American Journal of Epidemiology Advance Access originally published online on August 2, 2005
American Journal of Epidemiology 2005 162(6):603; doi:10.1093/aje/kwi252
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
LETTERS TO THE EDITOR |
THE AUTHORS REPLY
1 Clinical Research Unit, Hvidovre University Hospital, University of Copenhagen, DK-2650 Hvidovre, Denmark
2 Division of Public Health Community Medicine, Department of Clinical Science, Malmö University Hospital, Lund University, S-205 02 Malmö, Sweden
We read the letter by Dr. Kaufman (1
) with great interest, and we thank him for raising a point that is important for the motivation of one of the two measures (the interval odds ratio) that is advocated in our article (2).
Kaufman argues that the cluster-level fixed effects may be given a counterfactual interpretation. As a matter of fact, Kaufman goes further and writes as follows: "It is the causal interpretation that motivates the use of regression modeling in public health research, since this provides evidence on which to base policy decisions" (1, p. 602). While this may be true, it should be noted that the regression model does not itself ensure that the effects may be given causal interpretations. In this specific model, we find that the counterfactual interpretation of a cluster-level covariate is dubious, because it builds on assumptions that cannot be investigated and that are unlikely to hold. These two points are demonstrated in the example below. From an epidemiologic point of view, information conveyed by measures of variation provides evidence on which to base policy decisions. Using the median odds ratio and the interval odds ratio allowed us to evaluate the relative importance of the cluster level for various kinds of outcomes, and this may promote cluster-level interventions for those health outcomes that are—to a greater extent than others—determined by the cluster. These considerations are important when attempting to determine the efficacy of focusing an intervention on places instead of people. When the neighborhood variance is small, focusing the intervention on specific places may be a less optimal strategy. However, we still can find relevant association between cluster-level variables and health that may be of interest. By neglecting information on cluster-level variation or considering it as a nuisance, the exclusive consideration of measures of association between cluster characteristics and individual health in regression analysis may lead to inappropriate conclusions (3).
Kaufman provides an example, with community as the clustering factor and presence of a hospital in the community as a cluster-level covariate. Kaufman concludes: "If the factual condition from some community j is that there is no hospital present, then this estimated causal effect is interpretable as the effect of the hypothetical action of putting one there" (1, p. 602). This is indeed a tempting interpretation, because it is causal rather than associational. However, it relies entirely on the assumption that the random effects would remain the same under the hypothetical experiment of manipulating the cluster-level covariate. Not only can this assumption not be investigated, it constitutes an extrapolation of the data that is generally unwarranted in epidemiologic cross-sectional studies.
To illustrate our point, we extend the example given by Kaufman to a situation where the outcome Y is an indicator of individuals having asthma. In most Western European countries, asthma is more prevalent in urban areas than in rural areas. At the same time, people living in cities are more likely to have a hospital in their community than are people in the countryside. A causal interpretation of the effect of having a hospital in the community would infer that building a hospital in a rural area increases the prevalence of asthma in the area, a conclusion that is hardly trustworthy.
A causal interpretation of a cluster-level covariate may only be inferred, if the assumption of the random effects' remaining the same under the hypothetical experiment of manipulating the cluster-level covariate can be justified. This is usually not the case in epidemiologic studies because of confounding.
 |
ACKNOWLEDGMENTS |
|---|
Conflict of interest: none declared.
|
References |
|---|
 TOP References |
|---|
- Kaufman JS. Re: "Appropriate assessment of neighborhood effects on individual health: integrating random and fixed effects in multilevel logistic regression." Am J Epidemiol 2005;162:602–3.
[Free Full Text] - Larsen K, Merlo J. Appropriate assessment of neighborhood effects on individual health: integrating random and fixed effects in multilevel logistic regression. Am J Epidemiol 2005;161:81–8.
[Abstract/Free Full Text] - Merlo J. Multilevel analytical approaches in social epidemiology: measures of health variation compared with traditional measures of association. J Epidemiol Community Health 2003;57:550–2.
[Free Full Text]
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||