Send letter to journal:
Re: Multilevel models of schools' influence on pupils' smoking

Pokorny et al show that one must use a multilevel model to accurately identify contextual influences, such as school characteristics, on the behaviour of individuals. Neither aggregate models nor individual level only models will be accurate(1). This is a good point well made.

Unfortunately, Pokorny et al use aggregated school level perceived prevalence of smoking among peers as their contextual example variable. Using this variable makes the model difficult to interpret. A sentence from the Discussion (p306) highlights this. Abbreviated, it reads ‘students in schools with higher …perceived tobacco use… were more likely to be smokers than students in schools with lower …peer tobacco use.’ This statement could only be false if students’ perceptions of smoking prevalence were totally inaccurate. The observers and the observed have both been sampled and are reporting on one another. However, it is tempting for authors and readers to interpret these findings as support for the notion that peer perception is part of the mechanism for why some schools have a higher prevalence than others, when it simply restates that finding.

Aside from this, there is another important reason why the prevalence of smoking among peers, either at the student level or the school level, should not be included in a multilevel model seeking to understand inter- school variation in smoking prevalence. Students’ perception of smoking prevalence among their friends is a measure of actual school smoking prevalence. If we are seeking to understand why inter-school variation in smoking prevalence exists, then controlling for it will attenuate the phenomenon that we are seeking to explain. The papers published by Moore et al West et al illustrate this(2;3). In both cases, controlling for the number of friends smoking vastly diminished the unexplained school variation in smoking. This does not show that the phenomenon is explained; it shows the circular reasoning above. West et al appreciated this, and showed that turning the multilevel equation round to predict the prevalence of friends smoking showed the same inter-school variation and ranking of schools as was apparent when predicting individuals’ smoking habit. Notable individual level risk factors, such as peer smoking and sibling smoking, should not be included in multilevel models.

Many people I know feel that the school they attended shaped them as individuals and it seems reasonable to assume that it influences individuals’ smoking. It is welcome to have papers exploring this underdeveloped area. We have recently reviewed this literature and made suggestions as to which variables should be controlled in a multilevel model exploring inter-school variation, which variables should be excluded, and which variables should be used as potential causal variables in mediation analysis(4). The paper by West et al should be essential reading for anyone seeking to understand how schools influence students’ health behaviour(3).

Reference List

(1) Pokorny, S.B., Jason L.A., & Schoeny M.E. (2004) Current smoking among young adolescents: assessing school based contextual norms. Tobacco Control, 13, 301-307. (2) Moore, L., Roberts C., & Tudor-Smith C. (2001) School smoking policies and smoking prevalence among adolescents: multilevel analysis of cross-sectional data from Wales. Tobacco Control, 10, 117-123. (3) West, P., Sweeting H., & Leyland A. (2004) School effects on pupils' health behaviours: evidence in support of the health promoting school. Research Papers in Education, 19, 261-291. (4) Aveyard, P., Markham W., & Cheng K.K. (2004) A methodological and substantive review of the evidence that schools cause pupils to smoke. Social Science & Medicine, 58, 2253-2265.