Population distribution measurements of health indicators for chronic diseases can be used to target health care and community resources to areas with greatest need. The Centers for Disease Control and Prevention (CDC) and state health departments expend time, money, and energy measuring the distribution of health conditions and behavioral risk factors throughout the United States. The Behavioral Risk Factor Surveillance System (BRFSS), a cornerstone of these efforts with an annual budget of about $15 million, collects data on smoking, weight, diet, exercise, preventive medical care, and other behaviors (1,2). However, the BRFSS was designed to provide stable estimates at the metropolitan and state levels; because of limited sample sizes, creating stable estimates at the neighborhood level requires complex modeling and pooling several years of data (3).

Practice-based research networks (PBRNs) and integrated health delivery systems also collect patient behavioral health data in the course of routine clinical care (4,5). With increasing use of electronic health records (EHRs), behavioral data are more commonly available in standard formats (6). Data from larger health systems and PBRNs could provide a complementary or alternative method of monitoring neighborhood-level prevalence of behavioral risk factors.

Tobacco use remains the leading cause of preventable death in the United States (7). Knowledge about neighborhood-level smoking prevalence is necessary to guide the efficient deployment of community-based resources. Patient smoking status is routinely recorded in EHRs; EHR data derived from PBRNs and health systems may provide accurate community-level estimates of the prevalence of tobacco use. To determine if PBRN data could be used to make reasonable neighborhood-level estimates of smoking prevalence, we compared PBRN data to population-based, neighborhood-level estimates of smoking prevalence derived from the BRFSS.

Neighborhood-level smoking prevalence can be estimated using EHR data from a PBRN, collected in the routine course of clinical care. Generally, PBRN estimates were slightly lower than the BRFSS estimates, but they were higher in some neighborhoods in which the PBRN had higher market penetration.

PBRN smoking prevalence estimates may have been generally lower for several reasons. Practices did not have smoking status documented for 100% of patients. Though not directly comparable, our 73% documentation rate compares favorably to the BRFSS response rate of 48%, a rate that is declining (10). In addition, clinicians preferentially document the smoking status of smokers. Efforts to increase smoking status documentation, including local and national incentives and EHR reminders, preferentially add nonsmokers, leading to small changes in the measured smoking prevalence. Our previous intervention to increase smoking status documentation served to increase the rate of documentation only for patients who were former smokers and never smokers; there was no significant change in the proportion of documented smokers (11). The discrepancies between PBRN and BRFSS estimates may also be attributable to the differential coverage of the populations by PBRN and BRFSS. Also, our PBRN patient population differs from the greater Boston area adult population, the sampling frame for the BRFSS, by having a lower percentage of Medicaid beneficiaries, more women than men, and likely healthier or more health-conscious patients (12).

Other studies have noted similarities and differences between health system–based and population-based estimates of smoking prevalence. A study in Leicester, England, found that General Practice notes, computerized and manual, tended to overestimate population smoking prevalence (13). On a larger scale in the UK, a comparison of an EHR-based PBRN, containing approximately 6% of the UK population, to the population-based General Lifestyle Survey showed excellent agreement both nationally and regionally (14,15). However, this analysis is not geographically granular enough to allow targeting of neighborhood resources. In the United States, commercial EHRs with broad national penetration may have greater potential to make regional and national estimates of acute and chronic conditions, as well as behavioral risk factors (16).

Using PBRN or health system data to measure the neighborhood prevalence of smoking offers several advantages over traditional, population-based methods. First, the data are already collected as part of routine clinical care, so their collection is less expensive than using a separate infrastructure to conduct population-based surveys. Second, for small-area estimation, as our data show, the sample sizes available in PBRN data are much larger, resulting in smaller standard errors (although relative to neighborhood population, even the PBRN estimates might be considered small), which may allow for the targeting of community-based interventions in smaller areas than is feasible using the BRFSS. Third, given the larger sample size and collection during routine clinical care, the data are potentially more current than population-based surveys, for which time is needed for data to become available (14). Fourth, because data were taken from an EHR, the potential exists for smoking status to be linked to a richer data set consisting of other behaviors, comorbidities, medications, and health outcomes for varied patients (17). Similarly, EHR-based measurements can be linked with practice-based treatment interventions (11,18–20). Finally, the emergence and implementation of EHR data standards as part of the national Meaningful Use EHR Incentive program will allow for information pooling across multiple health systems and encourage routine documentation of smoking status as it did for the General Practice pay-for-performance contract in Britain (21). Some of the inherent data problems associated with market share of this single PBRN could be resolved by pooling data from multiple health systems.

Despite these advantages to PBRN and health system data, advantages exist to population-based estimates, such as those derived from BRFSS. Obviously, despite low and declining response rates, the BRFSS is population-based and can provide estimates regardless of whether an individual seeks care through a particular health system or seeks health care at all. The BRFSS is not dependent on “market share” to get more accurate estimates, is not dependent on the presence of health care facilities in neighborhoods, and is not subject to health system peculiarities that may limit the generalizability of PBRN data. Second, although it was not intended to provide the small-scale prevalence we calculated, the BRFSS is consistently administered across the United States. Third, the BRFSS potentially provides greater consistency over time. Although standards are emerging for the structure of EHR data, the BRFSS data are collected with greater attention to consistency in definitions and measurement. Population-based surveys are not free from bias, however. For example, smokers generally have lower response rates to surveys than do nonsmokers (13). Finally, patients may be more likely to report negative health behaviors like smoking to health care providers than via a population survey.

Our analysis suggests that PBRNs and health system data can be used to guide community resources to neighborhoods with greater need. However, an understanding of the limitations of estimates derived from nonpopulation-based sources, for example, data from a single health system with low market penetration, cannot be used for small-area estimation, because the results are potentially biased because of insufficient population coverage and selection bias.

The true public health benefit of using health system–based risk factor information will come from combining data from multiple health systems, which requires a change in view of health system data as a public health resource. A convergence of health system data with population-based results should be seen more often as documentation improves (14). Health system data could be part of a systematic surveillance system that would help in understanding the effectiveness of tobacco control programs (22), particularly with mandated reporting of behavioral risk factors such as tobacco use, obesity, alcohol use, and others (23). Such a system could afford greater opportunity for collaboration and synergy between health care systems, public health departments, community organizations, and other community-based resources. In effect, this may encourage the surrounding community to take a greater stake in the health system and the health system to take a greater stake in the surrounding community.