caused by choosing such articles. At present, the 2004 IARC monograph²⁰ on cancer risks from active and involuntary smoking, provides lists of relevant studies, but doesn't provide aggregate measures of risk based on the studies reviewed. Such aggregate numbers could provide stable internationally based estimates of risk and thus be used in calculating the burden due to tobacco.

Secondly, longitudinal cohort studies could provide population-specific estimates on mortality risks in specific countries.

Thirdly, research on distributions of lag-times for development of various diseases is needed. This is a complex issue, and in our article we made simplifying generalizations.

Fourthly, the use of age-specific RR should be more widely adopted.

Because good data on past and present exposure to second-hand smoke are not currently available, this report does not include an estimate of the harm due to second-hand smoke exposure. Information of levels of such exposure at a population-wide level could help estimate the true burden of damage from tobacco.

Different approaches to the estimation of SAM in Israel produced varying results. Mortality was underestimated using the SAMMEC approach. In order to accurately calculate SAM, it is essential to include a comprehensive set of diseases caused by tobacco use; use stable updated estimates of the additional risk for each condition caused by tobacco, use information on national tobacco use from the time when diseases originate and develop, and to include people aged 20–35. The most comprehensive estimate showed that active smoking causes 8664 deaths in Israel or 22.5% of the total estimated mortality in 2003 of 38 449 persons. This estimate was much higher than the un-lagged (3859) and lagged (6275) estimates obtained using the SAMMEC approach. Whichever estimate is used, active tobacco use alone kills more Israelis than the combined mortality from diabetes, infectious and parasitic diseases, motor vehicle ‘accidents’, all external causes (including homicides), pneumonia, suicides, renal diseases, liver diseases, perinatal conditions and congenital anomalies. There are already established funded bodies like the ‘authority for the war against road accidents’ and the ‘authority for the war against drugs’. The considerable mortality burden (10–20 times higher than that caused by traffic-related mortality) attributable to smoking in this article cries out for the establishment of an ‘authority’ to co-ordinate to identify and implement a multi-faceted intervention strategy to decrease the considerable burden from smoking in Israel.

Acknowledgements

This work was carried out by two salaried employees of the Israeli Ministry of Health (GG, ER), and a faculty member at Tel Aviv University School of Public Health (LR), as part of the Ministry of Health-initiated Healthy Israel 2020 programme relating to health promotion and disease prevention. We would like to acknowledge the assistance of Ziona Haklaii and Nehama Stein of the Health Ministry's Statistical Unit in supplying the raw mortality data.

Conflicts of interest: None declared.

References