The National Health Interview Survey (NHIS) is commonly used as one of the standards for determining current smoking prevalence in the U.S. adult population. The research findings based on the NHIS data indicate that the percentage of current smokers in the U.S. has declined over the past four decades from 42.2% in 1965 to 20.8% in 2006, Centers for Disease Control & Prevention (2008)
. Another survey that is used to provide valid estimates of current smoking prevalence in the U.S. adult population is the Tobacco Supplement to the Current Population Survey (TUS-CPS). Although the TUS-CPS is conducted less frequently than the NHIS, it has a much larger sample size than the NHIS and thus, it allows for estimates of smoking prevalence with smaller standard errors for small groups and geographical regions, National Cancer Institute (2007a)
. These characteristics are why smoking prevalence estimates based on the TUS-CPS data are widely used in many tobacco research studies, National Cancer Institute (2007b
) and have been used as one standard for state estimates, see Biener et al. (2004)
In this paper we present a detailed analysis of TUS-CPS data in which our primary goal was to identify survey methodologic factors that highly influence smoking prevalence trend results. Numerous studies show that current smoking prevalence depends on demographic characteristics. For example, Shavers et al. (2005)
assessed racial/ethnical differences in current smoking by occupation and industry based on the TUS-CPS for 1998-1999. They estimated that American Indians and Alaska Natives had the highest smoking prevalence, at 35.1%, whereas Asian and Pacific Islanders had the lowest rate, at 15.2%. These patterns were consistent with those of the NHIS, Centers for Disease Control and Prevention (2008)
. Fagan et al. (2007)
investigated the sociodemographic factors associated with smoking among unemployed adults based on the TUS-CPS data for 1998-1999 and 2001-2002. In particular, the authors examined employment attributes in relationship to current and former smoking and successful quitting among unemployed adults. Green et al. (2007)
used 2003 TUS-CPS data to examine the relationship between smoking behaviors of young adults and their education level, and concluded that education was an important predictor of current smoking prevalence and that young adults with a college education were half as likely to smoke as those without a college education. Pleis and Lethbridge-Çejku (2007)
investigated specific chronic conditions such as smoking with respect to multiple sociodemographic characteristics, including education level. They concluded that adults with at least a bachelor’s degree were less likely to smoke than the other adults and they were more likely to have never smoked in their past. These results reinforce a common research finding that race/ethnicity, employment status, and education are important predictors of current smoking prevalence.
To identify the unique contributions of survey methodologic factors on smoking prevalence, we investigated whether and to what degree respondent type and survey administration mode affect current smoking prevalence estimates. Our basic assumption in conducting this analysis was that self-reports provide a valid assessment of current smoking. This assumption is supported by the results of a comparison of self-reported smoking prevalence to the biochemical measurement of serum cotinine concentration, which used data from the Third National Health and Examination Survey for 1988-1994, Caraballo et al., (2001)
. The authors conclude that the overall smoking estimates based on self-reports and the biochemical measurements are approximately the same. These findings also confirm a meta-analysis conducted by Patrick et al. (1994)
, which investigated 26 publications presenting 51 comparisons between self-reported smoking and biochemical measures of smoking. The authors concluded that in most of the studies, self-reports of smoking provided an accurate measure of smoking. Hence, self-reports can be used to obtain viable information regarding current smoking status.
Our first research goal was to investigate whether in addition to self-responses proxy-responses can be relied on to accurately estimate current smoking prevalence. Including proxy-respondents in a survey is highly beneficial because it reduces survey costs and increases response rates, but proxy-responses usually result in higher measurement error. We review some general results regarding self- and proxy-respondents and then discuss this topic with respect to smoking, specifically.
Sudman, Bradburn and Schwarz (1996
, Chapter 10) state that although proxy-respondents may have very limited information on sensitive questions related to the individual for whom they are reporting, they may still be more honest in their reponses than when they report about themselves. The authors also point out that convergence of self- and proxy-responses highly depends on the joint participation or discussion between proxy-respondent and the individual, where the joint participation is related to activities that are naturally shared (e.g., viewing television, eating out) and the joint discussion is related to other activities (e.g., reading) or attitudes. Thus, location and distance can considerably affect the accuracy of a proxy-response for certain questions.
Another cause associated with differences between self- and proxy-responses is proxy-respondents’ reliance on inferences. Todorov (2003)
uses the NHIS on Disability data to conclude that proxy-respondents rely on inferences more than do self-respondents. For example, instead of trying to find out the exact number of doctor visits made by a person in the past 12 months, a proxy-respondent is more likely to provide an estimated number. The author argues that because proxy-respondents have less information than self-respondents a priori
, they rely more on inferences when responding to a question. Thus, self- and proxy-respondents differ not only in the amount of information known about a survey question, but also in the cognitive strategies used to generate a response to a question.Todorov’s (2003)
findings suggest that, with respect to assessing disabilities, proxy-responses can introduce a measurement bias. For further discussion of theoretical differences between self- and proxy-respondents, cognitive laboratory results of multiple studies, and accuracy of proxy-reporting, we refer the reader to Sudman, Bradburn and Schwarz (1996
, Chapter 10).
Findings on the difference between self- and proxy-responses specifically when assessing smoking status also provided useful background for our study. A number of studies have been carried out where both self- and proxy-responses for current smoking status have been obtained in selected populations, and the findings reported from these studies have been discordant. For example, Gilpin et al. (1994)
proposed to add questions about current smoking status to ongoing surveys and stated that “one adult could provide smoking status for all household members.” This conclusion was based on results of a California Tobacco Surveys data analysis. The authors explored potential differences in reporting smoking habits by adult self- and proxy-respondents. They concluded that although the highest discrepancy of responses was observed for a case in which self- and proxy-respondents were unrelated, none of the relationship groups, e.g., parent-guardian, spouse/partner, sibling, other relative, and unrelated were significantly different from the reference group, which was the child. The analysis was based on a logistic regression with main effects only, thus, no interaction terms were considered. Similarly, Hyland et al. (1997)
investigated the differences between self- and proxy-responses in terms of a number of self-respondent characteristics based on a cross-sectional telephone survey. They found that age, race, family income, and current smoking status (current smoker, recent quitter, long-term quitter, and never-smoker) of self-respondents were associated with discrepant results. However, because these differences were generally minimal, the authors concluded that “proxy-reported smoking status is an accurate and effective means to monitor population-wide smoking prevalence of adults.” In contrast, Navarro (1999)
examined the agreement between self- and proxy-reported smoking in relationship to race/ethnicity based on the 1990 California Tobacco Survey. The author concluded that this agreement was significantly different by race/ethnicity group. Thus, it may be important to control for race/ethnicity when self- and proxy-responses are used. Likewise, Harakeh et al. (2006)
assessed the correspondence between self- and proxy-respondents in a full family study and showed that although adolescents age 13 to 17 years could be used to obtain reliable information regarding their parents’ smoking habits, parents appeared to report less accurate information regarding their children’s smoking status.
Because the TUS-CPS survey allows both self- and proxy-responses for certain questions, it permits examination of possible discrepancies between self- and proxy-reported current smoking. In our study, we investigated differences in self- and proxy-reported current smoking prevalence and the degree to which any differences in prevalence depended on multiple sociodemographic characteristics of the person whose smoking status was being reported. Our other major question of interest was whether the survey mode (phone versus in-person) is related to the accuracy of current smoking prevalence estimates. Brick and Lepkowski (2008)
point out that generally, telephone assessments are less expensive than any other types of interviewer-administered assessments. Thus, telephone mode is commonly used in large survey studies. However, the potential for measurement bias associated with different survey modes may be a concern when current smoking status is assessed because reporting current smoking may be a sensitive question for many subjects. The most consistent finding from the earliest research using a variety of surveys and outcomes is that the mode effect is insignificant, see Groves et al. (1987)
and De Leeuw (2005)
. However, Simile, Stussman and Dahlhamer (2006)
discussed validity of telephone and in-person follow-up interviews using the 2005 NHIS data and showed that personal visits resulted in significantly different estimates than telephone responses with respect to multiple key health indicators, including current smoking status. Nevertheless, their analysis did not adjust for respondent type because the NHIS has a negligible fraction of proxy-responses and did not incorporate any interactions and time trends because it was based on a single survey year. St-Pierre and Beland (2004)
also explored the survey mode effect based on the Canadian Community Health Survey data. They showed a significant difference in a number of key health indicators, such as obesity, physical inactivity, current/occasional smoking status of respondents age 20 to 29 years and some others based on the personal visit and telephone surveys.
By incorporating survey mode in our analysis together with the most important sociodemographic characteristics and respondent type, we were able to examine the impact of this survey methodologic factor on current smoking prevalence estimates, thereby also confirming or disputing the contradictory findings regarding the effect of survey mode. In addition, because our data were taken from TUS-CPS survey waves from 1992 to 2003, we were able to adjust for decreasing smoking trends over time.