Elsevier

Addictive Behaviors

Volume 27, Issue 4, July–August 2002, Pages 509-527
Addictive Behaviors

Intraclass correlation among measures related to cigarette use by adolescents: Estimates from an urban and largely African American cohort

https://doi.org/10.1016/S0306-4603(01)00189-7Get rights and content

Abstract

This paper presents the first estimates of school-level intraclass correlation (ICC) for smoking-related variables from an urban and largely African American population. Seventh graders (n=6967) from 39 middle schools in Memphis, TN, were measured at baseline in 1994 and annually through 1997. Mixed model regression methods were used to estimate variance components for school and residual error. School-level ICCs were large enough, if ignored, to substantially inflate the Type I error rate in an analysis of treatment effects. We show how those correlations can be reduced using regression adjustments and used to determine sample size for future school-based smoking prevention studies.

Introduction

Smoking prevention studies directed toward youth often utilize designs in which classrooms or schools are assigned to conditions, while observations are taken on individual students (US Department of Health and Human Services, 1994). Such studies are an example of the broader class of comparative studies called group-randomized trials (GRTs). A GRT is a study in which identifiable groups are assigned to conditions, while observations are taken on members of those groups (Murray, 1998). Such studies exist in many disciplines and pose a number of design and analytic problems not present when individuals are randomized directly to conditions Murray, 1998, Murray et al., 1994. The most important is that observations taken from members of the same group are likely to be correlated (Kish, 1965). This intraclass correlation (ICC) reflects a component of variance attributable to the groups in addition to the usual variation attributable to their members. It also violates the independence of errors assumption associated with the most familiar analytic methods. Simulation studies have shown that any analysis that ignores the ICC will have a Type I error rate that is inflated, often badly Murray et al., 1996, Murray et al., 1998, Murray & Wolfinger, 1994, Zucker, 1990. This problem is exacerbated because the precision available to estimate the ICC is based on the number of groups assigned to each condition, and that number is often limited. These factors can reduce power so that it may be impossible to detect important intervention effects in an otherwise well-designed and properly executed study.

Even so, the GRT remains the best comparative design available when investigators wish to evaluate an intervention implemented at a group level or one that cannot be delivered to individuals. The best advice is to plan a large-enough study to allow for the expected ICC and to take advantage of other design and analytic strategies to limit its impact. To do so, investigators need good estimates of the ICC expected in their studies (Murray, McKinlay, et al., 1994). These problems are not new, as Cornfield (1978) identified them more than two decades ago.

Several recent studies have provided estimates for outcomes related to adolescent smoking Murray & Hannan, 1990, Murray et al., 1994, Murray & Short, 1997, Siddiqui et al., 1996. The estimates presented in these papers were based largely on samples of White adolescents; only Siddiqui et al. (1996) reported estimates for other racial or ethnic groups. In addition, most of these estimates were based on cross-sectional analyses; only Murray, Rooney et al. (1994) presented estimates from cohort analyses. Finally, most of the previous estimates were unadjusted; only Murray and Short (1997) documented the reduction in tobacco use ICCs available from regression adjustment for covariates, and then only for serial cross-sectional data.

This paper will address two major gaps in the literature. We will present the first estimates of school-level ICCs for measures related to cigarette use from a cohort study involving an urban and largely African American student population. We will also document for the first time the extent to which regression adjustment for baseline values can reduce these ICCs in cohort data. Importantly, we will illustrate the use of these estimates to plan future prevention studies.

Section snippets

Design, population, and survey methods

The data reported here were derived from an ongoing cohort study of adolescent smoking (Robinson, Klesges, Zbikowski, & Glaser, 1997). The entire 1994 seventh grade enrollment of a large mid-south urban school system comprised the target population for the cohort study; 39 schools were eligible and all participated in the project. The baseline survey was conducted in 1994 when the cohort was in the seventh grade and annual follow-up surveys were conducted each spring through 1997.

The data were

Participation

At baseline, 79% of the 8828 eligible seventh graders completed the survey. Only 3% declined to participate in the study, with another 2% withheld due to parental refusal or our inability to notify the parents of the research project. An additional 16% of the students did not complete the study because of absenteeism or problems with survey administration. The demographic characteristics of the 6967 respondents closely match those of the target population in this school system. The students

Comparison to previous studies

The only published report of school-level ICCs for smoking prevalence in a African American adolescent cohort appeared in Siddiqui et al. (1996). They reported that their school-level ICC estimates for smoking prevalence were zero, not only at baseline, but at each of three annual follow-up surveys. Those estimates were unusually low compared to data published for other samples (Murray, Rooney, et al., 1994), but were consistent with their overall findings, which suggested that African American

Acknowledgements

This work was supported by a grant from the National Heart, Lung, and Blood Institute (R01-HL-50723) to Robert Klesges, PhD, Principal Investigator.

References (28)

  • N.M. Laird et al.

    Random effects models for longitudinal data

    Biometrics

    (1982)
  • R.C. Littell et al.

    SAS system for MIXED models

    (1996)
  • D.M. Murray

    Design and analysis of group-randomized trials

    (1998)
  • D.M. Murray et al.

    Planning for the appropriate analysis in school-based drug use prevention studies

    Journal of Consulting and Clinical Psychology

    (1990)
  • Cited by (24)

    • An examination of the shift in school-level clustering of US adolescent electronic cigarette use and its multilevel correlates, 2011-2013

      2016, Health and Place
      Citation Excerpt :

      However, no research to our knowledge has investigated the school-level correlates of e-cigarette use or how the importance of schools as contexts for the production and reproduction of adolescent e-cigarette use has shifted over time. Guided by theory and past findings linking school contexts to youth tobacco use (Cole and Leatherdale, 2014; Murray et al., 2002), we expect that the increase in adolescent e-cigarette use has occurred alongside uneven increases in the rate of e-cigarette use across schools. Not all schools are likely to have contributed equally to the population-wide increase in teen e-cigarette use.

    • A Prospective Study of Perception in Adolescent Smoking

      2009, Journal of Adolescent Health
      Citation Excerpt :

      For current smoking this was .031. This is in line with those estimates that were found in prior studies on school effects in adolescent smoking [23]. Analyses in which the ICC was accounted for showed virtually the same results.

    View all citing articles on Scopus
    View full text