METHODS

Data source

The percentages of women giving birth who smoked during pregnancy for the years 1990 to 2000 in Wisconsin were obtained from US Standard Certificates of Birth maintained by the Wisconsin Vital Records Section, Wisconsin Bureau of Health Information. Information on smoking during pregnancy is recorded on the birth certificate by the attending physician, nurse, or other health professional at the time of delivery, and is obtained from the woman during prenatal care visits or at the time of delivery. The birth certificate records the answer to two questions: “Did you use tobacco at any time during the pregnancy?” (Yes/No); If “yes”, “What was the average number of cigarettes smoked per day?”.

Data analysis

Data were analysed separately at the state and individual county level. Analysis consisted of three steps: (1) modelling historical trends in the prevalence of smoking during pregnancy; (2) using models to predict future prevalence rates; and (3) using predicted prevalence to set community specific goals in tobacco control.

Modelling trends

First, the prevalence of smoking during pregnancy (that is, the proportion of births in which the birth certificate indicated that the mother smoked during the pregnancy) was calculated for each year, 1990–2000. Trends in prevalence for all counties (n = 72) and the state over the 11 year period were then modelled using linear least squares regression of log of prevalence on calendar year. Modelling log transformed rates over time allows for decreasing rates of decline in prevalence over time, avoids prediction of negative future prevalence, and, in Wisconsin, more accurately described trends (R² = 0.98 for the state) than models using non-transformed prevalence. However, the meaning of results derived from models of log transformed rates is less easily explained than models of non-transformed rates. Estimates of approximate relative annual percentage change in prevalence were calculated for each regression by taking the exponential of the slope of the regression line (m) and applying the formula (100*(e^m−1)), a method described and used by the National Cancer Institute to estimate trends in cancer incidence and mortality for the SEER reporting system.²³ Relative annual percentage change can be interpreted as a constant percentage change in prevalence from year to year.

Predicting future prevalence

The regression models were then used to predict county specific and overall state prevalence rates at a future point in time. The year 2005 was chosen for illustration, as it is the year the Wisconsin Tobacco Control Board will evaluate progress towards tobacco control initiatives, funded by an annual appropriation from the legislature of approximately $15 million ($3 per capita), beginning in 2001. To assess the ability of the models to predict future prevalence of smoking among pregnant women giving birth, a regression model of the first 10 years of data (1990–1999) (as described above) was used to predict county specific prevalence one year in the future (2000). These estimates were highly correlated with the actual (observed) county specific prevalence rates (Pearson correlation coefficient = 0.90). The regression model using all 11 years of data (1990–2000) was used to predict prevalence in the year 2005, with the expectation that accuracy of prediction decays for each additional year in the future. Ninety five per cent confidence intervals (95% CI) were computed for the model estimates.

Setting goals using predicted prevalence

Finally, predicted prevalence in the year 2005 was used to illustrate a pragmatic method for setting community specific objectives in tobacco control. Objectives are set according to a percentage of the rate predicted by historical trend analysis. For illustration, we selected 2005 prevalence objectives that are 20% less than the 2005 prevalence rates that would be expected if fitted historical trends during the 1990s were maintained (a time when Wisconsin’s annual tobacco control budget was approximately $1 million). County specific as well as overall state objectives are set in this manner.

RESULTS

Modelling trends

In the state of Wisconsin over the 11 year period, the prevalence of smoking during pregnancy declined from 22.9% of pregnant mothers (1990) to 16.5% (2000). When modelled, this represents an average relative annual decrease in prevalence of 3.3% of the rate per year (95% CI −3.7 to −2.9) (fig 1, table 1).

Across counties (n = 72) in the year 2000, the prevalence of smoking during pregnancy varied considerably, from 7.5% of pregnant women to 53.5% (median 18.8%). Relative annual percentage change in prevalence ranged from −5.5% per year to 1.6% per year (median −1.9%) (table 1). Most counties (n = 53) demonstrated trends that were significantly different from the state (that is, the 95% CI for the country’s relative annual percent change estimate did not overlap the point estimate for the state)—19 counties had faster rates of decline and 34 counties had slower rates of decline. Counties with larger numbers of annual births tended to have lower prevalence of smoking among pregnant women giving birth, and faster rates of decline in prevalence over time. Data from four counties and modelled trend estimates are shown in fig 2 and table 1. These counties illustrate county level variation in prevalence, annual relative percentage change in prevalence, and number of births, and highlight differences between county trends and the overall state average.

Predicting future prevalence

The regression models describing historical trends in prevalence of smoking during pregnancy were used to obtain estimates for prevalence in the year 2005. Overall, if the state trend continues, Wisconsin is predicted to have a prevalence of 13.8% of women giving birth who smoked during pregnancy (95% CI 13.2 to 14.4) (fig 1). County estimates range from 7–52% in 2005 (median 17.0%) (table 1).

Setting goals using predicted prevalence

Use of predicted prevalence to define tobacco control objectives at the state or county level is illustrated in figs 1 and 2. A model of state trends in smoking prevalence would suggest that Wisconsin is already predicted to come very close to meeting the current tobacco control board objective of 13.2% in 2005, an objective that reflects a 20% reduction in prevalence from the observed rate in 2000. At the county level, if trends continue as modelled, 16 counties will meet or beat a county goal of a 20% reduction from the observed countywide prevalence in 2000. Six counties show historical trends of increasing prevalence (relative annual percentage change greater than zero), and are predicted to have slightly higher prevalence than currently observed in 2000 (table 1).

Rather than setting tobacco control objectives as a reduction from a recently observed rate, we suggest setting objectives as a percentage less than the rate predicted by historical trend analysis. At the state level, 20% less than 13.8% (the predicted prevalence estimate for 2005) is 11.0%, the suggested goal for 2005 (fig 1). County specific goals were set in a similar manner, according to predicted prevalence in 2005 (fig 2, table 1).

DISCUSSION

This paper describes a method for utilising local data to set community specific tobacco control objectives. Using Wisconsin birth certificate data as an example, we demonstrated pronounced variation in the rates of smoking during pregnancy among women giving birth across all counties, and more importantly, variation in the rates of change over the past decade. Substantial local variation suggests the need for setting future tobacco control objectives at the local level that account for historical variation in progress across communities. Linear regression was used to model trends in prevalence over time, and models were used to predict future rates of smoking. This enables counties to set appropriate local objectives as a percent less than the predicted future prevalence, and enables states to assess variation in progress that may result from variation in tobacco control resources, population migration patterns, cultural beliefs, and changing sociodemographic characteristics of communities over time.

In practice, setting specific objectives would depend on the magnitude of planned interventions. For example, some communities may set an objective at the predicted future rate, assuming that the interventions and trends in population characteristics continue as in the past. Conversely, communities with substantial increases in tobacco control resources may set more aggressive objectives; or the choice of an appropriate percentage reduction from predicted might be set by the state programme. For example, counties in the lowest quartile of prevalence of smoking could be expected to achieve a 15% reduction from predicted, whereas counties in the highest quartile could be expected to achieve a 20–25% reduction, since it may be easier to achieve reductions when the overall prevalence is high. Counties that historically kept the state rate high could be targeted for additional funds and programmatic support. These decisions could also be informed by estimates such as the relative effectiveness of interventions aimed at helping pregnant smokers quit smoking, which are provided by the Public Health Service Clinical Practice Guideline.²⁴

Local planning efforts may depend on state or national objectives that are inappropriate to local conditions or are based on a percentage reduction from one recently observed rate without accounting for past trends in progress. In Wisconsin, models indicate that nearly a quarter of counties and the state itself are expected to meet or beat current 2005 objectives (based on a 20% reduction from the observed rate in 2000) merely by continuing at the rate of change seen over the last 11 years. A few counties have increasing rates of smoking during pregnancy, suggesting that meeting this objective will be more difficult. Guidelines from effective comprehensive tobacco control programmes, such as those found in California and Massachusetts, help to fill this gap, by presenting many useful indicators and example objectives to guide specific local programme planning, which include monitoring the mean number of in-store tobacco signs in retail stores, increasing the number of community events conducted with smoke-free policies, and increasing the proportion of schools that provide smoking cessation programmes for faculty, staff, and students.^25,26 However, these documents do not provide systematic guidance to communities to suggest how historical trends in local progress might modify choice of an appropriate objective. The objective setting approach presented in this paper has the important virtue of systematically incorporating both historic trends and local variation in prevalence and rates of change of prevalence of smoking using birth certificate data.

This method also permits states to assess county level variation in trends over time and guide programme planning. In Wisconsin, model projections will serve as the “expected” rates of smoking in 2005 for each county, based on past intervention efforts and/or changes in sociodemographic and economic community characteristics. The rates observed in 2005 can subsequently be compared to these expected rates, to assess potential variation in progress. Counties that beat their predicted prevalence could be compared with those that did not, and reasons for variation can then be explored. In Wisconsin, the more populated counties (for example, Dane and Waukesha) have the lowest rates of smoking during pregnancy whereas sparsely populated areas tended to have much higher rates (for example, Shawano county), illustrating geographic variation and potentially warranting more aggressive marketing of a telephone quitline in high prevalence areas, for example. Menominee county, an entirely native American Indian reservation, has the highest rate in the state, illustrating variation potentially explained by sociodemographic factors and suggesting a need for culturally appropriate programmes. Finally, variation in programmatic factors is evident—for example, in smoke-free restaurant policies, which have greater success in more urban communities, and advertising campaigns that vary across communities—indicating a need for increased funding in certain areas. Subgroup analyses could be performed to examine trends by age, race, education, and even financial (such as Medicaid) status. Ideally, policy makers would like to compare trends in tobacco use in their communities with other communities, to assess whether their programmes are more effective in reducing tobacco use. Greater progress compared with predicted trends would suggest that community based programmes are having a greater effect, although many factors must be considered in such an evaluation design.²⁷

Importantly, the method of data analysis and local objective setting described in this paper could be applied to other consistent sources of local data collected over time in communities. For example, the Behavioral Risk Factor Surveillance System (BRFSS) (all states since 1996), the National Youth Tobacco Survey (YTS) (since 1999), and/or the Youth Risk Behavior Survey (since 1991) all offer the flexibility of sub-state estimates of youth tobacco use prevalence. The YTS will be most useful in the future, when several years of data become available for trend analysis. Birth certificate data, on the other hand, provides information on all registered births (that is, it is not a survey estimate), and is more widely available over the past decade. Additionally, question items are consistently stated across jurisdictions in the USA, data are simple to analyse,¹ and, in general, trends in smoking rates based on birth certificate data have been confirmed with data from other sources, including the Pregnancy Risk Assessment Monitoring System, the National Survey of Family Growth, and the National Pregnancy and Health Survey in estimating the prevalence of smoking during pregnancy.^7–14 In validation studies, smoking ascertainment on the birth certificate was found to be 70–85% complete.^13,14,28

The choice of an appropriate data source should consider the relevance to local conditions, and communities choosing to use birth certificate data to set local objectives should be aware of some important limitations. First, the relative availability and accessibility of birth certificate data should not be the only reason for using it to inform community planning. Planning efforts aimed specifically at reducing the prevalence of smoking during pregnancy among women giving birth may lead to narrow cessation interventions. However, population based interventions could also influence the proportion of women smoking during pregnancy by reducing rates of initiation in children and adolescents, and targeting people of all age groups with broad based strategies.

Second, though most US states collect comparable information, a few do not. During the 1990s, the number of US states that collected comparable data on smoking during pregnancy from the birth certificate changed, with no reporting from California and South Dakota for the entire decade, and Indiana and New York State beginning reporting in 1999. This is not seen as a serious limitation, however, since the range in the percentage of all birth certificates collecting comparable information about smoking status increased from 75% in 1990 to 87% in 1999.¹⁶

Outside the USA, survey data are more commonly used to estimate prevalence of smoking during pregnancy. In England, for example, as part of a smoking and pregnancy initiative, the Health Education Authority carried out a series of annual surveys using quota sampling of pregnant women from 1992 to 1997 to measure behaviour in relation to smoking during pregnancy.²⁹ In Italy, the prevalence of smoking among pregnant women has been estimated through case–control studies.³⁰ Much more limited information is available in developing countries, where progress towards universal birth registration is underway, particularly in the Sub-Saharan Africa and South Asia regions.^31,32 Demonstrating effective use of birth certificate smoking data in the USA may be useful to these nations as they make progress towards ensuring that important information is collected and registered at the time of birth. However, the cost of implementing system-wide change is an important consideration.

Third, when modelling local trends in smoking among women giving birth, one inevitably encounters difficulty in communities where there are few absolute numbers of births. This scarceness of data can reduce the ability to accurately predict prevalence several years in the future. However, as illustrated in Wisconsin, even smaller rural counties with populations under 10 000 (n = 5), according to the US Census in 2000, have an average of 81 births each year. This limitation could be overcome by grouping adjacent communities matched on certain characteristics of interest. For example, data from neighbouring counties with shared programmes, organisational goals, or cultural characteristics might analyse a combined dataset to set group objectives. Alternatively, data from counties with similar overall prevalence of smoking or similar historical trends in prevalence might be combined to improve confidence in predicted rates.

Finally, estimates of the prevalence of smoking among pregnant women also tend to be somewhat lower when data collected on birth certificates is used.^11,14,33 Several studies have suggested that the lack of specificity in the birth certificate questions is the principal factor accounting for reported differences in levels of smoking during pregnancy.^13,14 Additionally, information is collected only on women whose pregnancies result in live birth. Smoking is a major cause of spontaneous abortion and still birth,^34–37 and these are not measured with birth certificate data. Underreporting may also result from women inaccurately reporting smoking behaviour to their health care provider. Particularly in cases of poor birth outcome, the mother might be less willing to admit to having smoked at all during pregnancy if asked about lifestyle factors after the delivery.¹⁶ Other reasons for underreporting may include the failure of the physician to inquire about smoking behaviour or the incorrect transfer of information from the prenatal care record to the certificate at the time of delivery, and these limitations should be considered.

In summary, this paper describes and illustrates the usefulness of a method for setting community specific tobacco control objectives that uses model estimates to set reasonable objectives for the future. While tobacco control programmes indicate that local initiatives are an important component of reducing tobacco use, few methods are available to guide objective setting that account for local variation in past progress. The method described uses the US Standard Birth Certificate as a source of routinely collected local data that can be used to determine the prevalence of smoking during pregnancy, an important indicator of the burden of tobacco use, and to set local objectives. Though the USA is the only country known to us to collect information on smoking during pregnancy on the birth certificate, the event and registration of birth may represent an equally useful opportunity for data collection on smoking for other countries throughout the developed and developing world.