Adult Health Behaviors

In the recent article, “Trends in Socioeconomic Inequalities in Adult Health Behaviors Among U.S. States, 1990–2004” (Public Health Reports, March/April 2007, p. 177–189), Sam Harper and John Lynch used multiple years of the Behavioral Risk Factor Surveillance System (BRFSS) to investigate the socioeconomic inequality of various health behaviors at the state level. The BRFSS is a valuable dataset in that it is the largest ongoing national telephone survey, with hundreds of thousands of participants each year. Analyses, such as Harper's and Lynch's, are imperative to the public health community's understanding of health disparities; more specifically to this article, how socioeconomic disparities translate to health behaviors.

When evaluating their study, we are concerned about state-level differences in BRFSS response rates. States have autonomy over how the BRFSS is executed (i.e., questionnaire length, whether data collection is in-house or contracted out, sampling design), which subsequently differentially impacts the final state-level response rates. Further, BRFSS response rates declined on both the national and state levels over the past decade. In 1994, the median response rate was 69.9%, with Michigan reporting the lowest (42.6%) and North Dakota reporting the highest (86.8%).¹ By 2004, the median response rate for the BRFSS was 52.7%, with New Jersey reporting the lowest response rate (32.2%) and Nebraska reporting the highest (66.6%).² If low response rates increase the possibility of survey error biasing results, how would this impact the calculation of the Relative Concentration Index and the reported results? Nonresponders are usually less educated, younger, and racially/ethnically diverse.³ Could increases in the socioeconomic inequality of health behaviors over time be due to nothing more than nonresponse?

The study attempts to correct for nonresponse by using a sample weight provided by the Centers for Disease Control and Prevention (CDC) in the analytic stage. This correction weight works by bumping up the BRFSS numbers to match the age, gender, and racial distribution of the targeted population for that state. Yet, the underlying assumption is that nonresponders answer all survey questions identically to those responders in the same age, gender, and racial category. However, the amount of bias introduced by nonresponse depends not only on the proportion of the sample that fails to respond, but also the extent to which the nonresponders are systematically different from the entire population. We do not know how nonresponders to the BRFSS differ beyond age, gender, and race from those who respond. It would be prudent to assume that the nonresponse correction weight will not effectively eliminate nonresponse bias.

The authors assert that their results “represent the ‘true’ population health and inequality burdens of these behaviors given the actual demographic composition of the state” (p. 187). Research evaluating the extent of nonresponse bias in the BRFSS and the validity of the missing completely at random assumption that the CDC-provided sampling weights imply is sorely needed. Until the time when such methodologic research is available, researchers using BRFSS data should consider the role of nonresponse bias as a noncausal explanation and heed caution when interpreting their findings.