Journal of Public Health Medicine DOI: 10.

1093/pubmed/fdg077

Vol. 25, No. 4, pp. 318324 Printed in Great Britain

A model to predict the results of changes in smoking behaviour on smoking prevalence

John Kemm

Abstract
Background Data are available on the prevalence of smoking states (never, current and ex). However, data on behaviour change rates (starting never to current, quitting current to ex and lapsing ex to current) are not readily available and cannot be simply derived from or related to prevalence data. Method A model was constructed to relate prevalence of smoking states to behaviour change rates. It was populated with prevalence of smoking status taken from the General Household Survey together with population structure, ageand sex-specic death rates, and birth rates for England and Wales. This model could be used to calculate past behaviour change given observed prevalence of smoking states or future prevalence of smoking given predicted rates of behaviour change. Results To t data it was necessary to assume that as they age some ex smokers reclassify themselves as never smokers. In the age band 1619 years about 9 per cent of never smokers start smoking, and about 5 per cent of current smokers quit. In the age band 2024 years the corresponding gures for starting are about 4 per cent in males and 2 per cent in females, and for quitting about 2 per cent in both. In older age bands the percentages starting are zero or less than zero (indicating reclassifying), and the percentage quitting rises with age. Net lapsing (shift from ex to current) occurred very infrequently and is quantitatively unimportant. If the current starting, quitting and lapsing rates are maintained the Smoking kills target will not be met. Future prevalence of smoking under different scenarios is examined. Conclusion The model is useful in calculating the proportions changing smoking state from serial cross-sectional data on prevalence and for predicting future prevalence. Keywords: smoking, modelling

follow-up).3 Both prevalence and quitting targets are useful but the relationship between them is unclear. If the quitting targets are met will the prevalence targets also be met or is additional effort required? What is the precise relationship between proportions changing smoking behaviour (starting, quitting and lapsing) and future prevalence of smoking? If x per cent of smokers are helped to quit each year for the next 10 years, what will be the prevalence of smoking at the end? What percentage of smokers will have to quit each year to bring prevalence of smoking down to x per cent in 10 years time? What difference will it make if efforts to encourage quitting are targeted at the young or the old? This paper describes a computer model, which allows such questions to be explored. The only data available for most populations are prevalence data for smoking state (current smoker, ex smoker, or never smoked). Prediction of smoking prevalence based on projection of past trends4 may be misleading, and projections based on rates of starting and quitting5 are theoretically more sound.

Downloaded from http://jpubhealth.oxfordjournals.org/ at Universiti Sains Malaysia on April 4, 2014

The model
The model relates prevalence of smoking status (current, ex and never) to the possible behaviour changes of starting (never to current), quitting (current to ex) and lapsing (ex to current) (Fig. 1). It takes account of differences between gender and age groups and the effects of ageing, differential mortality and changing population structure. It can be run in three modes: historic, target and predictive. In historic mode rates of starting, quitting, lapsing and reclassifying for each age group are tted to observed prevalence rates. In target mode target prevalence rates of smoking are specied for a particular year and the model calculates the starting, quitting and lapsing rates that would produce a progression to that target value. In predictive mode the starting, quitting and lapsing rates that will operate over a future period

Introduction
The White Paper Smoking kills1 set a target for England to reduce the prevalence of smoking in adults to 24 per cent by 2010 (from a baseline of 28 per cent in 1996). This was much less ambitious than the target set earlier in Health of the nation2 to reduce the prevalence of smoking in adults to no more than 20 per cent by the year 2000. In England considerable effort has been put into smoking cessation services and for 20022003 they were required to deliver 100 000 successful quitters (at 4 weeks

John Kemm, Honorary Clinical Lecturer Department of Public Health and Epidemiology, University of Birmingham, Birmingham B15 2TT. Address correspondence to Dr Kemm. E-mail: john.kemm@doh.gsi.gov.uk

Journal of Public Health Medicine 25(4) Faculty of Public Health 2003; all rights reserved.

MODEL TO PREDICT SMOKING BEHAVIOUR

319

Figure 1 Model of smoking states and changes between them.

Downloaded from http://jpubhealth.oxfordjournals.org/ at Universiti Sains Malaysia on April 4, 2014

of years are specied and the model calculates the resulting prevalences of current, ex and never smokers. In predictive mode it is also possible to specify a bolus intervention with starting, quitting and lapsing rates that operate for a specied number of years before reverting to the underlying rates. A selection of screens and controls embedded in the worksheets allow the user to control all relevant parameters including the mode, the start date of the model (1954 or later), the end year for the run, the year to start predictions in predictive and target mode, the years over which rates are smoothed (in historic mode), the relative risks of mortality in smokers and ex smokers, how the prevalence of current and ex smokers at age 15 is calculated, future behaviour change in predictive mode, or target smoking prevalence for each age band in target mode. The user can also control output, which can be in graphic or tabular form. The model is written in VBA (Visual Basic for Applications) embedded in an Excel spreadsheet. The model is populated with information on prevalence of current, ex and never smokers from 1954, age composition of the population, death rates and birth rates. The structure of the model and the data sources are further described in the Appendix.

percentage quitting in the 1619 years group is also high. Net lapsing (change from ex smoker to current) is quantitatively unimportant and between 1974 and 2000 was observed in only 14 of the 200 age band years for men and 16 of the 200 age band years for women. All behaviour change rates show considerable volatility, even when smoothed by averaging over 3 years, and this is reected in the broad condence intervals.

Sensitivity testing
In historic mode the model is fairly sensitive to the assumptions made about the proportion of 15-year-olds who are smokers and ex smokers. If one sets these proportions at 0.7 of that in the 1619 years age band rather than 0.5 then the estimated percentage of 1619-year-old never smokers starting is 5.52 per cent (men) and 5.76 per cent (women) rather than 8.61 per cent and 9.10 per cent. The justication for choosing 0.5 is made in the Discussion. The results presented are based on an assumption of relative risk for all-cause mortality for current smokers compared with non-smokers of 1.7 in men and 1.3 in women. Assuming equal mortality (a relative risk of one) made little difference except in the oldest age group.

Validation of the model

Results
Historic mode
The percentages starting, quitting and reclassifying have been calculated from prevalences of current, ex and never smokers from 1954 to 2000. Table 1 shows for 19902000 the mean net percentages changing behaviours. It can be seen that about 9 per cent of 1619-year-old never smokers started each year as did about 4 per cent of 2024-year-old male never smokers and about 2 per cent of 2024-year-old female never smokers. In all older age groups the number of ever smokers decreases, indicating that net reclassication (ex smokers describing themselves as never smokers) has taken place. Above age 20 years the percentage of smokers who quit tends to rise with age. The

The model can be tested by comparing predictions based on a particular year in the past with subsequently observed prevalences. Figure 2 shows that when the average starting and quitting rates for 19902000 (taken from Table 1) are used to predict the prevalence of smoking after 1990 they give reasonably good agreement between the predicted and observed prevalence rates. Agreement for the age bands not shown was similar to that for the age bands shown in the gure. This demonstrates that when appropriate rates are used predictions based on the model approximate to the true state. Predicted prevalence of ex smokers is higher than observed because the phenomenon of reclassication means that rates observed by survey underestimate the true prevalence of ex smokers. Exact agreement between predicted and observed prevalence should not be

320

JOURNAL OF PUBLIC HEALTH MEDICINE

Table 1 Mean annual percentage starting, reclassifying and quitting, 19902000

% Starting Age band Men 1619 2024 2534 3544 4554 5564 6574 75+ Women 1619 2024 2534 3544 4554 5564 6574 75+ Mean 95% CL % Reclassifying Mean 95% CL % Quitting Mean 95% CL

8.61 4.43 0 0 0 0 0 0 9.10 2.16 0 0 0 0 0 0

8.199.03 2.636.23

0 0 1.21 1.84 0.40 1.97 0.38 1.73 0 0 1.36 4.29 0.21 10.49 3.55 7.74

1.403.81 1.154.83 0.891.69 0.233.71 0.401.16 0.892.57

4.69 2.19 1.93 1.86 3.54 4.40 6.18 4.84 5.51 1.78 2.88 2.55 2.39 5.63 5.81 9.66

3.735.65 1.462.92 1.252.61 1.142.58 2.174.91 2.266.54 4.747.62 0.918.77 4.716.31 0.812.75 1.454.31 1.223.88 1.403.38 4.157.11 3.877.75 6.2113.11

7.8910.31 1.323.00

Downloaded from http://jpubhealth.oxfordjournals.org/ at Universiti Sains Malaysia on April 4, 2014

1.053.77 1.237.35 2.052.47 8.2412.74 2.884.22 5.4410.04

The table shows mean percentage net changes in smoking behaviour. Smokers starting are shown as percentage of never smokers. Smokers reclassifying or quitting are shown as percentage of ex smokers. Net lapsing is not observed in any group. In calculation of condence interval, n has been taken as 5 not 10, as data were available only for even-numbered years. Run conditions: mortality relative risk (RR) of smokers was 1.7 for males and 1.3 for females; prevalence in 15-year-olds was 50 per cent of that in 1619-year-olds, rates smoothed over 3 years. CL, condence limit.

expected, as the assumption of constant rates of behaviour change over the prediction period is unlikely to be true and the observed prevalence rates are subject to sampling and other errors.

bands over 45 years will fall by much more than the overall target 15 per cent reduction but the falls in younger age bands will be considerably less than the target.

Prediction mode
In prediction mode the model calculates the effect of different proportions changing smoking status from a specied start year on future prevalence of never, current and ex smokers. Figure 3 shows the effect on smoking prevalence of three scenarios no change, doubling quit rates in smokers aged under 25, and doubling quit rates in all ages at the same time as halving rates of starting smoking. A doubling of quit rates is of a similar order to that obtained in smoking cessation trials.6 Interventions directed at those under 25 years old (halving the rate of starting (not shown) or doubling the quit rates (shown)) have more effect than increasing the quit rate in older (over 25 years old) smokers (scenario not shown in gure). A sustained combination of doubling the quit rate for all ages and halving the start rate would reduce the prevalence of smoking to about 12 per cent by 2020. If the current starting and quitting rates (shown in Table 1) are continued the Smoking kills target of reducing prevalence by to 24 per cent by 2010 will not be met. On current trends in 2010 the prevalence of smokers will be 26.7 per cent in males and 24.1 per cent in females (European Standard Population). The age

Discussion
Use of the model in historic mode suggests that the proportion of never smokers who start and the proportion of current smokers who quit have changed little over the past 25 years. It is encouraging that the proportion of smokers who quit has not fallen despite a decrease in the pool of smokers. The quantitative unimportance of lapsing is also encouraging. The 1619 years group is very volatile, with high percentages of never smokers starting and of current smokers quitting. Some people start smoking after the age of 19, although very few do so after the age of 24. This agrees with ndings in the General Household Survey (table 8.24)7 but differs from the assumption for models of US populations that no smoking starts after age 18.8 The percentage of current smokers quitting rises with increasing age and this tallies with experience of older people giving up smoking for a variety of reasons. This analysis also puts quit smoking targets in proportion. Each year about 300 000 people shift from being current to ex smokers, i.e. quit smoking. The yearly target for quit smoking programmes is 100 000 quitters at 4 weeks, of whom a considerable number will not continue to become long-term quitters.9

MODEL TO PREDICT SMOKING BEHAVIOUR

321

Downloaded from http://jpubhealth.oxfordjournals.org/ at Universiti Sains Malaysia on April 4, 2014

Figure 2 Comparison of predicted and observed prevalence of smoking, for (a) males and (b) females, 19901998. Prediction from 1990 data applying mean rates given in Table 1. Only rates for 2024, 3544, 6574 and 75+ years age bands are shown. Other age bands have been omitted for clarity. The continuous line indicates the observed rates and the dashed line the predicted rates.

322

JOURNAL OF PUBLIC HEALTH MEDICINE

Downloaded from http://jpubhealth.oxfordjournals.org/ at Universiti Sains Malaysia on April 4, 2014

Figure 3 Projection of numbers of smokers. The gure show the projected percentage of men (a) and women (b) smokers (age adjusted to European Standard Population age 16 years) assuming three scenarios: no change; percentage quitting doubled in age bands under 25, percentage starting unchanged; percentage quitting doubled in all age bands, percentage starting halved. The percentages assumed in the no change model for the nine age bands (rst band 15 years, second band 1619 years, etc.) for starting were (in per cent): 14, 9, 4, 0, 0, 0, 0, 0 and 0, and for quitting 2.5, 5, 2, 2.5, 3, 3.5, 5, 6 and 7.

MODEL TO PREDICT SMOKING BEHAVIOUR

323

Using the model in prediction mode provides insights into how smoking patterns will evolve in the future and indicates that greater efforts are required if the modest Smoking kills targets are to be met by 2010. Smoking prevalence rates are the product of cohort effects, behaviour change with age and differential mortality. Ideally, one would want longitudinal data on smoking to deconstruct these elements but such data are usually unavailable. Cohort studies1015 can separate age and cohort effects but without further analysis do not indicate the proportion changing smoking state. Studies based on retrospective reports are subject to recall and other bias, whereas those using observed cohorts may not make full use of data that do not t the age interval. The quality of the data used to populate the model must be considered. The smoking data since 1974 were taken from the General Household Survey and are based on a well-established methodology and reasonable sample sizes. Less reliance can be placed on the Tobacco Advisory Council data used for years before 1974 but errors in these will have very little effect on the output of the model for 2000. There is a need for better data on smoking in 15-year-olds, especially as the model is sensitive to the assumptions made here. Analysis of data on 16-year-olds from the General Household Survey (19841998 pooled data, 1015 boys and 968 girls) shows 21 per cent to be current smokers, a prevalence that is 0.72 of the overall prevalence for the whole 1619 year age band. Surveys of secondary school children in England16 showed 23 per cent (range 1728 per cent) of 15-year-old boys and 27 per cent (range 2233 per cent) of girls to be current smokers. The two surveys are subject to different errors and social biases, and, with results so different, it is clear that they are measuring the prevalence of different states. In the absence of more reliable data, the assumption that the prevalence of current smoking and ex smoking in 15-year-olds was half that in the 1619 year age band is not unreasonable. The model works on net ows. It does not distinguish between balanced two-way ows and no change. For example, the situation in which equal numbers quit and lapse cannot be distinguished from the situation in which no one quits or lapses. Longitudinal data would be needed to distinguish these two situations. The instability in historic mode is explained by sampling errors in the underlying data, which are incompletely smoothed by the model. Mendez et al.8 have applied a similar model to US data. However, they used eight agesex groups as opposed to the 18 used in this model, and assumed no starting of smoking after age 18 and quitting rates that changed only once. Their approach produces much more stable estimates and still achieves very good t.

smokers who quit have not changed substantially over the past 20 years. The model also shows that above age 20 very few start smoking, that the percentage of smokers quitting rises with age and that net lapse from ex smoker to current is very infrequent. In prediction mode the model shows that on current behavioural trends the Smoking kills targets will not be achieved in 2010.

Disclaimer
The General Household Survey data are crown copyright. Neither the Ofce for National Statistics, Social Survey Division, nor the Data Archive, University of Essex, bears any responsibility for the analysis or interpretation of the data described in this paper.

References
1 Department of Health. Smoking kills: a White Paper on tobacco. London: The Stationery Ofce, 1998. 2 Department of Health. The health of the nation: a strategy for health. London: HMSO, 1992. 3 Department of Health. Statistics on smoking cessation services in England 2001 to March 2002. Statistical Bulletin 2002/25. London: The Stationery Ofce, 2002. 4 Pechmann C, Dixon P, Layne N. An assessment of US and Canadian smoking reduction objectives for the year 2000. Am J Publ Hlth 1998; 88: 13621367. 5 Mendez D, Warner KE. Smoking prevalence in 2010: why the Healthy People goal is unattainable. Am J Publ Hlth 2000; 90: 401403. 6 West R, McNeill A, Raw M. Smoking cessation guidelines for health professionals: an update. Thorax 2000; 55: 987999. 7 Walker A, Maher J, Coulthard M, Goddard E, Thomas M. Living in Britain: results from the 2000 General Household Survey. London: The Stationery Ofce, 2002. 8 Mendez D, Warner KE, Courant PN. Has smoking cessation ceased? Expected trends in prevalence of smoking in the United States. Am J Epidemiol 1998; 148: 249258. 9 Hughes JR, Gulliver SB, Fenwick JW, et al. Smoking cessation among self quitters. Hlth Psychol 1992; 11: 331334. 10 Vecchia CL, Decarli A, Pagano R. Prevalence of cigarette smoking among subsequent cohorts of Italian males and females. Prev Med 1986; 15: 606613. 11 Brenner H. A birth cohort analysis of the smoking epidemic in West Germany. J Epidemiol Commun Hlth 1993; 47: 5458. 12 Birkett NJ. Trends in smoking by birth cohort for births between 1940 and 1975: a reconstructed cohort analysis of the 1990 Ontario Health Survey. Prev Med 1997; 26: 534541. 13 Laaksonen M, Uutela A, Vartiainen E, et al. Development of smoking by birth cohort in the adult population in Eastern Finland 197297. Tobacco Control 1999; 8: 161168. 14 Kemm JR. A birth cohort analysis of smoking by adults in Great Britain 19741998. J Publ Hlth Med 2001; 23: 306311. 15 Ronneberg A, Lund KE, Hafstad A. Lifetime smoking habits among Norwegian men and women born between 1890 and 1974. Int J Epidemiol 1994; 23: 267276. 16 Goddard E, Higgins V. Smoking, drinking and drug use among young teenagers in 1998. Volume 1 England. London: The Stationery Ofce, 1999 (and earlier volumes in series).

Downloaded from http://jpubhealth.oxfordjournals.org/ at Universiti Sains Malaysia on April 4, 2014

Conclusion
In historic mode the model gives insight into the behavioural changes that determine smoking prevalence trends. It is shown that the proportions of never smokers who start and of current

324

JOURNAL OF PUBLIC HEALTH MEDICINE

17 Ofce of Population Censuses and Surveys, Social Survey Division. General Household Survey (computer les for various years). Colchester: The Data Archive (distributor). Computer les for: 1974 (SN 1027), 1976 (SN 1361), 1978 (SN 1465), 1980 (SN 1766), 1982 (SN 2034), 1984 (SN 2154), 1986 (SN 2569), 19881989 (SN 2724), 19901991 (SN 2937), 19921993 (SN 3166), 19941995 (SN 3538). 18 Ofce for National Statistics, Social Survey Division. General Household Survey, 19967 (computer les for various years). Colchester: The Data Archive (distributor). Computer les for: 19967 (2nd edn; SN3804), 19989 (SN 4134), 20001 (SN 4518). 19 Wald N, Nicolaides-Bouman A. UK smoking statistics, 2nd edn. Oxford: Oxford Medical Publications, 1991. 20 Ofce for National Statistics. Mortality statistics series 1998. Cause: England and Wales. Series DH2 No. 25. London: The Stationery Ofce, 1999 (and earlier volumes in series). 21 Ofce for National Statistics. Birth statistics 1998 England and Wales. Series FM1 No. 27. London: The Stationery Ofce, 1999 (and earlier volumes in series).

For each age band Xnev/ev = (Sobs + Eobs) (Spred + Epred) Xcur/ex = Eobs Epred where Xnev/ev are numbers changing from never to ever (current or ex) smokers in a year, Xcur/ex are numbers changing from current to ex in a year, Sobs and Eobs are numbers of current and ex smokers from survey data, Spred and Epred are respective numbers predicted from ageing of the previous years population calculated as follows:
YSa = (Y1)S(a1) (1 db)

Appendix: The structure of the model

The model represented the number of current smokers, ex smokers and never smokers in single years. Initially, all year groups in the same smoking age band (1619, 2024, 2534, 3544, ..., 6574, 75+ years) are set to have the same prevalence of behaviours but this changes as the model runs. Because of the data available, the model assumes that no one starts smoking before the age of 15. The prevalence of smokers and ex smokers among 15-year-olds was assumed to be a xed proportion of that in the 1619 years age band. Death rates for current smokers and non-smokers were calculated from age-specic overall death rate and prevalence of smoking in each 5-year age band, assuming a relative risk for all-cause mortality in smokers relative to non-smokers (never smokers and ex smokers) of 1.7 in men and 1.3 in women, as used by Brenner.11 The appropriate death rate is applied to each year group to estimate the number of survivors. The survivors are then moved to the next year group and the new births placed in the lowest age group (under 1 year). The number changing behaviour each year is calculated from the difference between the number in each state predicted as a result of ageing of the previous years population and the observed number from survey data. To prevent drift in the population structure, the model population was compared every 5 years with Census-based estimates and adjusted to match them. In addition to the logical changes of starting, quitting and lapsing, a further change from ex to never (termed reclassifying) was necessary to t the data.

where YSa is the number of smokers of age a at year Y, (Y1)S(a1) is the number of smokers of age (a 1) at year (Y 1), and db is the death rate applicable to that age band. A positive Xnev/ev represents numbers starting (never to current) and a negative Xnev/ev represents represents numbers reclassied (ex to never). A positive Xcur/ex represents numbers quitting and a negative Xcur/ex represents numbers lapsing. Those starting are expressed as a percentage of never smokers and those quitting as a percentage of current smokers. Those reclassied and lapsing are expressed as a percentage of ex-smokers.

Downloaded from http://jpubhealth.oxfordjournals.org/ at Universiti Sains Malaysia on April 4, 2014

Data to populate the model

Information on prevalence of current, ex and never smokers has been collected by the General Household Survey in even numbered years since 1972. Data from this survey kindly supplied by the Data Archive, University of Essex,17,18 were analysed to obtain information on current smoking status. The percentage of current, ex and never smokers in age bands (1619, 2024, 2534, 3544, 4554, 5564, 6574 and 75+ years) was obtained. Data for the period 19501972 (even-numbered years) were obtained from data collected by the Tobacco Advisory Council published in tables 4.1.1, 4.1.2, 6.1.1 and 6.1.2 of UK Smoking statistics.19 The earlier data were not available in the precise age bands required so prevalence for the age bands used was estimated from those available. Where data for the precise year was not available data for the adjacent year was taken. Prevalence in odd-numbered years was estimated by interpolation. Age composition of the populations, death rates in 5-year age bands (<1, 14, 59, 1019, ..., 8589, 9094, 95+ years) and birth rates for England and Wales were obtained from published sources.20,21

Accepted on 10 June 2003