Abstract
We compared projections from a dynamic model of US adult smoking prevalence with official estimates of prevalence from the National Health Interview Survey. Ten years after they were made, the model projections closely fit the National Health Interview Survey estimates for 2005 and 2010. We conclude that a verified model of adult smoking prevalence can assist governmental authorities in establishing aspirational but feasible targets for tobacco control. By extension, carefully crafted models can help in goal setting in multiple areas of public health.
There is considerable interest in the potential utility of systems models to assist governmental leaders in establishing and evaluating public health objectives.1,2 Objectives are used for multiple purposes, including to plan resource allocation within state and local health agencies or even at the national or international level; to encourage collaboration among governmental and nongovernmental agencies in pursuit of common goals; to evaluate progress, or the lack thereof, in addressing public health problems; and to set aspirational goals as a means of motivating the public health workforce. The question arises, therefore, as to whether modeling can produce accurate estimates of important public health phenomena and, if so, how those estimates can be employed in objective setting.
In at least 1 instance, the release of recent National Health Interview Survey (NHIS) findings allows assessment of the accuracy of systems model estimates and consideration of their applicability in the objective-setting process. In September 2011, the Centers for Disease Control and Prevention (CDC) reported smoking prevalence among US adults in 2010 and compared the figure with prevalence in 2005.3 In 2000, we published projections of smoking prevalence for both years,4 based on a dynamic simulation model of smoking prevalence developed in the mid-1990s.5 We updated the projections for 2010 in a later article.6 The release of the new data by CDC therefore permits an assessment of how accurate model projections can be for a period of up to a decade. The comparison also allows consideration of how model projections can be used in the development of public health objectives.
To project future smoking prevalence under different scenarios for smoking initiation and cessation rates, we employed a dynamic model of smoking prevalence. The model predicts future smoking prevalence by estimating future US population size by age and smoking status. For the model’s baseline predictions, we used age-specific smoking initiation and cessation rates estimated from data from the NHIS. We did not predict future prevalence by simply extrapolating current smoking trends. Instead, the model incorporates population dynamics structures that determine smoking prevalence. Based on best estimates of birth, mortality, and smoking initiation and cessation rates, the model simulates the state of the adult smoking population in the US population at a future time. In essence, the model tracks the number of smokers over time, with each year’s population of smokers enhanced the following year by the entry of new smokers and diminished by the departure of existing smokers because of smoking cessation or death.
We calculated annual smoking prevalence by comparing the number of smokers to the size of the adult population, also tracked by the model. We have used the model to develop projections of future smoking prevalence under status quo conditions (i.e., assuming continuation of existing initiation and cessation rates) and to calculate how much (and how quickly) initiation and cessation rates would need to change to achieve specific smoking prevalence targets.4 We have described the model in detail elsewhere.5
We compared projections from our model with actual (CDC-reported) smoking prevalence in 2005 and 2010. We also compared estimates and realized prevalence with the Healthy People 2020 goal for the nation with regard to adult smoking prevalence.7 We then considered the implications of the model’s accuracy for smoking prevalence target setting and, by extension, for objective setting in multiple areas of public health.
Table 1 shows actual smoking prevalence for each of 2005 and 2010, as reported by CDC, and, for the same years, projected prevalence from the Méndez et al. model.5 The model was remarkably accurate in forecasting smoking prevalence. For the version calibrated with actual data through 1995,4 one projection assumed that smoking initiation remained at 30%. With this assumption, the model projected 2005 NHIS-reported prevalence precisely, with both the forecast and the NHIS estimate at 20.9%. The projection for 2010 was 19.9%, 0.6 percentage points higher than the NHIS figure.
TABLE 1— Actual and Projected Adult Cigarette Smoking Prevalence: United States, 2005 and 2010
| Source | Model Calibrated Through | Initial Year of Projection | 2005, % | 2010, % |
| Actual prevalence: National Health Interview Survey3 | … | … | 20.9 | 19.3 |
| Projected prevalence | ||||
| Méndez and Warner,4 assuming continuation of 30% initiation rate | 1995 | 2000 | 20.9 | 19.9 |
| Méndez and Warner,4 assuming initiation rate declines from 30% to 15% from 2000 to 2010 | 1995 | 2000 | 20.5 | 18.4 |
| Méndez and Warner6 | 2005 | 2006 | … | 19.1 |
Smoking initiation has decreased since 2000. Using the data through 1995, we had made a second projection that assumed that initiation decreased from 30% to 15% from 2000 to 2010. (For the purposes of this model we defined initiation as smoking prevalence for the group aged 18–24 years.) The actual initiation rate decreased to 22%, implying a model projection for 2005 roughly halfway (20.7%) between the Mendez and Warner projections, assuming a continuation of the 30% initiation rate (20.9%) and a decline from 30% to 15% (20.5%). That figure is 0.2 percentage points below actual adult smoking prevalence of 20.9% in 2005. The same averaging of the initiation assumptions implies a prevalence of 19.1% in 2010–again, 0.2 percentage points below the actual prevalence of 19.3%, well within the 95% confidence limits. below the actual prevalence of 19.3%, well within the 95% confidence limits. As the model was calibrated with data earlier than 2000 for this study, the analysis demonstrates that projections made up to 10 years can be quite accurate.
Our projection for 2010 from a version of the model calibrated with actual data through 20056 was also 19.1%, again just 0.2 percentage points below the NHIS-reported actual smoking prevalence for 2010 and again well within the 95% confidence limits.
The comparison of actual and projected prevalence demonstrates that a dynamic simulation model can provide excellent estimates of smoking prevalence at least as far as a decade into the future. Another model calibrated between the 2 articles in Table 1, by Levy et al., also performed well.8 The question thus becomes, how should models like these be employed in establishing public health objectives?
The most visible and perhaps important of all public health objective-setting exercises in the United States is embodied in the decennial Healthy People reports. The original Healthy People report, which established goals for 1990, was published in 1979.9 Subsequent Healthy People reports defined objectives for the nation in health promotion and disease prevention for 2000,10 2010,11 and 2020.7 For many states, the reports have constituted a blueprint for establishing their own set of objectives, planning their health promotion programming and that of local health departments, and evaluating their progress.
The Web site for Healthy People 2020 describes the objective-setting process as “science-based,” with each objective having “a reliable data source.”7 Moreover, a document prepared by the Advisory Committee to the Secretary of Health and Human Services specifies that Healthy People 2020 should set realistic targets based on knowledge of what is potentially achievable given the health issue and current or emerging knowledge of interventions, programs, and policies that might result in improvement. However, targets should also represent “a reach,” and should be more than a continuation of the status quo [emphasis in original].12(p2)
The document also states: Setting targets for populations requires estimating what the level of the objective would be if the status quo were to continue. Projections should begin with an estimate of where the objective would be at the end of the specified time period (e.g., a decade) if nothing changed. A scientific assessment should then be completed to incorporate what is known about the level of change that could be anticipated as a result of employing known-effective or likely-effective interventions.12(p2)
We concur. We believe that a scientific approach, utilizing projections such as those from our model, would have improved target-setting for adult smoking prevalence in the 2010 Healthy People objectives. Shortly after release of the 2010 draft objectives, we published projections for where smoking prevalence would stand in 2010 under status quo conditions.4 Furthermore, we used the model to estimate what combinations of reductions in initiation rates and increases in cessation rates would be necessary to achieve the Healthy People draft goal of 13% (subsequently lowered to 12% in the final version, and retained for 2020). The estimates from the model demonstrated that the Healthy People goal was essentially impossible to achieve; it would have required a reduction in initiation and an increase in cessation far in excess of plausible improvements. We also projected smoking prevalence in 2010 under what we considered to be “stretches”—ambitious yet conceivable goals. Had the country been successful in cutting smoking initiation in half over the course of the decade, and also doubling the cessation rate by the end of the period, we projected that adult smoking prevalence would have been 16.7%. That is the kind of goal that would have been science-based.
To contemplate a reasonable smoking prevalence goal for the nation for 2020, we used our model to project smoking prevalence under 2 conditions, reporting the results in 20086: if initiation and cessation rates nationwide remained unchanged from their (then) present levels, we projected that adult smoking prevalence would decrease to 16.8% in 2020. By contrast, if all of the states could achieve California’s initiation and cessation rates by 2010, prevalence would decrease to 14.7%. California has the lowest smoking prevalence in the nation (14% in 2010) with the exception of Utah (9.3% in 2010). Utah’s smoking prevalence is heavily influenced by its high concentration of Mormons. Although California’s low prevalence reflects the state’s racial/ethnic composition in part, most is attributable to a highly effective tobacco control program.13 Given our analysis, we concluded that an adult smoking prevalence target of 14% for the United States by 2020, though a definite stretch, might have made sense. The Healthy People target of 12%, although still more of a stretch, is thus within the realm of possibility, if not likely.
Modeling is no panacea. For one thing, the modeling itself must be sound. There are multiple approaches to modeling, and the optimal approach will vary according to the research question to be addressed. And as the old adage puts it, “garbage in, garbage out”: effective modeling, regardless of the approach, requires sound logic and the use of appropriate and accurate data. But when done well, and especially when the accuracy of projections from models can be verified, modeling can facilitate rational objective setting and planning more generally. Unrealistic targets can produce cynicism and indeed the abandonment of the planning process associated with them. By contrast, feasible targets—targets that, with a real push, will be achieved in some jurisdictions—create motivation and with it striving to achieve. Public health has major and urgent problems to tackle. Models will never be the solution to those problems. But they can and should assist in finding solutions rapidly and efficiently.
Human Participant Protection
No human participant protection was required because no human participants were involved in this study. The research relied solely on model projections and comparisons using data from the National Health Interview Survey, itself cleared through standard federal government procedures.
