Work

[Commentary] REGRESSION TO THE MEAN IN ADDICTION RESEARCH

Year: 2007
Type: article
Abstract: The phrase ‘regression to the mean’ (RTM) is used in a number of contexts in addiction research. The word ‘regression’ arose from a description of the imperfect inheritance of characteristics from one generation to the next [1]. The relationship between parent and child is imperfect; e.g. the children of very tall parents will, on average, be less tall than their parents. A regression line provides a way to estimate the expected height of a child given the heights of the parents. It is easy to forget, however, that ‘regression to the mean’ is a description, not an explanation. RTM can arise in multiple ways. In the context of the height of children versus their parents, measurement error is a minor issue and the interpretation of the attenuated regression coefficient is that parents' heights are only two among many variables affecting the height of a child. In contexts where measurement error is a concern, regression to the mean can be expected because of the imperfect reliability of the instrument being used; an extremely high or low error in measurement at time 1 is likely to be followed by a more moderate error value at time 2. The third major mechanism by which RTM arises is measurement bias; that is, there is a true change from time 1 to time 2, but the scores we observe at time 1 and/or time 2 are not representative of the full distribution. Our outcome measures are typically indicators of substance consumption or clinical severity that vary over time. In most clinical studies, the participants are recruited at the time of seeking help for a clinical problem, which is usually when they are experiencing unusually high levels of distress or consequences. Because of this bias at the time of intake, we say that the dramatic improvement noted between pre-treatment and post-treatment is due at least partially to RTM. We less often say that the progressive deterioration that often occurs after treatment is partially attributable to RTM, but that change is also the result of a mixture of true change, random noise and bias. The three papers in this issue discuss how RTM phenomena may affect research conclusions, and discuss some ideas of how to address these regression phenomena by analytical and/or design methods [2–4]. First, it is crucial to be aware of the range of circumstances under which RTM can affect research conclusions, including the conclusions of literature reviews or meta-analyses. Dr John Finney's paper offers guidance on how RTM can arise and, under some circumstances, how it can be detected [2]. He also discusses limitations of some common approaches such as covariance adjustment that attempt to minimize bias arising from RTM. The other two papers address how to cope with RTM in longitudinal evaluation studies where random assignment is not feasible [3,4]. The paper by Gmel et al. contrasts four different methods for dealing with RTM effects in multiwave studies [3]. The paper by Ripatti & Makala considers autoregressive models, also for multi-wave studies [4]. We would like to note in passing that the three RTM papers in this issue do not represent an exhaustive overview of the statistical literature related to this topic. In particular, methodological readers may wish to consult the Bayesian literature, especially with regard to shrinkage estimators which are notably less biased than classical estimators when comparing three or more entities (for a technical introduction see [5], for a health-care application see [6]). We would like readers to draw two primary conclusions from studying these papers on RTM. First, under many circumstances RTM effects may dwarf intervention effects, and readers of the research literature need to be aware of its possible impact on reported findings. Secondly, we hope that readers will conclude that RTM is not an incomprehensible phenomenon, but one that, with the right design and analytical model, can be studied, its effects estimated and a better picture (although still an imperfect one) obtained of the impact of treatment or other events on outcome.
Source: Addiction
Language: English
Subfield: Epidemiology
Field: Medicine
Open Access status: bronze
OpenAlex ID: W1891988096
FWCI: 0.362
Cites: 6
Cited by: 11
Related to: 10