The privileged unit of analysis in psychology is the individual (Nesselroade, Gerstorf, Hardy, & Ram, 2007). Nevertheless, many data-analytic approaches coarsely aggregate data and tacitly assume group-average models to hold and to be interpreted in lieu of more fine-grained and, ultimately, person-specific models. For example, when a group of persons show an average increase of performance in a learning task, this does not mean that all persons follow a pattern of change similar to this average. In fact, none of the persons may be well represented by the average trend. In a similar vein, Tucker (1966) argued that the consideration of differences instead of averages will allow us to gain more information about the nature of basic functions underlying behavior. Ever since, researchers have been questioning coarse aggregation of data across persons (e.g., Lamiell, 1981; Nesselroade & Molenaar, 1999) as the estimates of averaged effects may not be representative of any single individual. In fact, strong inference about intra-individual variation from interindividual variation is only possible under the ergodic assumption (Molenaar, 2004), which assumes that the group model represents each individual's dynamics (homogeneity) and that those dynamics have constant characteristics in time (stationarity). In the same vein, Simpson (1951) pointed out that a statistical relationship observed in a population could be reversed within subgroups that form the population. For instance, “It may be universally true that drinking coffee increases one's level of neuroticism; then it may still be the case that people who drink more coffee are less neurotic” (Borsboom, Kievit, Cervone, & Hood, 2009, p. 72). Simpson's paradox may arise whenever inferences are drawn across different explanatory levels, for example, from populations to the individual, or from cross-sectional data to intraindividual change over time (see Kievit, Frankenhuis, Waldorp, & Borsboom, 2013, for further illustrations). Hence, there still is a need for focusing on individuals or subgroups of individuals to more accurately model individual process idiosyncrasies and similarities across persons. Particularly, in light of large-scale empirical data sets, aggregation is more likely to lead to models with low informative value about individual underlying processes as it is often difficult to expand prior hypotheses to account for the large number of potential explanatory variables.
Quote is from this article (click on image to enlarge)
- Posted using BlogPress from my iPad