Methods

The analysis was divided into two parts: a) modelling of baseline prevalence and b) development of prevalence estimates for 2020 and 2030 if MPOWER policies were implemented in 2010.

To develop the baseline prevalence estimates for 2020 and 2030, we first assembled prevalence time series for selected countries to compute the current trend of smoking rates. To construct the time series, we employed the tobacco use component of the WHO's Global InfoBase Database, a comprehensive compilation of historical tobacco use prevalence data in 191 countries. The countries are grouped into six main regions, corresponding to the official WHO regional divisions: AFRO (Africa), AMRO (the Americas), EMRO (the Eastern Mediterranean), EURO (Europe), SEARO (South-East Asia) and WPRO (the Western Pacific). The database includes prevalence rates obtained from published and non-published sources of governmental as well as non-governmental surveys. The database provides detailed information on each survey, including the definition of tobacco use (eg, current smokers vs. daily or ever-smokers), tobacco type, age groups, gender and geographical representation (national vs sub-national). In each region, we chose to focus on the 10 countries with the largest number of current cigarette smokers. This group of 60 target countries accounts for around 90% of the world's smokers and 85% of the world's population. Table 1 shows the selected countries and each country's share of the global smoking population.

Using the raw (unadjusted) data from the InfoBase, we followed a multistage process to construct time trends for each of the target countries. First, we chose from among the surveys available in the InfoBase those that were conducted at the national level and reported overall prevalence rates for current cigarette smokers. Second, for those countries that had only a few surveys satisfying the criteria of the first stage, we looked for national surveys that provided prevalence rates for current smokers by gender and age categories. We then used the appropriate population estimates to derive a weighted average overall estimate for current smokers. Third, for countries where stages 1 and 2 result in insufficient data points to develop a trend, we made use of the available subnational data or data on daily smokers (daily smoker figures exclude people who are current smokers but do not smoke every day), making the necessary adjustments. This entailed the use of assumptions that varied by country and the available data. The data and process we used to construct the time series for each country are described in the online appendix.

Additionally, for selected countries, we supplemented the time trends with data from the WHO's Global Adult Tobacco Survey (GATS),11 a nationally representative household survey, launched in 2007 in 16 low- and middle-income countries that account for more than half of the world's smokers. These data are also shown in the online appendix. While in many countries, men and women exhibit dramatically different smoking patterns, due to the lack of sufficient data in several countries, we were not able to perform separate analyses for men and women.

To compute expected base case prevalence for the years 2020 and 2030, we adapted the method we have successfully used to model and forecast the natural trajectory of adult smoking prevalence in the USA to accommodate data from other countries. As in the US case, we modelled prevalence as a stock/flow process by keeping track of the accumulation over time of smokers who enter the system through smoking initiation and leave it due to smoking cessation or death. Our approach has been discussed extensively elsewhere.12–15 The following expression describes the basic process:dPrev(t)/dt=Init-Prev(t)×(Cess+Mortality) (1)

with solution:Prev(t)=(Prev0−(Init/(Cess+Mortality)))×exp(−(Cess+Mortality)×t)+(Init/(Cess+Mortality)) (2)

Prev(t) stands for adult smoking prevalence at time t for the country under study; Init is the country-specific smoking initiation rate; Cess is the country-specific smoking cessation rate; Mortality is the country-specific mortality rate and Prev₀ is the estimated prevalence at the time we chose to initialise the country-specific time series.

To calibrate the model, we first obtained estimates for the adult initiation rate by averaging historical smoking prevalence at young ages (17–24) for each country with available data, thus assuming that little initiation occurs after age 24. (We recognise that there is initiation after age 24 in some countries, which is a limitation of our study. In these countries, both prevalence and policy effects will be underestimated.) Then, for countries with a time series of two or more historical prevalence data points, we obtained estimates for the cessation rate and the initial prevalence (Prev₀), employing the model above and using non-linear least squares according to the following objective function:MinCess,Prev0∑(Prev(Init,Cess,Prev0,Mortality,t)−Observed Prevalence(t))2. (3)

With the derived estimates, we used the model to project a smoking prevalence base case scenario for the years 2020 and 2030 for each country with available data, accounting for population growth. For countries with only one data point, we projected the same prevalence value to the years 2020 and 2030 as the base case but conducted a demanding sensitivity analysis around this assumption.

For each country in the study, we performed an extensive sensitivity analysis to develop worst- and best-case scenarios for the years 2020 and 2030. For countries with two or more observed prevalence data points, the sensitivity analysis consisted of conducting a Monte Carlo simulation to assess the variability of the 2020 and 2030 prevalence estimates according to the following procedure:

First, we fitted a triangular probability distribution to the data on initiation rates by selecting the minimum, maximum and mode statistics from the data as the distribution parameters. Then, we sampled from that distribution 10 000 times, each time obtaining a new estimate of the cessation rate according to expression (3) and a projected prevalence for the years 2020 and 2030. Finally, we collected the distribution of projected prevalence values for 2020 and 2030 and reported the 2.5 and 97.5 percentiles as the best- and worst-case scenarios for smoking prevalence.

For the sensitivity analysis for countries with only one data point we assumed, as the worst-case scenario, that their smoking rates would follow the same upward trajectory as the USA prior to the 1960s. As such, we used as worst-case values for 2020 and 2030, the prevalence attained by the USA in 10 and 20 years, respectively, following the time when the USA reached the target country's most current prevalence estimate. As the highest prevalence attained by the USA was <50% (in the late 1950s), we did not compute a worst-case scenario for countries that are currently at 50% or higher smoking prevalence.

To estimate global smoking prevalence for the years 2020 and 2030 along with best- and worst-case scenarios, we conducted a Monte Carlo simulation sampling from the individual country distributions for the 2020 and 2030 prevalence obtained in the previous stage of the analysis. Each country's sample was then weighted by the country's adult population projection for 2020 and 2030 to obtain regional and overall world prevalence.16 The mean and the 2.5 and 97.5 percentiles were taken as the point estimate, best-case and worst-case estimate for the base case scenario of global prevalence in years 2020 and 2030.

To estimate the global impact of applying the smoking control policies described in the MPOWER package, we first obtained, from published sources, estimates of each policy's effectiveness on smoking initiation and cessation rates.17–19 We derived the estimate for the effectiveness of cessation support by taking the average of the effectiveness of different types of cessation support interventions reported by Wilson and colleagues.17 Assuming that just a fraction of current smokers will use cessation support, we then multiplied this average effectiveness by a cessation participation factor of 10%, consistent with current literature, to obtain the 1.061 estimate we used in our calculations. The effectiveness values we used in our analysis are summarised in table 2.

Except for the impact of taxes on smoking rates, values in table 2 show the improvement in initiation or cessation rates due to the application of individual MPOWER policies. For example, the table shows that a national mass media and package warning campaign would increase cessation rates by 23% and reduce initiation rates by 20%. For the impact of taxation on smoking rates, the table reports the price elasticity of initiation rates and the price elasticity of smoking prevalence. For example, a tax increase that resulted in a 100% cigarette price increase would decrease prevalence by 20% within 1 or 2 years and produce a 70% reduction in initiation rates, both of which will be maintained as long as the real price reduction is sustained. We do not estimate reductions in daily smoking by continuing smokers.

We performed the policy impact analysis assuming that no MPOWER policies had been applied prior to 2010 and that in 2010, all MPOWER policies were implemented globally and simultaneously, with tax increases that would produce a 100% price increase for cigarettes. We also assume that this will be a real price increase, sustained throughout the time span of our analysis (ie, we assume that taxes will be adjusted for inflation annually to maintain the same proportional level above the price). We chose the 100% level as a plausible if ambitious target, recognising that this would require an extremely large tax increase in nearly all countries.

We also considered different potential interactions among the individual policies that affect initiation and cessation rates when applied simultaneously. First, for a best-case scenario, we assumed that the policies acted independently on the population; that is, the fact that an individual is affected by one policy does not indicate that he/she is more or less likely to be affected by other policies. Discarding the possibility of negative correlation policy effects among individuals (ie, a person affected by one policy is less likely to be affected by other policies), the independence assumption provides the maximum combined effect of the policies on the 2020 and 2030 prevalence estimates, and thus, it constitutes our best-case scenario. We discuss this in greater detail in the online appendix.

For the worst-case scenario, we assumed maximum correlation among the effects of policies on individuals. This implies that the same subset of individuals is likely to be influenced by any policy. In this case, the maximum possible effectiveness from all policies combined is limited to that of the most effective policy.

Finally, for a more plausible intermediate case, we performed a Monte Carlo simulation sampling values for the combined policy effectiveness from a uniform distribution bounded by the best- and worst-case scenarios just described. On each iteration of the simulation, the sampled effectiveness modifies a pair of initiation and cessation rates resampled from the base case simulation run.

As in the base case scenario, the global prevalence for the years 2020 and 2030 is calculated by weighting the individual country estimates by their projected population size in 2020 and 2030, respectively.16